Image Processing and Analysis Reference

Note:
The chapter describes functions for image processing and analysis. Most of the functions work with 2d arrays of pixels. We refer the arrays as "images" however they do not neccesserily have to be IplImage's, they may be CvMat's or CvMatND's as well.

Drawing Functions

Drawing functions work with arbitrary 8-bit images or single-channel images with larger depth: 16s, 32s, 32f, 64f All the functions include parameter color that means rgb value (that may be constructed with CV_RGB macro) for color images and brightness for grayscale images.

If a drawn figure is partially or completely outside the image, it is clipped.

CV_RGB

Constructs a color value

#define CV_RGB( r, g, b )  (int)((uchar)(b) + ((uchar)(g) << 8) + ((uchar)(r) << 16))

Line

Draws simple or thick line segment

void cvLine( CvArr* img, CvPoint pt1, CvPoint pt2, double color, int thickness=1, int connectivity=8 );

img
The image.
pt1
First point of the line segment.
pt2
Second point of the line segment.
color
Line color (RGB) or brightness (grayscale image).
thickness
Line thickness.
connectivity
Line connectivity, 8 (by default) or 4. It is possible to pass 0 instead of 8.

The function cvLine draws the line segment between pt1 and pt2 points in the image. The line is clipped by the image or ROI rectangle. The 8-connected or 4-connected Bresenham algorithm is used for simple line segments. Thick lines are drawn with rounding endings. To specify the line color, the user may use the macro CV_RGB( r, g, b ).

LineAA

Draws antialiased line segment

void cvLineAA( CvArr* img, CvPoint pt1, CvPoint pt2, double color, int scale=0 );

img
Image.
pt1
First point of the line segment.
pt2
Second point of the line segment.
color
Line color (RGB) or brightness (grayscale image).
scale
Number of fractional bits in the end point coordinates.

The function cvLineAA draws the 8-connected line segment between pt1 and pt2 points in the image. The line is clipped by the image or ROI rectangle. The algorithm includes some sort of Gaussian filtering to get a smooth picture. To specify the line color, the user may use the macro CV_RGB( r, g, b ).

Rectangle

Draws simple, thick or filled rectangle

void cvRectangle( CvArr* img, CvPoint pt1, CvPoint pt2, double color, int thickness=1 );

img
Image.
pt1
One of the rectangle vertices.
pt2
Opposite rectangle vertex.
color
Line color (RGB) or brightness (grayscale image).
thickness
Thickness of lines that make up the rectangle. Negative values, e.g. CV_FILLED, make the function to draw a filled rectangle.

The function cvRectangle draws a rectangle with two opposite corners pt1 and pt2.

Circle

Draws simple, thick or filled circle

void cvCircle( CvArr* img, CvPoint center, int radius, double color, int thickness=1 );

img
Image where the line is drawn.
center
Center of the circle.
radius
Radius of the circle.
color
Circle color (RGB) or brightness (grayscale image).
thickness
Thickness of the circle outline if positive, otherwise indicates that a filled circle has to be drawn.

The function cvCircle draws a simple or filled circle with given center and radius. The circle is clipped by ROI rectangle. The Bresenham algorithm is used both for simple and filled circles. To specify the circle color, the user may use the macro CV_RGB ( r, g, b ).

Ellipse

Draws simple or thick elliptic arc or fills ellipse sector

void cvEllipse( CvArr* img, CvPoint center, CvSize axes, double angle,
                double startAngle, double endAngle, double color, int thickness=1 );

img
Image.
center
Center of the ellipse.
axes
Length of the ellipse axes.
angle
Rotation angle.
startAngle
Starting angle of the elliptic arc.
endAngle
Ending angle of the elliptic arc.
color
Ellipse color (RGB) or brightness (grayscale image).
thickness
Thickness of the ellipse arc.

The function cvEllipse draws a simple or thick elliptic arc or fills an ellipse sector. The arc is clipped by ROI rectangle. The generalized Bresenham algorithm for conic section is used for simple elliptic arcs here, and piecewise-linear approximation is used for antialiased arcs and thick arcs. All the angles are given in degrees. The picture below explains the meaning of the parameters.

Parameters of Elliptic Arc

EllipseAA

Draws antialiased elliptic arc

void cvEllipseAA( CvArr* img, CvPoint center, CvSize axes, double angle,
                  double startAngle, double endAngle, double color, int scale=0 );

img
Image.
center
Center of the ellipse.
axes
Length of the ellipse axes.
angle
Rotation angle.
startAngle
Starting angle of the elliptic arc.
endAngle
Ending angle of the elliptic arc.
color
Ellipse color (RGB) or brightness (grayscale image).
scale
Specifies the number of fractional bits in the center coordinates and axes sizes.

The function cvEllipseAA draws an antialiased elliptic arc. The arc is clipped by ROI rectangle. The generalized Bresenham algorithm for conic section is used for simple elliptic arcs here, and piecewise-linear approximation is used for antialiased arcs and thick arcs. All the angles are in degrees.

FillPoly

Fills polygons interior

void cvFillPoly( CvArr* img, CvPoint** pts, int* npts, int contours, double color );

img
Image.
pts
Array of pointers to polygons.
npts
Array of polygon vertex counters.
contours
Number of contours that bind the filled region.
color
Polygon color (RGB) or brightness (grayscale image).

The function cvFillPoly fills an area bounded by several polygonal contours. The function fills complex areas, for example, areas with holes, contour self-intersection, etc.

FillConvexPoly

Fills convex polygon

void cvFillConvexPoly( CvArr* img, CvPoint* pts, int npts, double color );

img
Image.
pts
Array of pointers to a single polygon.
npts
Polygon vertex counter.
color
Polygon color (RGB) or brightness (grayscale image).

The function cvFillConvexPoly fills convex polygon interior. This function is much faster than the function cvFillPoly and fills not only the convex polygon but any monotonic polygon, that is, a polygon whose contour intersects every horizontal line (scan line) twice at the most.

PolyLine

Draws simple or thick polygons

void cvPolyLine( CvArr* img, CvPoint** pts, int* npts, int contours, int isClosed,
                 double color, int thickness=1, int connectivity=8 );

img
Image.
pts
Array of pointers to polylines.
npts
Array of polyline vertex counters.
contours
Number of polyline contours.
isClosed
Indicates whether the polylines must be drawn closed. If closed, the function draws the line from the last vertex of every contour to the first vertex.
color
Polygon color (RGB) or brightness (grayscale image).
thickness
Thickness of the polyline edges.
connectivity
The connectivity of polyline segments, 8 (by default) or 4.

The function cvPolyLine draws a set of simple or thick polylines.

PolyLineAA

Draws antialiased polygons

void cvPolyLineAA( CvArr* img, CvPoint** pts, int* npts, int contours,
                   int isClosed, int color, int scale =0);

img
Image.
pts
Array of pointers to polylines.
npts
Array of polyline vertex counters.
contours
Number of polyline contours.
isClosed
Indicates whether the polylines must be drawn closed. If closed, the function draws the line from the last vertex of every contour to the first vertex.
color
Polygon color (RGB) or brightness (grayscale image).
scale
Specifies number of fractional bits in the coordinates of polyline vertices.

The function cvPolyLineAA draws a set of antialiased polylines.

InitFont

Initializes font structure

void cvInitFont( CvFont* font, CvFontFace fontFace, float hscale,
                 float vscale, float italicScale, int thickness );

font
Pointer to the font structure initialized by the function.
fontFace
Font name identifier. Only the font CV_FONT_VECTOR0 is currently supported.
hscale
Horizontal scale. If equal to 1.0f, the characters have the original width depending on the font type. If equal to 0.5f, the characters are of half the original width.
vscale
Vertical scale. If equal to 1.0f, the characters have the original height depending on the font type. If equal to 0.5f, the characters are of half the original height.
italicScale
Approximate tangent of the character slope relative to the vertical line. Zero value means a non-italic font, 1.0f means ≈45° slope, etc. thickness Thickness of lines composing letters outlines. The function cvLine is used for drawing letters.

The function cvInitFont initializes the font structure that can be passed further into text drawing functions. Although only one font is supported, it is possible to get different font flavors by varying the scale parameters, slope, and thickness.

PutText

Draws text string

void cvPutText( CvArr* img, const char* text, CvPoint org, CvFont* font, int color );

img
Input image.
text
String to print.
org
Coordinates of the bottom-left corner of the first letter.
font
Pointer to the font structure.
color
Text color (RGB) or brightness (grayscale image).

The function cvPutText renders the text in the image with the specified font and color. The printed text is clipped by ROI rectangle. Symbols that do not belong to the specified font are replaced with the rectangle symbol.

GetTextSize

Retrieves width and height of text string

void cvGetTextSize( CvFont* font, const char* textString, CvSize* textSize, int* ymin );

font
Pointer to the font structure.
textString
Input string.
textSize
Resultant size of the text string. Height of the text does not include the height of character parts that are below the baseline.
ymin
Lowest y coordinate of the text relative to the baseline. Negative, if the text includes such characters as g, j, p, q, y, etc., and zero otherwise.

The function cvGetTextSize calculates the binding rectangle for the given text string when a specified font is used.

Gradients, Edges and Corners

Sobel

Calculates first, second, third or mixed image derivatives using extended Sobel operator

void cvSobel( const CvArr* I, CvArr* J, int dx, int dy, int apertureSize=3 );

I
Source image.
J
Destination image.
ox
Order of the derivative x .
oy
Order of the derivative y .
apertureSize
Size of the extended Sobel kernel, must be 1, 3, 5 or 7. In all cases except 1, apertureSize × apertureSize separable kernel will be used to calculate the derivative. For apertureSize=1 3x1 or 1x3 kernel is used (Gaussian smoothing is not done). There is also special value CV_SCHARR (=-1) that corresponds to 3x3 Scharr filter that may give more accurate results than 3x3 Sobel. Scharr aperture is: 
| -3 0  3|
|-10 0 10|
| -3 0  3|
for x-derivative or transposed for y-derivative.

