Develop Reference

Image Matching on Android Game's Board

The question comes 1st: I am looking for FAST approach to match images.
Now, the use case: I am developing a detector to detect orb on a 6x5 Match-3 game board for android platform. I have an array of the orb icon with transparent background, but the orb on the screen (screenshot) has different background color, probably different size too. I have to compare each orb on the screen with my array of icons (69 icons specifically) so it's a 69x30=2070 steps. I tried lazy implementation and group almost similar icon together to reduce the steps but still take a long time (10s at most) for computation. I also tried checking the channel and depth of image, resizing the images to have same size and tweaking the threshold value but still no luck.
I have tried Histogram Matching (seperate channel, grayscale), Template Matching (CCOEFF, SQDIFF, CCORR), AKAZE, ORB(unbounded, bounded), PHash all using OpenCV but histogram matching and PHash give me erroneous result (too much false positive), Template Matching consume 10s+ (considered too slow for user to wait) while AKAZE and ORB give better result than all other methods but still needs 6s+ per try. Is there any other method that can helps me cut down the computation time down to somewhere near 1s and can give better result considering the worst case scenario is 2070 steps?
Referrences that I have read that compares the performances of different feature matching algorithms:
A comparative analysis of SIFT, SURF, KAZE, AKAZE, ORB, and BRISK. It shows that ORB and BRISK should be averagely better than the other approach compared while AKAZE is moderately good for most cases. I deleted my Histogram comparison code as it is not really helpful but you may find the rest of it below.
Mat source = Utils.loadResource(this, R.drawable.orb_icon, Imgcodecs.CV_LOAD_IMAGE_UNCHANGED);
Mat tmp = new Mat();
Bitmap cropped_img = Bitmap.createBitmap(screenshot, x, y, width, height);
Utils.bitmapToMat(cropped_img, tmp);
//template matching code
int r_rows = source.rows() - tmp.rows() + 1;
int r_cols = source.cols() - tmp.cols() + 1;
Mat result = new Mat();
result.create(r_rows, r_cols, CvType.CV_32F);
Imgproc.matchTemplate(source, tmp, result, Imgproc.TM_CCOEFF_NORMED);
Core.MinMaxLocResult mmr = Core.minMaxLoc(result);
double maxVal = mmr.maxVal;
return maxVal;
//AKAZE
MatOfKeyPoint kp1 = new MatOfKeyPoint();
MatOfKeyPoint kp2 = new MatOfKeyPoint();
Mat desc1 = new Mat();
Mat desc2 = new Mat();
AKAZE akaze = AKAZE.create();
DescriptorMatcher matcher = DescriptorMatcher.create(DescriptorMatcher.BRUTEFORCE_HAMMING);
akaze.detectAndCompute(source, new Mat(), kp1, desc1);
akaze.detectAndCompute(tmp, new Mat(), kp2, desc2);
List<MatOfDMatch> knnMatches = new ArrayList<>();
matcher.knnMatch(desc1, desc2, knnMatches, 2);
float threshold = 0.7f;
int count = 0;
for(int i=0; i<knnMatches.size(); i++) {
if(knnMatches.get(i).rows() > 1) {
DMatch[] matches = knnMatches.get(i).toArray();
if(matches[0].distance < threshold * matches[1].distance) {
count++;
}
}
}
//ORB
ORB orb = ORB.create();
DescriptorMatcher matcher = DescriptorMatcher.create(DescriptorMatcher.BRUTEFORCE_HAMMING);
MatOfKeyPoint kp1 = new MatOfKeyPoint();
MatOfKeyPoint kp2 = new MatOfKeyPoint();
Mat desc1 = new Mat();
Mat desc2 = new Mat();
orb.detectAndCompute(source, new Mat(), kp1, desc1);
orb.detectAndCompute(tmp, new Mat(), kp2, desc2);
List<MatOfDMatch> knnMatches = new ArrayList<>();
matcher.knnMatch(desc1, desc2, knnMatches, 2);
float threshold = 0.8f;
int count = 0;
for(int i=0; i<knnMatches.size(); i++) {
if(knnMatches.get(i).rows() > 1) {
DMatch[] matches = knnMatches.get(i).toArray();
if(matches[0].distance < threshold * matches[1].distance) {
count++;
}
}
}
//PHash
Mat hash_source = new Mat();
Mat hash_tmp = new Mat();
Img_hash.pHash(tmp, hash_tmp);
Img_hash.pHash(source, hash_source);
Core.norm(source, tmp, Core.NORM_HAMMING);
Edit: As suggested, below is the game board, icon image, and orb screenshot sample.
ICON vs orb screenshot
Also, you may observe the simulation result of each approach by comparing the result(overlay smaller icon) on top of the orb on board:
Histogram Matching
,
Template Matching
and
AKAZE (similar to ORB)
After moving the variable initialization out of my comparison function to base class, detect keypoint and PHash of source icon images on class initialization, run detect and compute function in batch using List to reduce individual function call. It still takes up 4s+ for the image matching process. Time consumption is reduced but accuracy is still a major problem. You may observe my heap stack on below.

