I found this square detection code online and I'm trying to understand it, does anyone understand what the line at the end does? it states: "gray = gray0 >= (l+1)*255/N;"
Mat pyr, timg, gray0(image.size(), CV_8U), gray;
// down-scale and upscale the image to filter out the noise
pyrDown(image, pyr, Size(image.cols/2, image.rows/2));
pyrUp(pyr, timg, image.size());
vector<vector<Point> > contours;
// find squares in every color plane of the image
for( int c = 0; c < 3; c++ )
{
int ch[] = {c, 0};
mixChannels(&timg, 1, &gray0, 1, ch, 1);
// try several threshold levels
for( int l = 0; l < N; l++ )
{
// hack: use Canny instead of zero threshold level.
// Canny helps to catch squares with gradient shading
if( l == 0 )
{
// apply Canny. Take the upper threshold from slider
// and set the lower to 0 (which forces edges merging)
Canny(gray0, gray, 0, thresh, 5);
// dilate canny output to remove potential
// holes between edge segments
dilate(gray, gray, Mat(), Point(-1,-1));
}
else
{
// apply threshold if l!=0:
// tgray(x,y) = gray(x,y) < (l+1)*255/N ? 255 : 0
gray = gray0 >= (l+1)*255/N;
}
This code snipped is a part of the git repository CVSquares. From experience, I know that this library does not work well on real images though works good in computer generated images. Plus this method of detection on RGB without converting to grayscale is very computationally expensive.
Anyways the line of code you are asking about is basically a threshold filter that makes a binary gray Mat array based on the threshold applied to the array gray0 according to the level. If the condition is true, the array contains white pixel at that location else black pixel
A more generic detection code based on the grayscale image would work better like this:
Mat binary_img(color_img.size(),CV_8UC1);
vector<Vec4i> hierarchy;
vector<vector<Point> > contours;
vector<cv::Point> approx;
cv::GaussianBlur(color_img, color_img, cv::Size(9,9), 3);
cv::medianBlur(color_img, color_img, 9);
cv::medianBlur(color_img, color_img, 9);
cvtColor(color_img,binary_img,CV_BGR2GRAY);
IplImage tempBinary=IplImage(binary_img);
cvInRangeS(&tempBinary, Scalar(20), Scalar(100), &tempBinary);
Mat imgMat=Mat(&tempBinary);
findContours( imgMat, contours, hierarchy, CV_RETR_TREE, CV_CHAIN_APPROX_SIMPLE, Point(0, 0) );
for(int i = 0; i < contours.size(); i++ )
{
approxPolyDP(Mat(contours[i]), approx, arcLength(Mat(contours[i]), true)*0.0005, true);
if(approx.size()==4)
//draw approx
}
The second half of the line, gray0 >= (l+1)*255/N, is a condition. gray is a variable, which will contain the truth-fullness of this condition, thus it will be a boolean. But it is c/c++ code, and the booleans are integers here.
About this condition:
(l+1)*255/N will move the interval (1,l) to (0,255). Thus, for l=0 it will be 255/N. For l=N it will be exactly 255.
Depending on the value of l+1 (1 added to make the range started from 1 to 256, instead of 0 to 255) multiplied by 255 and integer divided by N being less than or equal to threshold value gray0, gray is set to 0 or 1 (pure black or pure white)
gray = gray0 >= (l+1)*255/N;
Related
What I'm trying to do is measure the thickness of the eyeglasses frames. I had the idea to measure the thickness of the frame's contours (may be a better way?). I have so far outlined the frame of the glasses, but there are gaps where the lines don't meet. I thought about using HoughLinesP, but I'm not sure if this is what I need.