The function cvSobel calculates the image derivative by convolving the image with the appropriate kernel:

J(x,y) = dox+oyI/dxox•dyoy |(x,y)
The Sobel operators combine Gaussian smoothing and differentiation so the result is more or less robust to the noise. Most often, the function is called with (ox=1, oy=0, apertureSize=3) or (ox=0, oy=1, apertureSize=3) to calculate first x- or y- image derivative. The first case corresponds to

  |-1  0  1|
  |-2  0  2|
  |-1  0  1|

kernel and the second one corresponds to

  |-1 -2 -1|
  | 0  0  0|
  | 1  2  1|
or
  | 1  2  1|
  | 0  0  0|
  |-1 -2 -1|
kernel, depending on the image origin (origin field of IplImage structure). No scaling is done, so the destination image usually has larger by absolute value numbers than the source image. To avoid overflow, the function requires 16-bit destination image if the source image is 8-bit. The result can be converted back to 8-bit using cvConvertScale or cvConvertScaleAbs functions. Besides 8-bit images the function can process 32-bit floating-point images. Both source and destination must be single-channel images of equal size or ROI size.

Laplace

Calculates Laplacian of the image

void cvLaplace( const CvArr* I, CvArr* J, int apertureSize=3 );

I
Source image.
J
Destination image.
apertureSize
Aperture parameter for Sobel operator (see cvSobel).

The function cvLaplace calculates Laplacian of the source image by summing second x- and y- derivatives calcualted using Sobel operator:

J(x,y) = d2I/dx2 + d2I/dy2

Specifying apertureSize=1 gives the fastest variant that is equal to convolving the image with the following kernel:

|0  1  0|
|1 -4  1|
|0  1  0|

As well as in cvSobel function, no scaling is done and the same combinations of input and output formats are supported.

Canny

Implements Canny algorithm for edge detection

void cvCanny( const CvArr* img, CvArr* edges, double threshold1,
              double threshold2, int apertureSize=3 );

img
Input image.
edges
Image to store the edges found by the function.
threshold1
The first threshold.
threshold2
The second threshold.
apertureSize
Aperture parameter for Sobel operator (see cvSobel).

The function cvCanny finds the edges on the input image img and marks them in the output image edges using the Canny algorithm. The smallest of threshold1 and threshold2 is used for edge linking, the largest - to find initial segments of strong edges.

PreCornerDetect

Calculates two constraint images for corner detection

void cvPreCornerDetect( const CvArr* img, CvArr* corners, int apertureSize=3 );

img
Input image.
corners
Image to store the corner candidates.
apertureSize
Aperture parameter for Sobel operator (see cvSobel).

The function cvPreCornerDetect finds the corners on the input image img and stores them in the corners image in accordance with Method 1 for corner detection desctibed in the guide.

CornerEigenValsAndVecs

Calculates eigenvalues and eigenvectors of image blocks for corner detection

void cvCornerEigenValsAndVecs( const CvArr* I, CvArr* eigenvv,
                               int blockSize, int apertureSize=3 );

I
Input image.
eigenvv
Image to store the results. It must be 6 times wider than the input image.
blockSize
Neighborhood size (see discussion).
apertureSize
Aperture parameter for Sobel operator (see cvSobel).

For every pixel the function cvCornerEigenValsAndVecs considers blockSize × blockSize neigborhood S(p). It calcualtes covariation matrix of derivatives over the neigborhood as:

    | sumS(p)(dI/dx)2   sumS(p)(dI/dx•dI/dy)|
M = |                                 |
    | sumS(p)(dI/dx•dI/dy)  sumS(p)(dI/dy)2 |

After that it finds eigenvectors and eigenvalues of the resultant matrix and stores them into destination image in form (λ1, λ2, x1, y1, x2, y2), where
λ1, λ2 - eigenvalues of M; not sorted
(x1, y1) - eigenvector corresponding to λ1
(x2, y2) - eigenvector corresponding to λ2

CornerMinEigenVal

Calculates minimal eigenvalue of image blocks for corner detection

void cvCornerMinEigenVal( const CvArr* img, CvArr* eigenvv, int blockSize, int apertureSize=3 );

img
Input image.
eigenvv
Image to store the minimal eigen values. Should have the same size as img
blockSize
Neighborhood size (see discussion of cvCornerEigenValsAndVecs).
apertureSize
Aperture parameter for Sobel operator (see cvSobel). format. In the case of floating-point input format this parameter is the number of the fixed float filter used for differencing.

The function cvCornerMinEigenVal is similar to cvCornerEigenValsAndVecs but it calculates and stores only the minimal eigen value of derivative covariation matrix for every pixel, i.e. min(λ1, λ2) in terms of the previous function.

FindCornerSubPix

Refines corner locations

void cvFindCornerSubPix( IplImage* I, CvPoint2D32f* corners,
                         int count, CvSize win, CvSize zeroZone,
                         CvTermCriteria criteria );

I
Input image.
corners
Initial coordinates of the input corners and refined coordinates on output.
count
Number of corners.
win
Half sizes of the search window. For example, if win=(5,5) then 5*2+1 × 5*2+1 = 11 × 11 search window is used.
zeroZone
Half size of the dead region in the middle of the search zone over which the summation in formulae below is not done. It is used sometimes to avoid possible singularities of the autocorrelation matrix. The value of (-1,-1) indicates that there is no such size.
criteria
Criteria for termination of the iterative process of corner refinement. That is, the process of corner position refinement stops either after certain number of iteration or when a required accuracy is achieved. The criteria may specify either of or both the maximum number of iteration and the required accuracy.

The function cvFindCornerSubPix iterates to find the sub-pixel accurate location of a corner, or "radial saddle point", as shown in on the picture below.

Sub-pixel accurate corner (radial saddle point) locator is based on the observation that any vector from q to p is orthogonal to the image gradient.

The core idea of this algorithm is based on the observation that every vector from the center q to a point p located within a neighborhood of q is orthogonal to the image gradient at p subject to image and measurement noise. Thus: 

εi=DIpiT•(q-pi)
where DIpi is the image gradient at the one of the points pi in a neighborhood of q . The value of q is to be found such that εi is minimized. A system of equations may be set up with εi' set to zero:

sumi(DIpi•DIpiT)•q - sumi(DIpi•DIpiT•pi) = 0

where the gradients are summed within a neighborhood ("search window") of q. Calling the first gradient term G and the second gradient term b gives:

q=G-1•b

The algorithm sets the center of the neighborhood window at this new center q and then iterates until the center keeps within a set threshold.

GoodFeaturesToTrack

Determines strong corners on image

void cvGoodFeaturesToTrack( IplImage* image, IplImage* eigImage, IplImage* tempImage,
                            CvPoint2D32f* corners, int* cornerCount,
                            double qualityLevel, double minDistance );

image
The source 8-bit or floating-point 32-bit, single-channel image.
eigImage
Temporary floating-point 32-bit image of the same size as image.
tempImage
Another temporary image of the same size and same format as eigImage.
corners
Output parameter. Detected corners.
cornerCount
Output parameter. Number of detected corners.
qualityLevel
Multiplier for the maxmin eigenvalue; specifies minimal accepted quality of image corners.
minDistance
Limit, specifying minimum possible distance between returned corners; Euclidian distance is used.

The function cvGoodFeaturesToTrack finds corners with big eigenvalues in the image. The function first calculates the minimal eigenvalue for every source image pixel using cvCornerMinEigenVal function and stores them in eigImage. Then it performs non-maxima suppression (only local maxima in 3x3 neighborhood remain). The next step is rejecting the corners with the minimal eigenvalue less than qualityLevel•max(eigImage(x,y)). Finally, the function ensures that all the corners found are distanced enough from one another by considering the corners (the most strongest corners are considered first) and checking that the distance between the newly considered feature and the features considered earlier is larger than minDistance. So, the function removes the features than are too close to the stronger features.

Sampling, Interpolation and Geometrical Transforms

InitLineIterator

Initializes line iterator

int cvInitLineIterator( const CvArr* img, CvPoint pt1, CvPoint pt2,
                        CvLineIterator* lineIterator, int connectivity=8 );

img
Image.
pt1
Starting the line point.
pt2
Ending the line point.
lineIterator
Pointer to the line iterator state structure.
connectivity
The scanned line connectivity, 4 or 8.

The function cvInitLineIterator initializes the line iterator and returns the number of pixels between two end points. Both points must be inside the image. After the iterator has been initialized, all the points on the raster line that connects the two ending points may be retrieved by successive calls of CV_NEXT_LINE_POINT point. The points on the line are calculated one by one using 4-connected or 8-connected Bresenham algorithm.

Example. Using line iterator to calculate pixel values along the color line

    CvScalar sum_line_pixels( IplImage* img, CvPoint pt1, CvPoint pt2 )
    {
        CvLineIterator iterator;
        int blue_sum = 0, green_sum = 0, red_sum = 0;
        int count = cvInitLineIterator( img, pt1, pt2, &iterator, 8 );

        for( int i = 0; i < count; i++ ){
            blue_sum += iterator.ptr[0];
            green_sum += iterator.ptr[1];
            red_sum += iterator.ptr[2];
            CV_NEXT_LINE_POINT(iterator);

            /* print the pixel coordinates: demonstrates how to calculate the coordinates */
            {
            int offset, x, y;
            /* assume that ROI is not set, otherwise need to take it into account. */
            offset = iterator.ptr - (uchar*)(img->imageData);
            y = offset/img->widthStep;
            x = (offset - y*img->widthStep)/(3*sizeof(uchar) /* size of pixel */);
            printf("(%d,%d)\n", x, y );
            }
        }
        return cvScalar( blue_sum, green_sum, red_sum );
    }

SampleLine

Reads raster line to buffer

int cvSampleLine( const CvArr* img, CvPoint pt1, CvPoint pt2,
                  void* buffer, int connectivity=8 );

img
Image.
pt1
Starting the line point.
pt2
Ending the line point.
buffer
Buffer to store the line points; must have enough size to store max( |pt2.x-pt1.x|+1, |pt2.y-pt1.y|+1 ) points in case of 8-connected line and |pt2.x-pt1.x|+|pt2.y-pt1.y|+1 in case of 4-connected line.
connectivity
The line connectivity, 4 or 8.