How can i improve openCV people detecting algorithm

I am trying to write human detector, it works now, but sometimes it reacts on cats/boxes etc., also i got like 5 fps. So the question is, how can i improve my algorithm for better fps and detection accuracy.
I have tried to use this one:
http://www.pyimagesearch.com/2015/11/09/pedestrian-detection-opencv/
But i couldnt find any way i could use this on android.
public Mat onCameraFrame(CvCameraViewFrame inputFrame) {
List<MatOfPoint> list = new ArrayList<>();
Mat frame = new Mat();
Mat gray = new Mat();
Mat hierarchy = new Mat();
Mat originalFrame = inputFrame.rgba();
Imgproc.medianBlur(originalFrame,originalFrame,3);
Imgproc.cvtColor(originalFrame, gray, Imgproc.COLOR_RGB2GRAY, 0);
HOGDescriptor hog = new HOGDescriptor();
//Получаем стандартный определитель людей и устанавливаем его нашему дескриптору
MatOfFloat descriptors = HOGDescriptor.getDefaultPeopleDetector();
hog.setSVMDetector(descriptors);
MatOfRect locations = new MatOfRect();
MatOfDouble weights = new MatOfDouble();
hog.detectMultiScale(gray, locations, weights);
Point rectPoint1 = new Point();
Point rectPoint2 = new Point();
Point fontPoint = new Point();
if (locations.rows() > 0) {
List<Rect> rectangles = locations.toList();
for (Rect rect : rectangles) {
rectPoint1.x = rect.x;
rectPoint1.y = rect.y;
fontPoint.x = rect.x;
fontPoint.y = rect.y - 4;
rectPoint2.x = rect.x + rect.width;
rectPoint2.y = rect.y + rect.height;
final Scalar rectColor = new Scalar( 0 , 0 , 0 );
// Добавляем на изображения найденную информацию
Imgproc.rectangle(originalFrame, rectPoint1, rectPoint2, rectColor, 2);
}
}
frame.release();
gray.release();
hierarchy.release();
list.clear();
return originalFrame;
}

You're using the HOG+SVM approach to detect people; it is inherently going to be quite slow. Never the less, you can use some of the suggestions in this question How to speed up svm.predict?
Depending on your problem, i.e. if the camera is static and the pedestrians are moving you could opt for a background subtraction approach this is probably the most efficient way but bear in mind that this will pick up any objects that are moving in the scene, so you could include thresholds to remove small objects. Some background subtraction algorithms include mixture of gaussian (MOG) or MOG2 or GMG. Also, an important thing to note is that these approaches rely on creating a background model of the scene, i.e. they assume static pixels over time to be part of the background, hence, when a pedestrian stands still for a while in the scene they get embedded into the background resulting in miss detection. There are many papers out there that provide potential solutions to that problem so you might want to have a look at them, here is one that produces decent results: Static and Moving Object Detection Using Flux Tensor with Split Gaussian Models
Additionally, you could opt for a data driven approach, either get a good pre-trained model and do your detection using that or train one yourself using TensorFlow, Caffe or Torch and use the dnn opencv_contrib module to do the detection.