So far I have conducted the following steps:
Convert image to grayscale
Create ROI around the eye/glasses area
Blur the image
Dilate the image (have done this to remove any thin framed glasses)
Conduct Canny edge detection
Found contours
These are the results:
This is my code so far:
//convert to grayscale
cv::Mat grayscaleImg;
cv::cvtColor( img, grayscaleImg, CV_BGR2GRAY );
//create ROI
cv::Mat eyeAreaROI(grayscaleImg, centreEyesRect);
cv::imshow("roi", eyeAreaROI);
//blur
cv::Mat blurredROI;
cv::blur(eyeAreaROI, blurredROI, Size(3,3));
cv::imshow("blurred", blurredROI);
//dilate thin lines
cv::Mat dilated_dst;
int dilate_elem = 0;
int dilate_size = 1;
int dilate_type = MORPH_RECT;
cv::Mat element = getStructuringElement(dilate_type,
cv::Size(2*dilate_size + 1, 2*dilate_size+1),
cv::Point(dilate_size, dilate_size));
cv::dilate(blurredROI, dilated_dst, element);
cv::imshow("dilate", dilated_dst);
//edge detection
int lowThreshold = 100;
int ratio = 3;
int kernel_size = 3;
cv::Canny(dilated_dst, dilated_dst, lowThreshold, lowThreshold*ratio, kernel_size);
//create matrix of the same type and size as ROI
Mat dst;
dst.create(eyeAreaROI.size(), dilated_dst.type());
dst = Scalar::all(0);
dilated_dst.copyTo(dst, dilated_dst);
cv::imshow("edges", dst);
//join the lines and fill in
vector<Vec4i> hierarchy;
vector<vector<Point>> contours;
cv::findContours(dilated_dst, contours, hierarchy, CV_RETR_TREE, CV_CHAIN_APPROX_SIMPLE);
cv::imshow("contours", dilated_dst);
I'm not entirely sure what the next steps would be, or as I said above, if I should use HoughLinesP and how to implement it. Any help is very much appreciated!
I think there are 2 main problems.
segment the glasses frame
find the thickness of the segmented frame
I'll now post a way to segment the glasses of your sample image. Maybe this method will work for different images too, but you'll probably have to adjust parameters, or you might be able to use the main ideas.
Main idea is:
First, find the biggest contour in the image, which should be the glasses. Second, find the two biggest contours within the previous found biggest contour, which should be the glasses within the frame!
I use this image as input (which should be your blurred but not dilated image):
// this functions finds the biggest X contours. Probably there are faster ways, but it should work...
std::vector<std::vector<cv::Point>> findBiggestContours(std::vector<std::vector<cv::Point>> contours, int amount)
{
std::vector<std::vector<cv::Point>> sortedContours;
if(amount <= 0) amount = contours.size();
if(amount > contours.size()) amount = contours.size();
for(int chosen = 0; chosen < amount; )
{
double biggestContourArea = 0;
int biggestContourID = -1;
for(unsigned int i=0; i<contours.size() && contours.size(); ++i)
{
double tmpArea = cv::contourArea(contours[i]);
if(tmpArea > biggestContourArea)
{
biggestContourArea = tmpArea;
biggestContourID = i;
}
}
if(biggestContourID >= 0)
{
//std::cout << "found area: " << biggestContourArea << std::endl;
// found biggest contour
// add contour to sorted contours vector:
sortedContours.push_back(contours[biggestContourID]);
chosen++;
// remove biggest contour from original vector:
contours[biggestContourID] = contours.back();
contours.pop_back();
}
else
{
// should never happen except for broken contours with size 0?!?