The function cvSampleLine implements a particular case of application of line iterators. The function reads all the image points lying on the line between pt1 and pt2, including the ending points, and stores them into the buffer.

GetRectSubPix

Retrieves pixel rectangle from image with sub-pixel accuracy

void cvGetRectSubPix( const CvArr* I, CvArr* J, CvPoint2D32f center );

I
Source image.
J
Extracted rectangle.
center
Floating point coordinates of the extracted rectangle center within the source image. The center must be inside the image.

The function cvGetRectSubPix extracts pixels from I:

J( x+width(J)/2, y+height(J)/2 )=I( x+center.x, y+center.y )

where the values of pixels at non-integer coordinates ( x+center.x, y+center.y ) are retrieved using bilinear interpolation. Every channel of multiple-channel images is processed independently. Whereas the rectangle center must be inside the image, the whole rectangle may be partially occluded. In this case, the replication border mode is used to get pixel values beyond the image boundaries.

GetQuadrangeSubPix

Retrieves pixel quadrangle from image with sub-pixel accuracy

void cvGetQuadrangeSubPix( const CvArr* I, CvArr* J, const CvArr* M,
                           int fillOutliers=0, CvScalar fillValue=cvScalarAll(0) );

I
Source image.
J
Extracted quadrangle.
M
The transformation 3 × 2 matrix [A|b] (see the discussion).
fillOutliers
The flag indicating whether to interpolate values of pixel taken from outside of the source image using replication mode (fillOutliers=0) or set them a fixed value (fillOutliers=1).
fillValue
The fixed value to set the outlier pixels to if fillOutliers=1.

The function cvGetQuadrangleSubPix extracts pixels from I at sub-pixel accuracy and stores them to J as follows:

J( x+width(J)/2, y+height(J)/2 )= I( A11x+A12y+b1, A21x+A22y+b2 ),

where A and b are taken from M
    | A11 A12  b1 |
M = |            |
    | A21 A22  b2 |

where the values of pixels at non-integer coordinates A•(x,y)T+b are retrieved using bilinear interpolation. Every channel of multiple-channel images is processed independently.

Example. Using cvGetQuadrangeSubPix for image rotation.

#include "cv.h"
#include "highgui.h"
#include "math.h"

int main( int argc, char** argv )
{
    IplImage* src;
    /* the first command line parameter must be image file name */
    if( argc==2 && (src = cvLoadImage(argv[1], -1))!=0)
    {
        IplImage* dst = cvCloneImage( src );
        int delta = 1;
        int angle = 0;

        cvNamedWindow( "src", 1 );
        cvShowImage( "src", src );

        for(;;)
        {
            float m[6];
            double factor = (cos(angle*CV_PI/180.) + 1.1)*3;
            CvMat M = cvMat( 2, 3, CV_32F, m );
            int w = src->width;
            int h = src->height;

            m[0] = (float)(factor*cos(-angle*2*CV_PI/180.));
            m[1] = (float)(factor*sin(-angle*2*CV_PI/180.));
            m[2] = w*0.5f;
            m[3] = -m[1];
            m[4] = m[0];
            m[5] = h*0.5f;

            cvGetQuadrangleSubPix( src, dst, &M, 1, cvScalarAll(0));

            cvNamedWindow( "dst", 1 );
            cvShowImage( "dst", dst );

            if( cvWaitKey(5) == 27 )
                break;

            angle = (angle + delta) % 360;
        }
    }
    return 0;
}

Resize

Resizes image

void cvResize( const CvArr* I, CvArr* J, int interpolation=CV_INTER_LINEAR );

I
Source image.
J
Destination image.
interpolation
Interpolation method:

The function cvResize resizes image I so that it fits exactly to J. If ROI is set, the function consideres the ROI as supported as usual. the source image using the specified structuring element B that determines the shape of a pixel neighborhood over which the minimum is taken:

C=erode(A,B): C(I)=min(K in BI)A(K)

The function supports the in-place mode when the source and destination pointers are the same. Erosion can be applied several times iterations parameter. Erosion on a color image means independent transformation of all the channels.

Morphological Operations

CreateStructuringElementEx

Creates structuring element

IplConvKernel* cvCreateStructuringElementEx( int nCols, int nRows, int anchorX, int anchorY,
                                             CvElementShape shape, int* values );

nCols
Number of columns in the structuring element.
nRows
Number of rows in the structuring element.
anchorX
Relative horizontal offset of the anchor point.
anchorY
Relative vertical offset of the anchor point.
shape
Shape of the structuring element; may have the following values:
values
Pointer to the structuring element data, a plane array, representing row-by-row scanning of the element matrix. Non-zero values indicate points that belong to the element. If the pointer is NULL , then all values are considered non-zero, that is, the element is of a rectangular shape. This parameter is considered only if the shape is CV_SHAPE_CUSTOM .

The function cv CreateStructuringElementEx allocates and fills the structure IplConvKernel , which can be used as a structuring element in the morphological operations.

ReleaseStructuringElement

Deletes structuring element

void cvReleaseStructuringElement( IplConvKernel** ppElement );

ppElement
Pointer to the deleted structuring element.

The function cv ReleaseStructuringElement releases the structure IplConvKernel that is no longer needed. If *ppElement is NULL , the function has no effect. The function returns created structuring element.

Erode

Erodes image by using arbitrary structuring element

void cvErode( const CvArr* A, CvArr* C, IplConvKernel* B=0, int iterations=1 );

A
Source image.
C
Destination image.
B
Structuring element used for erosion. If it is NULL, a 3×3 rectangular structuring element is used.
iterations
Number of times erosion is applied.

The function cvErode erodes the source image using the specified structuring element B that determines the shape of a pixel neighborhood over which the minimum is taken:

C=erode(A,B):  C(x,y)=min((x',y') in B(x,y))A(x',y')

The function supports the in-place mode when the source and destination pointers are the same. Erosion can be applied several times iterations parameter. Erosion on a color image means independent transformation of all the channels.

Dilate

Dilates image by using arbitrary structuring element

void cvDilate( const CvArr* A, CvArr* C, IplConvKernel* B=0, int iterations=1 );

A
Source image.
C
Destination image.
B
Structuring element used for erosion. If it is NULL, a 3×3 rectangular structuring element is used.
iterations
Number of times erosion is applied.

The function cvDilate dilates the source image using the specified structuring element B that determines the shape of a pixel neighborhood over which the maximum is taken:

C=dilate(A,B):  C(x,y)=max((x',y') in B(x,y))A(x',y')

The function supports the in-place mode when the source and destination pointers are the same. Dilation can be applied several times iterations parameter. Dilation on a color image means independent transformation of all the channels.

MorphologyEx

Performs advanced morphological transformations

void cvMorphologyEx( const CvArr* A, CvArr* C, CvArr* temp,
                     IplConvKernel* B, CvMorphOp op, int iterations );

A
Source image.
C
Destination image.
temp
Temporary image, required in some cases.
B
Structuring element.
op
Type of morphological operation (see the discussion).
iterations
Number of times erosion and dilation are applied.

The function cvMorphologyEx performs advanced morphological transformations using on erosion and dilation as basic operations.

Opening:
C=open(A,B)=dilate(erode(A,B),B),   if op=CV_MOP_OPEN

Closing:
C=close(A,B)=erode(dilate(A,B),B),  if op=CV_MOP_CLOSE

Morphological gradient:
C=morph_grad(A,B)=dilate(A,B)-erode(A,B),  if op=CV_MOP_GRADIENT

"Top hat":
C=tophat(A,B)=A-erode(A,B),   if op=CV_MOP_TOPHAT

"Black hat":
C=blackhat(A,B)=dilate(A,B)-A,   if op=CV_MOP_BLACKHAT

The temporary image temp is required if op=CV_MOP_GRADIENT or if A=C (inplace operation) and op=CV_MOP_TOPHAT or op=CV_MOP_BLACKHAT

Filters and Color Conversion

Smooth

Smooths the image in one of several ways

void cvSmooth( const CvArr* src, CvArr* dst,
               int smoothtype=CV_GAUSSIAN,
               int param1=3, int param2=0 );

src
The source image.
dst
The destination image.
smoothtype
Type of the smoothing:
param1
The first parameter of smoothing operation.
param2
The second parameter of smoothing operation. In case of simple scaled/non-scaled and Gaussian blur if param2 is zero, it is set to param1.

The function cvSmooth smooths image using one of several methods. Every of the methods has some features and restrictions listed below

Blur with no scaling works with single-channel images only and supports accumulation of 8-bit to 16-bit format (similar to cvSobel and cvLaplace) and 32-bit floating point to 32-bit floating-point format.

Simple blur and Gaussian blur support 1- or 3-channel, 8-bit and 32-bit floating point images. These two methods can process images in-place.

Median and bilateral filters work with 1- or 3-channel 8-bit images and can not process images in-place.

Integral

Calculates integral images

void cvIntegral( const CvArr* I, CvArr* S, CvArr* Sq=0, CvArr* T=0 );

I
The source image, w×h, single-channel, 8-bit, or floating-point (32f or 64f).
S
The sum image, w+1×h+1, single-channel, 32-bit integer or double precision floating-point (64f).
Sq
The square sum image, w+1×h+1, single-channel, double precision floating-point (64f).
T
The tilted sum image (sum of rotated by 45° image), w+1×h+1, single-channel, the same data type as sum.