opencv4android template matching using camera

I have downloaded and successfully run the example provided in opencv4android sdk.
I am able to simply display the camera frames without any processing,
public Mat onCameraFrame(CvCameraViewFrame inputFrame) {
return inputFrame.rgba();
}
I want to process live frame with some predefined image template to recognize that template. I have taken reference from this post and implemented accordingly. But I get black screen only.
private Mat mCameraMat = new Mat();
private Mat mTemplateMat;
public Mat onCameraFrame(CvCameraViewFrame inputFrame) {
mCameraMat = inputFrame.rgba();
initialize();
int match_method = Imgproc.TM_SQDIFF;
// Create the result matrix
int result_cols = mCameraMat.cols() - mTemplateMat.cols() + 1;
int result_rows = mCameraMat.rows() - mTemplateMat.rows() + 1;
Log.d(TAG, " mCameraMat cols "+mCameraMat.cols());
Log.d(TAG, " mCameraMat rows "+mCameraMat.rows());
Log.d(TAG, " mTemplateMat cols "+mTemplateMat.cols());
Log.d(TAG, " mTemplateMat rows "+mTemplateMat.rows());
Mat result = new Mat(result_rows, result_cols, CvType.CV_32F);
// Do the Matching and Normalize
Imgproc.matchTemplate(mCameraMat, mTemplateMat, result, match_method);
Core.normalize(result, result, 0, 1, Core.NORM_MINMAX, -1, new Mat());
// Localizing the best match with minMaxLoc
MinMaxLocResult mmr = Core.minMaxLoc(result);
Point matchLoc;
if (match_method == Imgproc.TM_SQDIFF || match_method == Imgproc.TM_SQDIFF_NORMED) {
matchLoc = mmr.minLoc;
} else {
matchLoc = mmr.maxLoc;
}
Rect roi = new Rect((int) matchLoc.x, (int) matchLoc.y, mTemplateMat.cols(), mTemplateMat.rows());
Core.rectangle(mCameraMat, new Point(roi.x, roi.y), new Point(roi.width - 2, roi.height - 2), new Scalar(255, 0, 0, 255), 2);
return result;
}
public void initialize(){
try {
if (mCameraMat.empty())
return;
if(mTemplateMat == null){
Mat temp = Utils.loadResource(Tutorial1Activity.this, R.drawable.icon);
mTemplateMat = new Mat(temp.size(), CvType.CV_32F);
Imgproc.cvtColor(temp, mTemplateMat, Imgproc.COLOR_BGR2RGBA);
Log.d(TAG, "initialize mTemplateMat cols "+mTemplateMat.cols());
Log.d(TAG, "initialize mTemplateMat rows "+mTemplateMat.rows());
}
} catch (IOException e) {
e.printStackTrace();
}
}
Note:
My ultimate goal is to recognize the playing cards from live camera. Kindly suggest best approach. Should I use image templates or any other thing to make things faster?
This is how I want to recognize multiple cards from live camera:
Result should be: ♠A ♠K ♠Q ♠J ♠10 when camera preview seems like below

Template matching is unlikely to be the best approach here.
Try aSIFT to do an affine invariant SIFT matching or a normal SIFT (OpenCV implementation exists). However,
since these are in C++, you may want to use JNI to make calls to it from Java on an Android device. This is probably the best way to detect the suit of the card from the 4 symbols.
Another option to detect and recognize the numbers/alphabets on the cards is to use a text detector like MSER and then a text recognizer on the regions of interest indicated by the MSER filter.
In any case, you are unlikely to be able produce the best of results from the kind of image you've shown in the image. You may be able to get acceptable performance for full frontal, upright images with the first method.