return sortedContours;
}
}
return sortedContours;
}
int main()
{
cv::Mat input = cv::imread("../Data/glass2.png", CV_LOAD_IMAGE_GRAYSCALE);
cv::Mat inputColors = cv::imread("../Data/glass2.png"); // used for displaying later
cv::imshow("input", input);
//edge detection
int lowThreshold = 100;
int ratio = 3;
int kernel_size = 3;
cv::Mat canny;
cv::Canny(input, canny, lowThreshold, lowThreshold*ratio, kernel_size);
cv::imshow("canny", canny);
// close gaps with "close operator"
cv::Mat mask = canny.clone();
cv::dilate(mask,mask,cv::Mat());
cv::dilate(mask,mask,cv::Mat());
cv::dilate(mask,mask,cv::Mat());
cv::erode(mask,mask,cv::Mat());
cv::erode(mask,mask,cv::Mat());
cv::erode(mask,mask,cv::Mat());
cv::imshow("closed mask",mask);
// extract outermost contour
std::vector<cv::Vec4i> hierarchy;
std::vector<std::vector<cv::Point>> contours;
//cv::findContours(mask, contours, hierarchy, CV_RETR_TREE, CV_CHAIN_APPROX_SIMPLE);
cv::findContours(mask, contours, hierarchy, CV_RETR_EXTERNAL, CV_CHAIN_APPROX_SIMPLE);
// find biggest contour which should be the outer contour of the frame
std::vector<std::vector<cv::Point>> biggestContour;
biggestContour = findBiggestContours(contours,1); // find the one biggest contour
if(biggestContour.size() < 1)
{
std::cout << "Error: no outer frame of glasses found" << std::endl;
return 1;
}
// draw contour on an empty image
cv::Mat outerFrame = cv::Mat::zeros(mask.rows, mask.cols, CV_8UC1);
cv::drawContours(outerFrame,biggestContour,0,cv::Scalar(255),-1);
cv::imshow("outer frame border", outerFrame);
// now find the glasses which should be the outer contours within the frame. therefore erode the outer border ;)
cv::Mat glassesMask = outerFrame.clone();
cv::erode(glassesMask,glassesMask, cv::Mat());
cv::imshow("eroded outer",glassesMask);
// after erosion if we dilate, it's an Open-Operator which can be used to clean the image.
cv::Mat cleanedOuter;
cv::dilate(glassesMask,cleanedOuter, cv::Mat());
cv::imshow("cleaned outer",cleanedOuter);
// use the outer frame mask as a mask for copying canny edges. The result should be the inner edges inside the frame only
cv::Mat glassesInner;
canny.copyTo(glassesInner, glassesMask);
// there is small gap in the contour which unfortunately cant be closed with a closing operator...
cv::dilate(glassesInner, glassesInner, cv::Mat());
//cv::erode(glassesInner, glassesInner, cv::Mat());
// this part was cheated... in fact we would like to erode directly after dilation to not modify the thickness but just close small gaps.
cv::imshow("innerCanny", glassesInner);
// extract contours from within the frame
std::vector<cv::Vec4i> hierarchyInner;
std::vector<std::vector<cv::Point>> contoursInner;
//cv::findContours(glassesInner, contoursInner, hierarchyInner, CV_RETR_TREE, CV_CHAIN_APPROX_SIMPLE);
cv::findContours(glassesInner, contoursInner, hierarchyInner, CV_RETR_EXTERNAL, CV_CHAIN_APPROX_SIMPLE);
// find the two biggest contours which should be the glasses within the frame
std::vector<std::vector<cv::Point>> biggestInnerContours;
biggestInnerContours = findBiggestContours(contoursInner,2); // find the one biggest contour
if(biggestInnerContours.size() < 1)
{
std::cout << "Error: no inner frames of glasses found" << std::endl;
return 1;
}
// draw the 2 biggest contours which should be the inner glasses
cv::Mat innerGlasses = cv::Mat::zeros(mask.rows, mask.cols, CV_8UC1);
for(unsigned int i=0; i<biggestInnerContours.size(); ++i)
cv::drawContours(innerGlasses,biggestInnerContours,i,cv::Scalar(255),-1);
cv::imshow("inner frame border", innerGlasses);
// since we dilated earlier and didnt erode quite afterwards, we have to erode here... this is a bit of cheating :-(
cv::erode(innerGlasses,innerGlasses,cv::Mat() );
// remove the inner glasses from the frame mask
cv::Mat fullGlassesMask = cleanedOuter - innerGlasses;
cv::imshow("complete glasses mask", fullGlassesMask);
// color code the result to get an impression of segmentation quality
cv::Mat outputColors1 = inputColors.clone();
cv::Mat outputColors2 = inputColors.clone();
for(int y=0; y<fullGlassesMask.rows; ++y)
for(int x=0; x<fullGlassesMask.cols; ++x)
{
if(!fullGlassesMask.at<unsigned char>(y,x))
outputColors1.at<cv::Vec3b>(y,x)[1] = 255;
else
outputColors2.at<cv::Vec3b>(y,x)[1] = 255;
}
cv::imshow("output", outputColors1);
/*
cv::imwrite("../Data/Output/face_colored.png", outputColors1);
cv::imwrite("../Data/Output/glasses_colored.png", outputColors2);
cv::imwrite("../Data/Output/glasses_fullMask.png", fullGlassesMask);
*/
cv::waitKey(-1);
return 0;
}
I get this result for segmentation:
the overlay in original image will give you an impression of quality:
and inverse:
There are some tricky parts in the code and it's not tidied up yet. I hope it's understandable.