The function cvIntegral calculates one or more integral images for the source image as following:

S(X,Y)=sumx<X,y<YI(x,y)

Sq(X,Y)=sumx<X,y<YI(x,y)2

T(X,Y)=sumy<Y,abs(x-X)<yI(x,y)

After that the images are calculated, they can be used to calculate sums of pixels over an arbitrary rectangles, for example:

sumx1<=x<x2,y1<=y<y2I(x,y)=S(x2,y2)-S(x1,y2)-S(x2,y1)+S(x1,x1)

It makes possible to do a fast blurring or fast block correlation with variable window size etc.

CvtColor

Converts image from one color space to another

void cvCvtColor( const CvArr* src, CvArr* dst, int code );

src
The source 8-bit image.
dst
The destination 8-bit image.
code
Color conversion operation that can be specifed using CV_<src_color_space>2<dst_color_space> constants (see below).

The function cvCvtColor converts input image from one color space to another. The function ignores colorModel and channelSeq fields of IplImage header, so the source image color space should be specified correctly (including order of the channels in case of RGB space, e.g. BGR means 24-bit format with B0 G0 R0 B1 G1 R1 ... layout, whereas RGB means 24-format with R0 G0 B0 R1 G1 B1 ... layout). The function can do the following transformations:

Threshold

Applies fixed-level threshold to array elements

void cvThreshold( const CvArr* src, CvArr* dst, double threshold,
                  double maxValue, int thresholdType );

src
Source array (single-channel, 8-bit of 32-bit floating point).
dst
Destination array; must be either the same type as src or 8-bit.
threshold
Threshold value.
maxValue
Maximum value to use with CV_THRESH_BINARY, CV_THRESH_BINARY_INV, and CV_THRESH_TRUNC thresholding types.
thresholdType
Thresholding type (see the discussion)

The function cvThreshold applies fixed-level thresholding to single-channel array. The function is typically used to get bi-level (binary) image out of grayscale image or for removing a noise, i.e. filtering out pixels with too small or too large values. There are several types of thresholding the function supports that are determined by thresholdType:

thresholdType=CV_THRESH_BINARY:
dst(x,y) = maxValue, if src(x,y)>threshold
           0, otherwise

thresholdType=CV_THRESH_BINARY_INV:
dst(x,y) = 0, if src(x,y)>threshold
           maxValue, otherwise

thresholdType=CV_THRESH_TRUNC:
dst(x,y) = threshold, if src(x,y)>threshold
           src(x,y), otherwise

thresholdType=CV_THRESH_TOZERO:
dst(x,y) = src(x,y), if (x,y)>threshold
           0, otherwise

thresholdType=CV_THRESH_TOZERO_INV:
dst(x,y) = 0, if src(x,y)>threshold
           src(x,y), otherwise

And this is the visual description of thresholding types:

AdaptiveThreshold

Applies adaptive threshold to array

void cvAdaptiveThreshold( const CvArr* src, CvArr* dst, double maxValue,
                          int adaptiveMethod, int thresholdType,
                          int blockSize, double param1 );

src
Source image.
dst
Destination image.
maxValue
Maximum value that is used with CV_THRESH_BINARY and CV_THRESH_BINARY_INV.
adaptiveMethod
Adaptive thresholding algorithm to use: CV_ADAPTIVE_THRESH_MEAN_C or CV_ADAPTIVE_THRESH_GAUSSIAN_C (see the discussion).
thresholdType
Thresholding type; must be one of
blockSize
The size of a pixel neighborhood that is used to calculate a threshold value for the pixel: 3, 5, 7, ...
param1
The method-dependent parameter. For the methods CV_ADAPTIVE_THRESH_MEAN_C and CV_ADAPTIVE_THRESH_GAUSSIAN_C it is a constant subtracted from mean or weighted mean (see the discussion), though it may be negative.

The function cvAdaptiveThreshold transforms grayscale image to binary image according to the formulae:

thresholdType=CV_THRESH_BINARY:
dst(x,y) = maxValue, if src(x,y)>T(x,y)
           0, otherwise

thresholdType=CV_THRESH_BINARY_INV:
dst(x,y) = 0, if src(x,y)>T(x,y)
           maxValue, otherwise

where TI is a threshold calculated individually for each pixel.

For the method CV_ADAPTIVE_THRESH_MEAN_C it is a mean of blockSize × blockSize pixel neighborhood, subtracted by param1.

For the method CV_ADAPTIVE_THRESH_GAUSSIAN_C it is a weighted sum (gaussian) of blockSize × blockSize pixel neighborhood, subtracted by param1.

LUT

Performs look-up table transformation on image

CvMat* cvLUT( const CvArr* A, CvArr* B, const CvArr* lut );

A
Source array of 8-bit elements.
B
Destination array of arbitrary depth and of the same number of channels as the source array.
lut
Look-up table of 256 elements; should be of the same depth as the destination array.

The function cvLUT fills the destination array with values of look-up table entries. Indices of the entries are taken from the source array. That is, the function processes each pixel as follows:

B(x,y)=lut[A(x,y)+Δ]
where Δ is 0 for 8-bit unsigned source image type and 128 for 8-bit signed source image type.

Pyramids and the Applications

PyrDown

Downsamples image

void cvPyrDown( const CvArr* src, CvArr* dst, int filter=CV_GAUSSIAN_5x5 );

src
The source image.
dst
The destination image, should have 2x smaller width and height than the source.
filter
Type of the filter used for convolution; only CV_GAUSSIAN_5x5 is currently supported.

The function cvPyrDown performs downsampling step of Gaussian pyramid decomposition. First it convolves source image with the specified filter and then downsamples the image by rejecting even rows and columns.

PyrUp

Upsamples image

void cvPyrUp( const CvArr* src, CvArr* dst, int filter=CV_GAUSSIAN_5x5 );

src
The source image.
dst
The destination image, should have 2x smaller width and height than the source.
filter
Type of the filter used for convolution; only CV_GAUSSIAN_5x5 is currently supported.

The function cvPyrUp performs up-sampling step of Gaussian pyramid decomposition. First it upsamples the source image by injecting even zero rows and columns and then convolves result with the specified filter multiplied by 4 for interpolation. So the destination image is four times larger than the source image.

PyrSegmentation

Implements image segmentation by pyramids

void cvPyrSegmentation( IplImage* src, IplImage* dst,
                        CvMemStorage* storage, CvSeq** comp,
                        int level, double threshold1, double threshold2 );

src
The source image.
dst
The destination image.
storage
Storage; stores the resulting sequence of connected components.
comp
Pointer to the output sequence of the segmented components.
level
Maximum level of the pyramid for the segmentation.
threshold1
Error threshold for establishing the links.
threshold2
Error threshold for the segments clustering.

The function cvPyrSegmentation implements image segmentation by pyramids. The pyramid builds up to the level level. The links between any pixel a on level i and its candidate father pixel b on the adjacent level are established if

p(c(a),c(b))<threshold1. After the connected components are defined, they are joined into several clusters. Any two segments A and B belong to the same cluster, if
p(c(A),c(B))<threshold2. The input image has only one channel, then
p(c¹,c²)=|c¹-c²|. If the input image has three channels (red, green and blue), then
p(c¹,c²)=0,3·(c¹r-c²r)+0,59·(c¹g-c²g)+0,11·(c¹b-c²b) . There may be more than one connected component per a cluster.
The images src and dst should be 8-bit single-channel or 3-channel images or equal size

Connected components

CvConnectedComp

Connected component

    typedef struct CvConnectedComp
    {
        double area; /* area of the segmented component */
        float value; /* gray scale value of the segmented component */
        CvRect rect; /* ROI of the segmented component */
    } CvConnectedComp;

FloodFill

Fills a connected component with given color

void cvFloodFill( CvArr* img, CvPoint seed, double newVal,
                  double lo=0, double up=0, CvConnectedComp* comp=0,
                  int flags=4, CvArr* mask=0 );
#define CV_FLOODFILL_FIXED_RANGE (1 << 16)
#define CV_FLOODFILL_MASK_ONLY   (1 << 17)

img
Input image, either 1-,3-channel 8-bit, or single-channel floating-point image. It is modified by the function unless CV_FLOODFILL_MASK_ONLY flag is set (see below).
seed
Coordinates of the seed point inside the image ROI.
newVal
New value of repainted domain pixels. For 8-bit color images it is a packed color (e.g. using CV_RGB macro).
lo
Maximal lower brightness/color difference between the currently observed pixel and one of its neighbor belong to the component or seed pixel to add the pixel to component. In case of 8-bit color images it is packed value.
up
Maximal upper brightness/color difference between the currently observed pixel and one of its neighbor belong to the component or seed pixel to add the pixel to component. In case of 8-bit color images it is packed value.
comp
Pointer to structure the function fills with the information about the repainted domain.
flags
The operation flags. Lower bits contain connectivity value, 4 (by default) or 8, used within the function. Connectivity determines which neighbors of a pixel are considered. Upper bits can be 0 or combination of the following flags:
  • CV_FLOODFILL_FIXED_RANGE - if set the difference between the current pixel and seed pixel is considered, otherwise difference between neighbor pixels is considered (the range is floating).
  • CV_FLOODFILL_MASK_ONLY - if set, the function does not fill the image (newVal is ignored), but the fills mask (that must be non-NULL in this case).
mask
Operation mask, should be singe-channel 8-bit image, 2 pixels wider and 2 pixels taller than img. If not NULL, the function uses and updates the mask, so user takes responsibility of initializing mask content. Floodfilling can't go across non-zero pixels in the mask, for example, an edge detector output can be used as a mask to stop filling at edges. Or it is possible to use the same mask in multiple calls to the function to make sure the filled area do not overlap.