counting objects & better way to filling holes

I am new to OpenCV and am trying to count the number of objects in an image. I have done this before using MATLAB Image Processing Toolbox and adapted the same approach in OpenCV (Android) also.
The first step was to convert an image to gray scale. Then to threshold it and then counting the number of blobs. In Matlab there is a command - "bwlabel", which gives the number of blobs. I couldn't find such thing in OpenCV (again, I am a noob in OpenCV as well as Android).
Here is my code,
//JPG to Bitmap to MAT
Bitmap i = BitmapFactory.decodeFile(imgPath + "mms.jpg");
Bitmap bmpImg = i.copy(Bitmap.Config.ARGB_8888, false);
Mat srcMat = new Mat ( bmpImg.getHeight(), bmpImg.getWidth(), CvType.CV_8UC3);
Utils.bitmapToMat(bmpImg, srcMat);
//convert to gray scale and save image
Mat gray = new Mat(srcMat.size(), CvType.CV_8UC1);
Imgproc.cvtColor(srcMat, gray, Imgproc.COLOR_RGB2GRAY,4);
//write bitmap
Boolean bool = Highgui.imwrite(imgPath + "gray.jpg", gray);
//thresholding
Mat threshed = new Mat(bmpImg.getWidth(),bmpImg.getHeight(), CvType.CV_8UC1);
Imgproc.adaptiveThreshold(gray, threshed, 255, Imgproc.ADAPTIVE_THRESH_MEAN_C, Imgproc.THRESH_BINARY, 75, 5);//15, 8 were original tests. Casey was 75,10
Core.bitwise_not(threshed, threshed);
Utils.matToBitmap(threshed, bmpImg);
//write bitmap
bool = Highgui.imwrite(imgPath + "threshed.jpg", threshed);
Toast.makeText(this, "Thresholded image saved!", Toast.LENGTH_SHORT).show();
In the next step, I tried to fill the holes and letters using dilation followed by an erosion but the blobs gets attached to each other which will ultimately give a wrong count. There is a tradeoff between filling holes and getting the blobs attached to each other on tuning the parameters for dilation and erosion.
Here is the code,
//morphological operations
//dilation
Mat dilated = new Mat(bmpImg.getWidth(),bmpImg.getHeight(), CvType.CV_8UC1);
Imgproc.dilate(threshed, dilated, Imgproc.getStructuringElement(Imgproc.MORPH_ELLIPSE, new org.opencv.core.Size (16, 16)));
Utils.matToBitmap(dilated, bmpImg);
//write bitmap
bool = Highgui.imwrite(imgPath + "dilated.jpg", dilated);
Toast.makeText(this, "Dilated image saved!", Toast.LENGTH_SHORT).show();
//erosion
Mat eroded = new Mat(bmpImg.getWidth(),bmpImg.getHeight(), CvType.CV_8UC1);
Imgproc.erode(dilated, eroded, Imgproc.getStructuringElement(Imgproc.MORPH_ELLIPSE, new org.opencv.core.Size(15, 15)));
Utils.matToBitmap(eroded, bmpImg);
//write bitmap
bool = Highgui.imwrite(imgPath + "eroded.jpg", eroded);
Toast.makeText(this, "Eroded image saved!", Toast.LENGTH_SHORT).show();
Because sometimes my M&Ms might be just next to each other! ;)
I also tried to use Hough Circles but the result is very unreliable (tested with coin images as well as real coins)
Here is the code,
//hough circles
Mat circles = new Mat();
// parameters
int iCannyUpperThreshold = 100;
int iMinRadius = 20;
int iMaxRadius = 400;
int iAccumulator = 100;
Imgproc.HoughCircles(gray, circles, Imgproc.CV_HOUGH_GRADIENT,
1.0, gray.rows() / 8, iCannyUpperThreshold, iAccumulator,
iMinRadius, iMaxRadius);
// draw
if (circles.cols() > 0)
{
Toast.makeText(this, "Coins : " +circles.cols() , Toast.LENGTH_LONG).show();
}
else
{
Toast.makeText(this, "No coins found", Toast.LENGTH_LONG).show();
}
The problem with this approach is that the algorithm is limited to perfect circles only (AFAIK). So, it doesn't work well when I try to scan and count M&Ms or coins lying on my desk (because angle of the device changes). With this approach, sometimes I get less no. of coins detected and sometimes more (I don't get it why more??).
On scanning this image the app sometimes shows 19 coins and sometimes 38 coins counted...I know there are other features which may be detected as circles but I totally don't get it why 38..?
So my questions...
Is there a better way to fill holes without joining adjacent blobs?
How do I count the number of objects accurately? I don't want to limit my app to counting only circles with HoughCircles approach.
FYI : OpenCV-2.4.9-android-sdk. Kindly keep in mind that I am a newbie in OpenCV and Android too.
Any help is much appreciated.
Thanks & Cheers!
Jainam