The next step would be to compute the thickness of the the segmented frame. My suggestion is to compute the distance transform of the inversed mask. From this you will want to compute a ridge detection or skeletonize the mask to find the ridge. After that use the median value of ridge distances.
Anyways I hope this posting can help you a little, although it's not a solution yet.
Depending on lighting, frame color etc this may or may not work but how about simple color detection to separate the frame ? Frame color will usually be a lot darker than human skin. You'll end up with a binary image (just black and white) and by calculating the number (area) of black pixels you get the area of the frame.
Another possible way is to get better edge detection, by adjusting/dilating/eroding/both until you get better contours. You will also need to differentiate the contour from the lenses and then apply cvContourArea.
I am trying to track a custom circular marker in an image, and I need to check that a circle contains a minimum number of other circles/objects. My code for finding circles is below:
void findMarkerContours( int, void* )
{
vector<vector<Point> > contours;
vector<Vec4i> hierarchy;
vector<Point> approx;
cv::Mat dst = src.clone();
cv::Mat src_gray;
cv::cvtColor(src, src_gray, CV_BGR2GRAY);
//Reduce noise with a 3x3 kernel
blur( src_gray, src_gray, Size(3,3));
//Convert to binary using canny
cv::Mat bw;
cv::Canny(src_gray, bw, thresh, 3*thresh, 3);
imshow("bw", bw);
findContours(bw.clone(), contours, CV_RETR_TREE, CV_CHAIN_APPROX_SIMPLE);
Mat drawing = Mat::zeros( bw.size(), CV_8UC3 );
for (int i = 0; i < contours.size(); i++)
{
Scalar color = Scalar( rng.uniform(0, 255), rng.uniform(0,255), rng.uniform(0,255) );
// contour
drawContours( drawing, contours, i, color, 1, 8, vector<Vec4i>(), 0, Point() );
//Approximate the contour with accuracy proportional to contour perimeter
cv::approxPolyDP(cv::Mat(contours[i]), approx, cv::arcLength(cv::Mat(contours[i]), true) *0.02, true);
//Skip small or non-convex objects
if(fabs(cv::contourArea(contours[i])) < 100 || !cv::isContourConvex(approx))
continue;
if (approx.size() >= 8) //More than 6-8 vertices means its likely a circle
{
drawContours( dst, contours, i, Scalar(0,255,0), 2, 8);
}
imshow("Hopefully we should have circles! Yay!", dst);
}
namedWindow( "Contours", CV_WINDOW_AUTOSIZE );
imshow( "Contours", drawing );
}
As you can see the code to detect circles works quite well:
But now I need to filter out markers that I do not want. My marker is the bottom one. So once I have found a contour that is a circle, I want to check if there are other circular contours that exist within the region of the first circle and finally check the color of the smallest circle.
What method can I take to say if (circle contains 3+ smaller circles || smallest circle is [color] ) -> do stuff?
Take a look at the documentation for
findContours(InputOutputArray image, OutputArrayOfArrays contours, OutputArray hierarchy, int mode, int method, Point offset=Point())
You'll see that there's an optional hierarchy output vector which should be handy for your problem.
hierarchy – Optional output vector, containing information about the image topology. It has as many elements as the number of contours.