The function cvFloodFill fills a connected component starting from the seed pixel where all pixels within the component have close to each other values (prior to filling). The pixel is considered to belong to the repainted domain if its value I(x,y) meets the following conditions (the particular cases are specifed after commas):

I(x',y')-lo<=I(x,y)<=I(x',y')+up, grayscale image + floating range
I(seed.x,seed.y)-lo<=I(x,y)<=I(seed.x,seed.y)+up, grayscale image + floating range

I(x',y')r-lor<=I(x,y)r<=I(x',y')r+upr and
I(x',y')g-log<=I(x,y)g<=I(x',y')g+upg and
I(x',y')b-lob<=I(x,y)b<=I(x',y')b+upb, color image + floating range

I(seed.x,seed.y)r-lor<=I(x,y)r<=I(seed.x,seed.y)r+upr and
I(seed.x,seed.y)g-log<=I(x,y)g<=I(seed.x,seed.y)g+upg and
I(seed.x,seed.y)b-lob<=I(x,y)b<=I(seed.x,seed.y)b+upb, color image + fixed range
where I(x',y') is value of one of pixel neighbors (to be added to the connected component in case of floating range, a pixel should have at least one neigbor with similar brightness)

FindContours

Finds contours in binary image

int cvFindContours( CvArr* img, CvMemStorage* storage, CvSeq** firstContour,
                    int headerSize=sizeof(CvContour), CvContourRetrievalMode mode=CV_RETR_LIST,
                    CvChainApproxMethod method=CV_CHAIN_APPROX_SIMPLE );

image
The source 8-bit single channel image. Non-zero pixels are treated as 1's, zero pixels remain 0's - that is image treated as binary. To get such a binary image from grayscale, one may use cvThreshold, cvAdaptiveThreshold or cvCanny. The function modifies the source image content.
storage
Container of the retrieved contours.
firstContour
Output parameter, will contain the pointer to the first outer contour.
headerSize
Size of the sequence header, >=sizeof(CvChain) if method=CV_CHAIN_CODE, and >=sizeof(CvContour) otherwise.
mode
Retrieval mode.
  • CV_RETR_EXTERNALretrives only the extreme outer contours
  • CV_RETR_LISTretrieves all the contours and puts them in the list
  • CV_RETR_CCOMPretrieves all the contours and organizes them into two-level hierarchy: top level are external boundaries of the components, second level are bounda boundaries of the holes
  • CV_RETR_TREEretrieves all the contours and reconstructs the full hierarchy of nested contours
method
Approximation method.
  • CV_CHAIN_CODEoutputs contours in the Freeman chain code. All other methods output polygons (sequences of vertices).
  • CV_CHAIN_APPROX_NONEtranslates all the points from the chain code into points;
  • CV_CHAIN_APPROX_SIMPLEcompresses horizontal, vertical, and diagonal segments, that is, the function leaves only their ending points;
  • CV_CHAIN_APPROX_TC89_L1,
    CV_CHAIN_APPROX_TC89_KCOS applies one of the flavors of Teh-Chin chain approximation algorithm.
  • CV_LINK_RUNS uses completely different (from the previous methods) algorithm - linking of horizontal segments of 1's. Only CV_RETR_LIST retrieval mode is allowed by the method.

The function cvFindContours retrieves contours from the binary image and returns the number of retrieved contours. The pointer firstContour is filled by the function. It will contain pointer to the first most outer contour or NULL if no contours is detected (if the image is completely black). Other contours may be reached from firstContour using h_next and v_next links. The sample in cvDrawContours discussion shows how to use contours for connected component detection. Contours can be also used for shape analysis and object recognition - see squares sample in CVPR 2001 tutorial course located at SourceForge site.

StartFindContours

Initializes contour scanning process

CvContourScanner cvStartFindContours( IplImage* img, CvMemStorage* storage,
                                      int headerSize, CvContourRetrievalMode mode,
                                      CvChainApproxMethod method );

image
The source 8-bit single channel binary image.
storage
Container of the retrieved contours.
headerSize
Size of the sequence header, >=sizeof(CvChain) if method=CV_CHAIN_CODE, and >=sizeof(CvContour) otherwise.
mode
Retrieval mode, has the same meaning as in cvFindContours.
method
Approximation method, the same as in cvFindContours except that CV_LINK_RUNS can not be used here.

The function cvStartFindContours initializes and returns pointer to the contour scanner. The scanner is used further in cvFindNextContour to retrieve the rest of contours.

FindNextContour

Finds next contour in the image

CvSeq* cvFindNextContour( CvContourScanner scanner );

scanner
Contour scanner initialized by the function cvStartFindContours .

The function cvFindNextContour locates and retrieves the next contour in the image and returns pointer to it. The function returns NULL, if there is no more contours.

SubstituteContour

Replaces retrieved contour

void cvSubstituteContour( CvContourScanner scanner, CvSeq* newContour );

scanner
Contour scanner initialized by the function cvStartFindContours .
newContour
Substituting contour.

The function cvSubstituteContour replaces the retrieved contour, that was returned from the preceding call of the function cvFindNextContour and stored inside the contour scanner state, with the user-specified contour. The contour is inserted into the resulting structure, list, two-level hierarchy, or tree, depending on the retrieval mode. If the parameter newContour=NULL, the retrieved contour is not included into the resulting structure, nor all of its children that might be added to this structure later.

EndFindContours

Finishes scanning process

CvSeq* cvEndFindContours( CvContourScanner* scanner );

scanner
Pointer to the contour scanner.

The function cvEndFindContours finishes the scanning process and returns the pointer to the first contour on the highest level.

DrawContours

Draws contour outlines or interiors in the image

void cvDrawContours( CvArr *image, CvSeq* contour,
                     double external_color, double hole_color,
                     int max_level, int thickness=1,
                     int connectivity=8 );

image
Image where the contours are to be drawn. Like in any other drawing function, the contours are clipped with the ROI.
contour
Pointer to the first contour.
externalColor
Color to draw external contours with.
holeColor
Color to draw holes with.
maxLevel
Maximal level for drawn contours. If 0, only contour is drawn. If 1, the contour and all contours after it on the same level are drawn. If 2, all contours after and all contours one level below the contours are drawn, etc. If the value is negative, the function does not draw the contours following after contour but draws child contours of contour up to abs(maxLevel)-1 level.
thickness
Thickness of lines the contours are drawn with. If it is negative (e.g. =CV_FILLED), the contour interiors are drawn.
connectivity
Connectivity of line segments of the contour outlines.

The function cvDrawContours draws contour outlines in the image if thickness>=0 or fills area bounded by the contours if thickness<0.

Example. Connected component detection via contour functions

#include "cv.h"
#include "highgui.h"

int main( int argc, char** argv )
{
    IplImage* src;
    // the first command line parameter must be file name of binary (black-n-white) image
    if( argc == 2 && (src=cvLoadImage(argv[1], 0))!= 0)
    {
        IplImage* dst = cvCreateImage( cvGetSize(src), 8, 3 );
        CvMemStorage* storage = cvCreateMemStorage(0);
        CvSeq* contour = 0;

        cvThreshold( src, src, 1, 255, CV_THRESH_BINARY );
        cvNamedWindow( "Source", 1 );
        cvShowImage( "Source", src );

        cvFindContours( src, storage, &contour, sizeof(CvSeq), CV_RETR_CCOMP, CV_CHAIN_APPROX_SIMPLE );
        cvZero( dst );

        for( ; contour != 0; contour = contour->h_next )
        {
            int color = CV_RGB( rand(), rand(), rand() );
            /* replace CV_FILLED with 1 to see the outlines */
            cvDrawContours( dst, contour, color, color, -1, CV_FILLED, 8 );
        }

        cvNamedWindow( "Components", 1 );
        cvShowImage( "Components", dst );
        cvWaitKey(0);
    }
}

Replace CV_FILLED with 1 in the sample below to see the contour outlines

Image and contour moments

Moments

Calculates all moments up to third order of a polygon or rasterized shape

void cvMoments( const CvArr* arr, CvMoments* moments, int isBinary=0 );

arr
Image (1-channel or 3-channel with COI set) or polygon (CvSeq of points of a vector of points).
moments
Pointer to returned moment state structure.
isBinary
(For images only) If the flag is non-zero, all the zero pixel values are treated as zeroes, all the others are treated as ones.

The function cvMoments calculates spatial and central moments up to the third order and writes them to moments. The moments may be used then to calculate gravity center of the shape, its area, main axises and various shape characeteristics including 7 Hu invariants.

GetSpatialMoment

Retrieves spatial moment from moment state structure

double cvGetSpatialMoment( CvMoments* moments, int j, int i );

moments
The moment state, calculated by cvMoments.
j
x-order of the retrieved moment, j >= 0.
i
y-order of the retrieved moment, i >= 0 and i + j <= 3.

The function cvGetSpatialMoment retrieves the spatial moment, which in case of image moments is defined as:

Mji=sumx,y(I(x,y)•xj•yi)

where I(x,y) is the intensity of the pixel (x, y).

GetCentralMoment

Retrieves central moment from moment state structure

double cvGetCentralMoment( CvMoments* moments, int j, int i );

moments
Pointer to the moment state structure.
j
x-order of the retrieved moment, j >= 0.
i
y-order of the retrieved moment, i >= 0 and i + j <= 3.

The functioncvGetCentralMoment retrieves the central moment, which in case of image moments is defined as:

μij=sumx,y(I(x,y)•(x-xc)j•(y-yc)i),

where xc=M10/M00, yc=M01/M00 - coordinates of the gravity center

GetNormalizedCentralMoment

Retrieves normalized central moment from moment state structure

double cvGetNormalizedCentralMoment( CvMoments* moments, int x_order, int y_order );

moments
Pointer to the moment state structure.
j
x-order of the retrieved moment, j >= 0.
i
y-order of the retrieved moment, i >= 0 and i + j <= 3.