So to proceed we take your threshold image which you have generated as input and further modify it. The present code is in C++ but I guess you can easily convert it into android platform
Now instead of dilation or blurring you can try flood fill
which results in
Finally now applying the contour detection algorithm algorithm we get
The code for the above is
Mat dst = imread($path to the threshold image); // image should be single channel black and white image
imshow("dst",dst);
cv::Mat mask = cv::Mat::zeros(dst.rows + 2, dst.cols + 2, CV_8U);
// A image with size greater than the present object is created
cv::floodFill(dst, mask, cv::Point(0,0), 255, 0, cv::Scalar(), cv::Scalar(), 4 + (255 << 8) + cv::FLOODFILL_MASK_ONLY);
erode(mask,mask,Mat());
// Now to remove the outer boundary
rectangle(mask,Rect(0,0,mask.cols,mask.rows), Scalar(255,255,255),2,8,0);
imshow("Mask",mask);
Mat copy;
mask.copyTo(copy);
vector<vector<Point> > contours;
vector<Vec4i> hierarchy;
findContours( copy, contours, hierarchy, CV_RETR_TREE, CV_CHAIN_APPROX_SIMPLE, Point(0, 0) );
vector<vector<Point> > contours_poly( contours.size() );
vector<Rect> boundRect( contours.size() );
vector<Point2f>center( contours.size() );
vector<float>Distance( contours.size() );
vector<float>radius( contours.size() );
Mat drawing = cv::Mat::zeros(mask.rows, mask.cols, CV_8U);
int num_object = 0;
for( int i = 0; i < contours.size(); i++ ){
approxPolyDP( Mat(contours[i]), contours_poly[i], 3, true );
// To get rid of the smaller object and the outer rectangle created
//because of the additional mask image we enforce a lower limit on area
//to remove noise and an upper limit to remove the outer border.
if (contourArea(contours_poly[i])>(mask.rows*mask.cols/10000) && contourArea(contours_poly[i])<mask.rows*mask.cols*0.9){
boundRect[i] = boundingRect( Mat(contours_poly[i]) );
minEnclosingCircle( (Mat)contours_poly[i], center[i], radius[i] );
circle(drawing,center[i], (int)radius[i], Scalar(255,255,255), 2, 8, 0);
rectangle(drawing,boundRect[i], Scalar(255,255,255),2,8,0);
num_object++;
}
}
cout <<"No. of object detected =" <<num_object<<endl;
imshow("drawing",drawing);
waitKey(2);
char key = (char) waitKey(20);
if(key == 32){
// You can save your images here using a space
}
I hope this helps you in solving your problem

Just check it out,
Blur source.
Threshold binary inverted on gray.
Find contours, note that you should use CV_RETR_EXTERNAL as contour retrieval mode.
You can take the contours size as your object count.
Code:
Mat tmp,thr;
Mat src=imread("img.jpg",1);
blur(src,src,Size(3,3));
cvtColor(src,tmp,CV_BGR2GRAY);
threshold(tmp,thr,220,255,THRESH_BINARY_INV);
imshow("thr",thr);
vector< vector <Point> > contours; // Vector for storing contour
vector< Vec4i > hierarchy;
findContours( thr, contours, hierarchy,CV_RETR_EXTERNAL, CV_CHAIN_APPROX_SIMPLE ); // Find the contours in the image
for( int i = 0; i< contours.size(); i=hierarchy[i][0] ) // iterate through each contour.
{
Rect r= boundingRect(contours[i]);
rectangle(src,r, Scalar(0,0,255),2,8,0);
}
cout<<"Numeber of contour = "<<contours.size()<<endl;
imshow("src",src);
waitKey();

Optical flow in Android

We have been dealing with OpenCV for two weeks to make it work on Android.
Do you know where can we find an Android implementation of optical flow? It would be nice if it's implemented using OpenCV.