For each i-th contour contours[i] , the elements hierarchy[i][0] ,
hiearchyi , hiearchyi , and hiearchyi are set to
0-based indices in contours of the next and previous contours at the
same hierarchical level, the first child contour and the parent
contour, respectively. If for the contour i there are no next,
previous, parent, or nested contours, the corresponding elements of
hierarchy[i] will be negative.
When calling findCountours using CV_RETR_TREE you'll be getting the full hierarchy of each contour that was found.
This doc explains the hierarchy format pretty well.
You are already searching for circles of a certain size
//Skip small or non-convex objects
if(fabs(cv::contourArea(contours[i])) < 100 || !cv::isContourConvex(approx))
continue;
So you can use that to look for smaller circles than the one youve got, instead of looking for < 100 look for contours.size
I imagine there is the same for color also...
I am working in C++ and opencv
I am detecting the big contour in an image because I have a black area in it.
In this case, the area is only horizontally, but it can be in any place.
Mat resultGray;
cvtColor(result,resultGray, COLOR_BGR2GRAY);
medianBlur(resultGray,resultGray,3);
Mat resultTh;
Mat canny_output;
vector<vector<Point> > contours;
vector<Vec4i> hierarchy;
Canny( resultGray, canny_output, 100, 100*2, 3 );
findContours( canny_output, contours, hierarchy, CV_RETR_TREE, CV_CHAIN_APPROX_SIMPLE, Point(0, 0) );
Vector<Point> best= contours[0];
int max_area = -1;
for( int i = 0; i < contours.size(); i++ ) {
Scalar color = Scalar( 0, 0, 0 );
if(contourArea(contours[i])> max_area)
{
max_area=contourArea(contours[i]);
best=contours[i];
}
}
Mat approxCurve;
approxPolyDP(Mat(best),approxCurve,0.01*arcLength(Mat(best),true),true);
Wiht this, i have the big contour and it approximation (in approxCurve). Now, I want to obtain the corners of this approximation and get the image inside this contour, but I dont know how can I do it.
I am using this How to remove black part from the image?
But the last part I dont understad very well.
Anyone knows how can I obtain the corners? It is another way more simple that this?
Thanks for your time,
One much simpler way you could do that is to check the image pixels and find the minimum/maximum coordinates of non-black pixels.
Something like this:
int maxx,maxy,minx,miny;
maxx=maxy=-std::numeric_limits<int>::max();
minx=miny=std::numeric_limits<int>::min();
for(int y=0; y<img.rows; ++y)
{
for(int x=0; x<img.cols; ++x)
{
const cv::Vec3b &px = img.at<cv::Vec3b>(y,x);
if(px(0)==0 && px(1)==0 && px(2)==0)
continue;
if(x<minx) minx=x;
if(x>maxx) maxx=x;
if(y<miny) miny=y;
if(y>maxy) maxy=y;
}
}
cv::Mat subimg;
img(cv::Rect(cv::Point(minx,miny),cv::Point(maxx,maxy))).copyTo(subimg);
In my opinion, this approach is more reliable since you don't have to detect any contour, which could lead to false detections depending on the input image.
In a very efficient way, you can sample the original image until you find a pixel on, and from there move along a row and along a column to find the first (0,0,0) pixel. It will work, unless in the good part of the image you can have (0,0,0) pixels. If this is the case (e.g.: dead pixel), you can add a double check checking the neighbourhood of this (0,0,0) pixel (it should contain other (0,0,0) pixels.
Say I have the following binary image created from the output of cv::watershed():
Now I want to find and fill the contours, so I can separate the corresponding objects from the background in the original image (that was segmented by the watershed function).