The function cvGetNormalizedCentralMoment retrieves the normalized central moment, which in case of image moments is defined as:

ηij= μij/M00((i+j)/2+1)

GetHuMoments

Calculates seven Hu invariants

void cvGetHuMoments( CvMoments* moments, CvHuMoments* HuMoments );

moments
Pointer to the moment state structure.
HuMoments
Pointer to Hu moments structure.

The function cvGetHuMoments calculates seven Hu invariants that are defined as: 

 h1=η20+η02

 h2=(η20-η02)²+4η11²

 h3=(η30-3η12)²+ (3η21-η03)²

 h4=(η30+η12)²+ (η21+η03)²

 h5=(η30-3η12)(η30+η12)[(η30+η12)²-3(η21+η03)²]+(3η21-η03)(η21+η03)[3(η30+η12)²-(η21+η03)²]

 h6=(η20-η02)[(η30+η12)²- (η21+η03)²]+4η11(η30+η12)(η21+η03)

 h7=(3η21-η03)(η21+η03)[3(η30+η12)²-(η21+η03)²]-(η30-3η12)(η21+η03)[3(η30+η12)²-(η21+η03)²]

These values are proved to be invariants to the image scale, rotation, and reflection except the seventh one, whose sign is changed by reflection.

Special Image Transforms

HoughLines

Finds lines in binary image using Hough transform

CvSeq* cvHoughLines2( CvArr* image, void* lineStorage, int method,
                      double dRho, double dTheta, int threshold,
                      double param1=0, double param2 );

image
Source 8-bit single-channel (binary) image. It may be modified by the function.
lineStorage
The storage for the lines detected. It can be a memory storage (in this case a sequence of lines is created in the storage and returned by the function) or single row/single column matrix (CvMat*) of a particular type (see below) where the lines' parameters are written. The matrix header is modified by the function so its cols/rows contains a number of lines detected (that is a matrix is truncated to fit exactly the detected lines, though no data is deallocated - only the header is modified). In the latter case if the actual number of lines exceeds the matrix size, the maximum possible number of lines is returned (the lines are not sorted by length, confidence or whatever criteria).
method
The Hough transform variant, one of:
  • CV_HOUGH_STANDARD - classical or standard Hough transform. Every line is represented by two floating-point numbers (ρ, θ), where ρ is a distance between (0,0) point and the line, and θ is the angle between x-axis and the normal to the line. Thus, the matrix must be (the created sequence will be) of CV_32FC2 type.
  • CV_HOUGH_PROBABILISTIC - probabilistic Hough transform (more efficient in case if picture contains a few long linear segments). It returns line segments rather than the whole lines. Every segment is represented by starting and ending points, and the matrix must be (the created sequence will be) of CV_32SC4 type.
  • CV_HOUGH_MULTI_SCALE - multi-scale variant of classical Hough transform. The lines are encoded the same way as in CV_HOUGH_CLASSICAL.
dRho
Distance resolution in pixel-related units.
dTheta
Angle resolution measured in radians.
threshold
Threshold parameter. A line is returned by the function if the corresponding accumulator value is greater than threshold.
param1
The first method-dependent parameter:
  • For classical Hough transform it is not used (0).
  • For probabilistic Hough transform it is the minimum line length.
  • For multi-scale Hough transform it is divisor for distance resolution dRho. (The coarse distance resolution will be dRho and the accurate resolution will be (dRho / param1)).
param2
The second method-dependent parameter:
  • For classical Hough transform it is not used (0).
  • For probabilistic Hough transform it is the maximum gap between line segments lieing on the same line to treat them as the single line segment (i.e. to join them).
  • For multi-scale Hough transform it is divisor for angle resolution dTheta. (The coarse angle resolution will be dTheta and the accurate resolution will be (dTheta / param2)).

The function cvHoughLines2 implements a few variants of Hough transform for line detection.

Example. Detecting lines with Hough transform.

/* This is a standalone program. Pass an image name as a first parameter of the program.
   Switch between standard and probabilistic Hough transform by changing "#if 1" to "#if 0" and back */
#include <cv.h>
#include <highgui.h>
#include <math.h>

int main(int argc, char** argv)
{
    IplImage* src;
    if( argc == 2 && (src=cvLoadImage(argv[1], 0))!= 0)
    {
        IplImage* dst = cvCreateImage( cvGetSize(src), 8, 1 );
        IplImage* color_dst = cvCreateImage( cvGetSize(src), 8, 3 );
        CvMemStorage* storage = cvCreateMemStorage(0);
        CvSeq* lines = 0;
        int i;
        cvCanny( src, dst, 50, 200, 3 );
        cvCvtColor( dst, color_dst, CV_GRAY2BGR );
#if 1
        lines = cvHoughLines2( dst, storage, CV_HOUGH_CLASSICAL, 1, CV_PI/180, 150, 0, 0 );

        for( i = 0; i < lines->total; i++ )
        {
            float* line = (float*)cvGetSeqElem(lines,i);
            float rho = line[0];
            float theta = line[1];
            CvPoint pt1, pt2;
            double a = cos(theta), b = sin(theta);
            if( fabs(a) < 0.001 )
            {
                pt1.x = pt2.x = cvRound(rho);
                pt1.y = 0;
                pt2.y = color_dst->height;
            }
            else if( fabs(b) < 0.001 )
            {
                pt1.y = pt2.y = cvRound(rho);
                pt1.x = 0;
                pt2.x = color_dst->width;
            }
            else
            {
                pt1.x = 0;
                pt1.y = cvRound(rho/b);
                pt2.x = cvRound(rho/a);
                pt2.y = 0;
            }
            cvLine( color_dst, pt1, pt2, CV_RGB(255,0,0), 3, 8 );
        }
#else
        lines = cvHoughLines2( dst, storage, CV_HOUGH_PROBABILISTIC, 1, CV_PI/180, 80, 30, 10 );
        for( i = 0; i < lines->total; i++ )
        {
            CvPoint* line = (CvPoint*)cvGetSeqElem(lines,i);
            cvLine( color_dst, line[0], line[1], CV_RGB(255,0,0), 3, 8 );
        }
#endif
        cvNamedWindow( "Source", 1 );
        cvShowImage( "Source", src );

        cvNamedWindow( "Hough", 1 );
        cvShowImage( "Hough", color_dst );

        cvWaitKey(0);
    }
}

This is the sample picture the function parameters have been tuned for:

And this is the output of the above program in case of probabilistic Hough transform ("#if 0" case):

DistTransform

Calculates distance to closest zero pixel for all non-zero pixels of source image

void cvDistTransform( const CvArr* src, CvArr* dst, CvDisType disType=CV_DIST_L2,
                      int maskSize=3, float* mask=0 );

src
Source 8-bit single-channel (binary) image.
dst
Output image with calculated distances (32-bit floating-point, single-channel).
disType
Type of distance; can be CV_DIST_L1, CV_DIST_L2, CV_DIST_C or CV_DIST_USER.
maskSize
Size of distance transform mask; can be 3 or 5. In case if CV_DIST_L1 or CV_DIST_C the parameter is forced to 3, because 5×5 mask gives the same result as 3×3 in this case yet it is slower.
mask
User-defined mask in case of user-defined distance, it consists of 2 numbers (horizontal/vertical shift cost, diagonal shift cost) in case of 3×3 mask and 3 numbers (horizontal/vertical shift cost, diagonal shift cost, knight's move cost) in case of 5×5 mask.

The function cvDistTransform calculates the approximated distance from every binary image pixel to the nearest zero pixel. For zero pixels the function sets the zero distance, for others it finds the shortest path consisting of basic shifts: horizontal, vertical, diagonal or knight's move (the latest is available for 5×5 mask). The overal distance is calculated as a sum of these basic distances. Because the distance function should be symmetric, all the horizontal and vertical shifts must have the same cost (that is denoted as a), all the diagonal shifts must have the same cost (denoted b), and all knight's moves' must have the same cost (denoted c). For CV_DIST_C and CV_DIST_L1 types the distance is calculated precisely, whereas for CV_DIST_L2 (Euclidian distance) the distance can be calculated only with some relative error (5×5 mask gives more accurate results), OpenCV uses the values suggested in [Borgefors86]:

CV_DIST_C (3×3):
a=1, b=1

CV_DIST_L1 (3×3):
a=1, b=2

CV_DIST_L2 (3×3):
a=0.955, b=1.3693

CV_DIST_L2 (5×5):
a=1, b=1.4, c=2.1969

And below are samples of distance field (black (0) pixel is in the middle of white square) in case of user-defined distance:

User-defined 3×3 mask (a=1, b=1.5)

4.543.533.544.5
432.522.534
3.52.51.511.52.53.5
3210123
3.52.51.511.52.53.5
432.522.534
4.543.533.544.5

User-defined 5×5 mask (a=1, b=1.5, c=2)

4.53.53333.54.5
3.5322233.5
321.511.523
3210123
321.511.523
3.5322233.5
43.53333.54

Typically, for fast coarse distance estimation CV_DIST_L2, 3×3 mask is used, and for more accurate distance estimation CV_DIST_L2, 5×5 mask is used.

[Borgefors86] Gunilla Borgefors, "Distance Transformations in Digital Images". Computer Vision, Graphics and Image Processing 34, 344-371 (1986).