Openframeworks has openCV baked in, as well as many other interesting libraries. It has a very elegant strucutre, and I have used it with android to make a virtual mouse of the phone using motion estimation from the camera.
See the ports to android here http://openframeworks.cc/setup/android-studio/
Seems they recently added support for android studio, otherwise eclipse works great.

Try this
#Override
public Mat onCameraFrame(CvCameraViewFrame inputFrame) {
mRgba = inputFrame.rgba();
if (mMOP2fptsPrev.rows() == 0) {
//Log.d("Baz", "First time opflow");
// first time through the loop so we need prev and this mats
// plus prev points
// get this mat
Imgproc.cvtColor(mRgba, matOpFlowThis, Imgproc.COLOR_RGBA2GRAY);
// copy that to prev mat
matOpFlowThis.copyTo(matOpFlowPrev);
// get prev corners
Imgproc.goodFeaturesToTrack(matOpFlowPrev, MOPcorners, iGFFTMax, 0.05, 20);
mMOP2fptsPrev.fromArray(MOPcorners.toArray());
// get safe copy of this corners
mMOP2fptsPrev.copyTo(mMOP2fptsSafe);
}
else
{
//Log.d("Baz", "Opflow");
// we've been through before so
// this mat is valid. Copy it to prev mat
matOpFlowThis.copyTo(matOpFlowPrev);
// get this mat
Imgproc.cvtColor(mRgba, matOpFlowThis, Imgproc.COLOR_RGBA2GRAY);
// get the corners for this mat
Imgproc.goodFeaturesToTrack(matOpFlowThis, MOPcorners, iGFFTMax, 0.05, 20);
mMOP2fptsThis.fromArray(MOPcorners.toArray());
// retrieve the corners from the prev mat
// (saves calculating them again)
mMOP2fptsSafe.copyTo(mMOP2fptsPrev);
// and save this corners for next time through
mMOP2fptsThis.copyTo(mMOP2fptsSafe);
}
/*
Parameters:
prevImg first 8-bit input image
nextImg second input image
prevPts vector of 2D points for which the flow needs to be found; point coordinates must be single-precision floating-point numbers.
nextPts output vector of 2D points (with single-precision floating-point coordinates) containing the calculated new positions of input features in the second image; when OPTFLOW_USE_INITIAL_FLOW flag is passed, the vector must have the same size as in the input.
status output status vector (of unsigned chars); each element of the vector is set to 1 if the flow for the corresponding features has been found, otherwise, it is set to 0.
err output vector of errors; each element of the vector is set to an error for the corresponding feature, type of the error measure can be set in flags parameter; if the flow wasn't found then the error is not defined (use the status parameter to find such cases).
*/
Video.calcOpticalFlowPyrLK(matOpFlowPrev, matOpFlowThis, mMOP2fptsPrev, mMOP2fptsThis, mMOBStatus, mMOFerr);
cornersPrev = mMOP2fptsPrev.toList();
cornersThis = mMOP2fptsThis.toList();
byteStatus = mMOBStatus.toList();
y = byteStatus.size() - 1;
for (x = 0; x < y; x++) {
if (byteStatus.get(x) == 1) {
pt = cornersThis.get(x);
pt2 = cornersPrev.get(x);
Core.circle(mRgba, pt, 5, colorRed, iLineThickness - 1);
Core.line(mRgba, pt, pt2, colorRed, iLineThickness);
}
}
return mRgba;
}