To segment the image and find the contours I use the code below:
cv::Mat bgr = cv::imread("test.png");
// Some function that provides the rough outline for the segmented regions.
cv::Mat markers = find_markers(bgr);
cv::watershed(bgr, markers);
cv::Mat_<bool> boundaries(bgr.size());
for (int i = 0; i < bgr.rows; i++) {
for (int j = 0; j < bgr.cols; j++) {
boundaries.at<bool>(i, j) = (markers.at<int>(i, j) == -1);
}
}
std::vector<std::vector<cv::Point> > contours;
std::vector<cv::Vec4i> hierarchy;
cv::findContours(
boundaries, contours, hierarchy,
CV_RETR_EXTERNAL, CV_CHAIN_APPROX_SIMPLE
);
So far so good. However, if I pass the contours acquired above to cv::drawContours() as below:
cv::Mat regions(bgr.size(), CV_32S);
cv::drawContours(
regions, contours, -1, cv::Scalar::all(255),
CV_FILLED, 8, hierarchy, INT_MAX
);
This is what I get:
The leftmost contour was left open by cv::findContours(), and as a result it is not filled by cv::drawContours().
Now I know this is a consequence of cv::findContours() clipping off the 1-pixel border around the image (as mentioned in the documentation), but what to do then? It seems an awful waste to discard a contour just because it happened to brush off the image's border. And anyway how can I even find which contour(s) fall in this category? cv::isContourConvex() is not a solution in this case; a region can be concave but "closed" and thus not have this problem.
Edit: About the suggestion to duplicate the pixels from the borders. The problem is that my marking function is also painting all pixels in the "background", i.e. those regions I'm sure aren't part of any object:
This results in a boundary being drawn around the output. If I somehow avoid cv::findContours() to clip off that boundary:
The boundary for the background gets merged with that leftmost object:
Which results in a nice white-filled box.
Solution number 1: use image extended by one pixel in each direction:
Mat extended(bgr.size()+Size(2,2), bgr.type());
Mat markers = extended(Rect(1, 1, bgr.cols, bgr.rows));
// all your calculation part
std::vector<std::vector<Point> > contours;
findContours(boundaries, contours, CV_RETR_EXTERNAL, CV_CHAIN_APPROX_SIMPLE);
Mat regions(bgr.size(), CV_8U);
drawContours(regions, contours, -1, Scalar(255), CV_FILLED, 8, Mat(), INT_MAX, Point(-1,-1));
Note that contours were extracted from extended image, i.e. their x and y values are bigger by 1 from what they should be. This is why I use drawContours with (-1,-1) pixel offset.
Solution number 2: add white pixels from boundary of image to the neighbor row/column:
bitwise_or(boundaries.row(0), boundaries.row(1), boundaries.row(1));
bitwise_or(boundaries.col(0), boundaries.col(1), boundaries.col(1));
bitwise_or(boundaries.row(bgr.rows()-1), boundaries.row(bgr.rows()-2), boundaries.row(bgr.rows()-2));
bitwise_or(boundaries.col(bgr.cols()-1), boundaries.col(bgr.cols()-2), boundaries.col(bgr.cols()-2));
Both solution are half-dirty workarounds, but this is all I could think about.
Following Burdinov's suggestions I came up with the code below, which correctly fills all extracted regions while ignoring the all-enclosing boundary:
cv::Mat fill_regions(const cv::Mat &bgr, const cv::Mat &prospective) {
static cv::Scalar WHITE = cv::Scalar::all(255);
int rows = bgr.rows;
int cols = bgr.cols;
// For the given prospective markers, finds
// object boundaries on the given BGR image.
cv::Mat markers = prospective.clone();
cv::watershed(bgr, markers);
// Copies the boundaries of the objetcs segmented by cv::watershed().
// Ensures there is a minimum distance of 1 pixel between boundary
// pixels and the image border.
cv::Mat borders(rows + 2, cols + 2, CV_8U);
for (int i = 0; i < rows; i++) {
uchar *u = borders.ptr<uchar>(i + 1) + 1;
int *v = markers.ptr<int>(i);
for (int j = 0; j < cols; j++, u++, v++) {
*u = (*v == -1);
}
}
// Calculates contour vectors for the boundaries extracted above.