Histogram Functions

CvHistogram

Muti-dimensional histogram

    typedef struct CvHistogram
    {
        int header_size; /* header's size */
        CvHistType type; /* type of histogram */
        int flags; /* histogram's flags */
        int c_dims; /* histogram's dimension */
        int dims[CV_HIST_MAX_DIM]; /* every dimension size */
        int mdims[CV_HIST_MAX_DIM]; /* coefficients for fast access to element */
        /* &m[a,b,c] = m + a*mdims[0] + b*mdims[1] + c*mdims[2] */
        float* thresh[CV_HIST_MAX_DIM]; /* bin boundaries arrays for every dimension */
        float* array; /* all the histogram data, expanded into the single row */
    struct CvNode* root; /* root of balanced tree storing histogram bins */
        CvSet* set; /* pointer to memory storage (for the balanced tree) */
        int* chdims[CV_HIST_MAX_DIM]; /* cache data for fast calculating */
    } CvHistogram;

CreateHist

Creates histogram

CvHistogram* cvCreateHist( int cDims, int* dims, int type,
                           float** ranges=0, int uniform=1 );

cDims
Number of histogram dimensions.
dims
Array of histogram dimension sizes.
type
Histogram representation format: CV_HIST_ARRAY means that histogram data is represented as an multi-dimensional dense array CvMatND; CV_HIST_TREE means that histogram data is represented as a multi-dimensional sparse array CvSparseMat.
ranges
Array of ranges for histogram bins. Its meaning depends on the uniform parameter value. The ranges are used for when histogram is calculated or backprojected to determine, which histogram bin corresponds to which value/tuple of values from the input image[s].
uniform
Uniformity flag; if not 0, the histogram has evenly spaced bins and for every 0<=i<cDims ranges[i] is array of two numbers: lower and upper boundaries for the i-th histogram dimension. The whole range [lower,upper] is split then into dims[i] equal parts to determine i-th input tuple value ranges for every histogram bin. And if uniform=0, then i-th element of ranges array contains dims[i]+1 elements: lower0, upper0, lower1, upper1 == lower2, ..., upperdims[i]-1, where lowerj and upperj are lower and upper boundaries of i-th input tuple value for j-th bin, respectively. In either case, the input values that are beyond the specified range for a histogram bin, are not counted by cvCalcHist and filled with 0 by cvCalcBackProject.

The function cvCreateHist creates a histogram of the specified size and returns the pointer to the created histogram. If the array ranges is 0, the histogram bin ranges must be specified later via the function cvSetHistBinRanges, though cvCalcHist and cvCalcBackProject may process 8-bit images without setting bin ranges, they assume equally spaced in 0..255 bins.

SetHistBinRanges

Sets bounds of histogram bins

void cvSetHistBinRanges( CvHistogram* hist, float** ranges, int uniform=1 );

hist
Histogram.
ranges
Array of bin ranges arrays, see cvCreateHist.
uniform
Uniformity flag, see cvCreateHist.

The function cvSetHistBinRanges is a stand-alone function for setting bin ranges in the histogram. For more detailed description of the parameters ranges and uniform see cvCalcHist function, that can initialize the ranges as well. Ranges for histogram bins must be set before the histogram is calculated or backproject of the histogram is calculated.

ReleaseHist

Releases histogram

void cvReleaseHist( CvHistogram** hist );

hist
Double pointer to the released histogram.

The function cvReleaseHist releases the histogram (header and the data). The pointer to histogram is cleared by the function. If *hist pointer is already NULL, the function does nothing.

ClearHist

Clears histogram

void cvClearHist( CvHistogram* hist );

hist
Histogram.

The function cvClearHist sets all histogram bins to 0 in case of dense histogram and removes all histogram bins in case of sparse array.

MakeHistHeaderForArray

Makes a histogram out of array

void cvMakeHistHeaderForArray( int cDims, int* dims, CvHistogram* hist,
                               float* data, float** ranges=0, int uniform=1 );

cDims
Number of histogram dimensions.
dims
Array of histogram dimension sizes.
hist
The histogram header initialized by the function.
data
Array that will be used to store histogram bins.
ranges
Histogram bin ranges, see cvCreateHist.
uniform
Uniformity flag, see cvCreateHist.

The function cvMakeHistHeaderForArray initializes the histogram, which header and bins are allocated by user. No cvReleaseHist need to be called afterwards. The histogram will be dense, sparse histogram can not be initialized this way.

QueryHistValue_1D

Queries value of histogram bin

#define cvQueryHistValue_1D( hist, idx0 ) \
    cvGetReal1D( (hist)->bins, (idx0) )
#define cvQueryHistValue_2D( hist, idx0, idx1 ) \
    cvGetReal2D( (hist)->bins, (idx0), (idx1) )
#define cvQueryHistValue_3D( hist, idx0, idx1, idx2 ) \
    cvGetReal3D( (hist)->bins, (idx0), (idx1), (idx2) )
#define cvQueryHistValue_nD( hist, idx ) \
    cvGetRealND( (hist)->bins, (idx) )

hist
Histogram.
idx0, idx1, idx2, idx3
Indices of the bin.
idx
Array of indices

The macros cvQueryHistValue_*D return the value of the specified bin of 1D, 2D, 3D or nD histogram. In case of sparse histogram the function returns 0, if the bin is not present in the histogram, and no new bin is created.

GetHistValue_1D

Returns pointer to histogram bin

#define cvGetHistValue_1D( hist, idx0 ) \
    ((float*)(cvPtr1D( (hist)->bins, (idx0), 0 ))
#define cvGetHistValue_2D( hist, idx0, idx1 ) \
    ((float*)(cvPtr2D( (hist)->bins, (idx0), (idx1), 0 ))
#define cvGetHistValue_3D( hist, idx0, idx1, idx2 ) \
    ((float*)(cvPtr3D( (hist)->bins, (idx0), (idx1), (idx2), 0 ))
#define cvGetHistValue_nD( hist, idx ) \
    ((float*)(cvPtrND( (hist)->bins, (idx), 0 ))

hist
Histogram.
idx0, idx1, idx2, idx3
Indices of the bin.
idx
Array of indices

The macros cvGetHistValue_*D return pointer to the specified bin of 1D, 2D, 3D or nD histogram. In case of sparse histogram the function creates a new bins and fills it with 0, if it does not exists.

GetMinMaxHistValue

Finds minimum and maximum histogram bins

void cvGetMinMaxHistValue( const CvHistogram* hist,
                           float* minVal, float* maxVal,
                           int* minIdx =0, int* maxIdx =0);

hist
Histogram.
minVal
Pointer to the minimum value of the histogram; can be NULL.
maxVal
Pointer to the maximum value of the histogram; can be NULL.
minIdx
Pointer to the array of coordinates for minimum. If not NULL, must have hist->c_dims elements to store the coordinates.
maxIdx
Pointer to the array of coordinates for maximum. If not NULL, must have hist->c_dims elements to store the coordinates.

The function cvGetMinMaxHistValue finds the minimum and maximum histogram bins and their positions. In case of several maximums or minimums the earliest in lexicographical order extrema locations are returned.

NormalizeHist

Normalizes histogram

void cvNormalizeHist( CvHistogram* hist, double factor );

hist
Pointer to the histogram.
factor
Normalization factor.

The function cvNormalizeHist normalizes the histogram bins by scaling them, such that the sum of the bins becomes equal to factor.

ThreshHist

Thresholds histogram

void cvThreshHist( CvHistogram* hist, double thresh );

hist
Pointer to the histogram.
thresh
Threshold level.

The function cvThreshHist clears histogram bins that are below the specified level.

CompareHist

Compares two dense histograms

double cvCompareHist( const CvHistogram* H1, const CvHistogram* H2,
                      CvCompareMethod method );

H1
The first dense histogram.
H2
The second dense histogram.
method
Comparison method, one of:
  • CV_COMP_CORREL;
  • CV_COMP_CHISQR;
  • CV_COMP_INTERSECT.

The function cvCompareHist compares two histograms using specified method and returns the comparison result. It processes as following:

Correlation (method=CV_COMP_CORREL):
d(H1,H2)=sumI(H'1(I)•H'2(I))/sqrt(sumI[H'1(I)2]•sumI[H'2(I)2])
where
H'k(I)=Hk(I)-1/N•sumJHk(J) (N=number of histogram bins)

Chi-Square (method=CV_COMP_CHISQR):
d(H1,H2)=sumI[(H1(I)-H2(I))/(H1(I)+H2(I))]

Intersection (method=CV_COMP_INTERSECT):
d(H1,H2)=sumImax(H1(I),H2(I))

Note, that the function can operate on dense histogram only. To compare sparse histogram or more general sparse configurations of weighted points, consider cvCalcEMD function.

CopyHist

Copies histogram

void cvCopyHist( CvHistogram* src, CvHistogram** dst );

src
Source histogram.
dst
Pointer to destination histogram.

The function cvCopyHist makes a copy of the histogram. If the second histogram pointer *dst is NULL, a new histogram of the same size as src is created. Otherwise, both histograms must have equal types and sizes. Then the function copies the source histogram bins values to destination histogram and sets the same as src's value ranges.

CalcHist

Calculates histogram of image(s)

void cvCalcHist( IplImage** img, CvHistogram* hist,
                 int doNotClear=0, const CvArr* mask=0 );

img
Source images (though, you may pass CvMat** as well).
hist
Pointer to the histogram.
doNotClear
Clear flag, if it is non-zero, the histogram is not cleared before calculation. It may be useful for iterative histogram update.
mask
The operation mask, determines what pixels of the source images are counted.

The function cvCalcHist calculates the histogram of one or more single-channel images. The elements of a tuple that is used to increment a histogram bin are taken at the same location from the corresponding input images.

Sample. Calculating and displaying 2D Hue-Saturation histogram of a color image

#include <cv.h>
#include <highgui.h>

int main( int argc, char** argv )
{
    IplImage* src;
    if( argc == 2 && (src=cvLoadImage(argv[1], 1))!= 0)
    {
        IplImage* h_plane = cvCreateImage( cvGetSize(src), 8, 1 );
        IplImage* s_plane = cvCreateImage( cvGetSize(src), 8, 1 );
        IplImage* v_plane = cvCreateImage( cvGetSize(src), 8, 1 );
        IplImage* planes[] = { h_plane, s_plane };
        IplImage* hsv = cvCreateImage( cvGetSize(src), 8, 3 );
        int h_bins = 30, s_bins = 32;
        int hist_size[] = {h_bins, s_bins};
        float h_ranges[] = { 0, 180 }; /* hue varies from 0 (~0°red) to 180 (~360°red again) */
        float s_ranges[] = { 0, 255 }; /* saturation varies from 0 (black-gray-white) to 255 (pure spectrum color) */
        float* ranges[] = { h_ranges, s_ranges };
        int scale = 10;
        IplImage* hist_img = cvCreateImage( cvSize(h_bins*scale,s_bins*scale), 8, 3 );
        CvHistogram* hist;
        float max_value = 0;
        int h, s;

        cvCvtColor( src, hsv, CV_BGR2HSV );
        cvCvtPixToPlane( hsv, h_plane, s_plane, v_plane, 0 );
        hist = cvCreateHist( 2, hist_size, CV_HIST_ARRAY, ranges, 1 );
        cvCalcHist( planes, hist, 0, 0 );
        cvGetMinMaxHistValue( hist, 0, &max_value, 0, 0 );
        cvZero( hist_img );

        for( h = 0; h < h_bins; h++ )
        {
            for( s = 0; s < s_bins; s++ )
            {
                float bin_val = cvQueryHistValue_2D( hist, h, s );
                int intensity = cvRound(bin_val*255/max_value);
                cvRectangle( hist_img, cvPoint( h*scale, s*scale ),
                             cvPoint( (h+1)*scale - 1, (s+1)*scale - 1),
                             CV_RGB(intensity,intensity,intensity), /* graw a grayscale histogram.
                                                                       if you have idea how to do it
                                                                       nicer let us know */
                             CV_FILLED );
            }
        }

        cvNamedWindow( "Source", 1 );
        cvShowImage( "Source", src );

        cvNamedWindow( "H-S Histogram", 1 );
        cvShowImage( "H-S Histogram", hist_img );

        cvWaitKey(0);
    }
}

CalcBackProject

Calculates back projection

void cvCalcBackProject( IplImage** img, CvArr* backProject, const CvHistogram* hist );

img
Source images (though you may pass CvMat** as well).
backProject
Destination back projection image of the same type as the source images.
hist
Histogram.

The function cvCalcBackProject calculates the back project of the histogram. For each tuple of pixels at the same position of all input single-channel images the function puts the value of the histogram bin, corresponding to the tuple, to the destination image. In terms of statistics, the value of each output image pixel is probability of the observed tuple given the distribution (histogram). For example, to find a red object in the picture, one may do the following:

  1. Calculate a hue histogram for the red object assuming the image contains only this object. The histogram is likely to have a strong maximum, corresponding to red color.
  2. Calculate back projection of a hue plane of input image where the object is searched, using the histogram. Threshold the image.
  3. Find connected components in the resulting picture and choose the right component using some additional criteria, for example, the largest connected component.
That is the approximate algorithm of Camshift color object tracker, except for the last step, where CAMSHIFT algorithm is used to locate the object on the back projection given the previous object position.

CalcBackProjectPatch

Locates a template within image by histogram comparison

void cvCalcBackProjectPatch( IplImage** img, CvArr* dst,
                             CvSize patchSize, CvHistogram* hist,
                             int method, float normFactor );

img
Source images (though, you may pass CvMat** as well)
dst
Destination image.
patchSize
Size of patch slid though the source image.
hist
Histogram
method
Compasion method, passed to cvCompareHist (see description of that function).
normFactor
Normalization factor for histograms, will affect normalization scale of destination image, pass 1. if unsure.

The function cvCalcBackProjectPatch calculates back projection by comparing histograms of the source image patches with the given histogram. Taking measurement results from some image at each location over ROI creates an array img. These results might be one or more of hue, x derivative, y derivative, Laplacian filter, oriented Gabor filter, etc. Each measurement output is collected into its own separate image. The img image array is a collection of these measurement images. A multi-dimensional histogram hist is constructed by sampling from the img image array. The final histogram is normalized. The hist histogram has as many dimensions as the number of elements in img array.

Each new image is measured and then converted into an img image array over a chosen ROI. Histograms are taken from this img image in an area covered by a "patch" with anchor at center as shown in the picture below. The histogram is normalized using the parameter norm_factor so that it may be compared with hist. The calculated histogram is compared to the model histogram; hist uses the function cvCompareHist with the comparison method=method). The resulting output is placed at the location corresponding to the patch anchor in the probability image dst. This process is repeated as the patch is slid over the ROI. Iterative histogram update by subtracting trailing pixels covered by the patch and adding newly covered pixels to the histogram can save a lot of operations, though it is not implemented yet.

Back Project Calculation by Patches

CalcProbDensity

Divides one histogram by another

void  cvCalcProbDensity( const CvHistogram* hist1, const CvHistogram* hist2,
                         CvHistogram* histDens, double scale=255 );

hist1
first histogram (divisor).
hist2
second histogram.
histDens
destination histogram.

The function cvCalcProbDensity calculates the object probability density from the two histograms as:

histDens(I)=0  if hist1(I)==0
            scale  if hist1(I)!=0 && hist2(I)>hist1(I)
            hist2(I)*scale/hist1(I) if hist1(I)!=0 && hist2(I)<=hist1(I)

So the destination histogram bins are within [0,scale).

CalcEMD2

Computes "minimal work" distance between two weighted point configurations

float cvCalcEMD2( const CvArr* signature1, const CvArr* signature2, CvDisType distType,
                  float (*distFunc)(const float* f1, const float* f2, void* userParam ),
                  const CvArr* costMatrix, CvArr* flow,
                  float* lowerBound, void* userParam );

signature1
First signature, size1×dims+1 floating-point matrix. Each row stores the point weight followed by the point coordinates. The matrix is allowed to have a single column (weights only) if the user-defined cost matrix is used.
signature2
Second signature of the same format as signature1, though the number of rows may be different. The total weights may be different, in this case an extra "dummy" point is added to either signature1 or signature2.
distType
Metrics used; CV_DIST_L1, CV_DIST_L2, and CV_DIST_C stand for one of the standard metrics; CV_DIST_USER means that a user-defined function distFunc or pre-calculated costMatrix is used.
distFunc
The user-defined distance function. It takes coordinates of two points and returns the distance between the points.
costMatrix
The user-defined size1×size2 cost matrix. At least one of costMatrix and distFunc must be NULL. Also, if a cost matrix is used, lower boundary (see below) can not be calculated, because it needs a metric function.
flow
The resultant size1×size2 flow matrix: flowij is a flow from i-th point of signature1 to j-th point of signature2
lowerBound
Optional output parameter: lower boundary of distance between the two signatures that is a distance between mass centers. The lower boundary may not be calculated if the user-defined cost matrix is used, the total weights of point configurations are not equal, or there is the signatures consist of weights only (i.e. the matrices have a single column).
userParam
Pointer to optional data that is passed into the user-defined distance function.

The function cvCalcEMD2 computes earth mover distance and/or a lower boundary of the distance between the two weighted point configurations. One of the application desctibed in [RubnerSept98] is multi-dimensional histogram comparison for image retrieval. EMD is a transportation problem that is solved using some modification of simplex algorithm, thus the complexity is exponential in the worst case, though, it is much faster in average. In case of real metric the lower boundary can be calculated even faster (using linear-time algorithm) and it can be used to determine roughly whether the two signatures are far enough so that they cannot relate to the same object.

[RubnerSept98] Y. Rubner. C. Tomasi, L.J. Guibas. The Earth Mover's Distance as a Metric for Image Retrieval. Technical Report STAN-CS-TN-98-86, Department of Computer Science, Stanford University, September 1998.

Utility Functions

MatchTemplate

Compares template against overlapped image regions

void cvMatchTemplate( const CvArr* I, const CvArr* T,
                      CvArr* result, int method );

I
Image where the search is running. It should be single-chanel, 8-bit or 32-bit floating-point.
T
Searched template; must be not greater than the source image and the same data type as the image.
R
Image of comparison results; single-channel 32-bit floating-point. If I is W×H and T is w×h then R must be W-w+1×H-h+1.
method
Specifies the way the template must be compared with image regions (see below).

The function cvMatchTemplate is similiar to cvCalcBackProjectPatch. It slids through I, compares w×h patches against T using the specified method and stores the comparison results to result. Here are the formular for the different comparison methods one may use (the summation is done over template and/or the image patch: x'=0..w-1, y'=0..h-1):

method=CV_TM_SQDIFF:
R(x,y)=sumx',y'[T(x',y')-I(x+x',y+y')]2

method=CV_TM_SQDIFF_NORMED:
R(x,y)=sumx',y'[T(x',y')-I(x+x',y+y')]2/sqrt[sumx',y'T(x',y')2•sumx',y'I(x+x',y+y')2]

method=CV_TM_CCORR:
R(x,y)=sumx',y'[T(x',y')•I(x+x',y+y')]

method=CV_TM_CCORR_NORMED:
R(x,y)=sumx',y'[T(x',y')•I(x+x',y+y')]/sqrt[sumx',y'T(x',y')2•sumx',y'I(x+x',y+y')2]

method=CV_TM_CCOEFF:
R(x,y)=sumx',y'[T'(x',y')•I'(x+x',y+y')],

where T'(x',y')=T(x',y') - 1/(w•h)•sumx",y"T(x",y") (mean template brightness=>0)
      I'(x+x',y+y')=I(x+x',y+y') - 1/(w•h)•sumx",y"I(x+x",y+y") (mean patch brightness=>0)

method=CV_TM_CCOEFF_NORMED:
R(x,y)=sumx',y'[T'(x',y')•I'(x+x',y+y')]/sqrt[sumx',y'T'(x',y')2•sumx',y'I'(x+x',y+y')2]
After the function finishes comparison, the best matches can be found as global minimums (CV_TM_SQDIFF*) or maximums (CV_TM_CCORR* and CV_TM_CCOEFF*) using cvMinMaxLoc function.