std::vector<std::vector<cv::Point> > contours;
std::vector<cv::Vec4i> hierarchy;
cv::findContours(
borders, contours, hierarchy,
CV_RETR_LIST, CV_CHAIN_APPROX_SIMPLE
);
int area = bgr.size().area();
cv::Mat regions(borders.size(), CV_32S);
for (int i = 0, n = contours.size(); i < n; i++) {
// Ignores contours for which the bounding rectangle's
// area equals the area of the original image.
std::vector<cv::Point> &contour = contours[i];
if (cv::boundingRect(contour).area() == area) {
continue;
}
// Draws the selected contour.
cv::drawContours(
regions, contours, i, WHITE,
CV_FILLED, 8, hierarchy, INT_MAX
);
}
// Removes the 1 pixel-thick border added when the boundaries
// were first copied from the output of cv::watershed().
return regions(cv::Rect(1, 1, cols, rows));
}
Hi I am working on computer vision project and trying to detect square by using openCV/C++ from camera. I had download the source code from openCV library but it seems like losing fps so hard. Does anybody have idea how to fix this problem? There is a video link about my testing below, check it out:
http://magicbookproject.blogspot.co.uk/2012/12/detect-paper-demo.html
Here is the code and just found on another post:
void find_squares(Mat& image, vector<vector<Point> >& squares)
{
// blur will enhance edge detection
Mat blurred(image);
medianBlur(image, blurred, 9);
Mat gray0(blurred.size(), CV_8U), gray;
vector<vector<Point> > contours;
// find squares in every color plane of the image
for (int c = 0; c < 3; c++)
{
int ch[] = {c, 0};
mixChannels(&blurred, 1, &gray0, 1, ch, 1);
// try several threshold levels
const int threshold_level = 2;
for (int l = 0; l < threshold_level; l++)
{
// Use Canny instead of zero threshold level!
// Canny helps to catch squares with gradient shading
if (l == 0)
{
Canny(gray0, gray, 10, 20, 3); //
// Dilate helps to remove potential holes between edge segments
dilate(gray, gray, Mat(), Point(-1,-1));
}
else
{
gray = gray0 >= (l+1) * 255 / threshold_level;
}
// Find contours and store them in a list
findContours(gray, contours, CV_RETR_LIST, CV_CHAIN_APPROX_SIMPLE);
// Test contours
vector<Point> approx;
for (size_t i = 0; i < contours.size(); i++)
{
// approximate contour with accuracy proportional
// to the contour perimeter
approxPolyDP(Mat(contours[i]), approx, arcLength(Mat(contours[i]), true)*0.02, true);
// Note: absolute value of an area is used because
// area may be positive or negative - in accordance with the
// contour orientation
if (approx.size() == 4 &&
fabs(contourArea(Mat(approx))) > 1000 &&
isContourConvex(Mat(approx)))
{
double maxCosine = 0;
for (int j = 2; j < 5; j++)
{
double cosine = fabs(angle(approx[j%4], approx[j-2], approx[j-1]));
maxCosine = MAX(maxCosine, cosine);
}
if (maxCosine < 0.3)
squares.push_back(approx);
}
}
}
}
}
You can speed it up if you don't mind losing accuracy. For example
// find squares in every color plane of the image
for (int c = 0; c < 3; c++)
You are looping through three color planes. Just examine one color (as if the image is grayscale), that should triple the speed.
Also try without Canny, which is quite slow. Set a use_canny parameter,
if (l == 0 && use_canny)
{
Canny(gray0, gray, 10, 20, 3); //
Compare with and without. I am getting acceptable results, considerably faster.
A good rule of thumb for computer vision is to convert your image to grayscale before doing any intensive processing. Only loop through color channels if you find it absolutely necessary. I recommend the following pattern for object recognition:
Convert image to grayscale
Filter grayscale image to a simpler format (canny, threshold, edge-detect)
Do heavy processing (detect square shapes)
Reconstruct original image with the processed values (draw/store your squares)
Remember that you are doing all these steps for each and every frame, so be sure to remove whatever you find is unnecessary. Since this code will run so often, you will see huge performance benefits from even a minor optimization, so it's worth spending some time optimizing.