I have a sequence of images for which I want to calculate the median image (as to remove moving elements). Intuitively, hard-coding a loop to go through all the pixels would have a gross running time, as well as fairly large memory usage. Is there a way to easily do this in OpenCV? (I'm not interested in averaging, I need to do a median). I'm writing this for Android (using OpenCV4Android) so obviously computing power is limited.
As far as I know, there no OpenCV function that creates median image from sequence of images. I needed the same feature couple of years ago and I had to implement this myself. It is relatively slow because for each pixel you need to extract relevant pixel from multiple images (inefficient memory access) and calculate median (also a time consuming process).
Possible ways to increase efficiency are:
There no need to compute median from all images. Small subset of images will be enough.
You can find more efficient algorithms for finding median of some small groups. For example I used algorithm that can efficiently find median in group of nine values.
If the mean is ok:
Mat result(CV_64FC3, listImages[0].size());
for(int i = 0; i < listImages.size(); i++) {
result += listImages[i];
}
result /= listImages.size();
result.convertTo(result, CV_8UC3);
EDIT:
This quick pseudo-median should make the trick:
// Following algorithm will retain the pixel which is the closest to the mean
// Computing Mean
Mat tmpResult = Mat.zeros(listImages[0].size(), CV_64FC3);
for(int i = 0; i < listImages.size(); i++) {
tmpResult += listImages[i];
}
tmpResult /= listImages.size();
tmpResult.convertTo(tmpResult, CV_8UC3);
// We will now, for each pixel retain the closest to the mean
// Initializing result with the first image
Mat result(listImages[0].clone());
Mat diff1, diff2, minDiff;
for(int i = 1; i < listImages.size(); i++) {
// Computing diff between mean/newImage and mean/lastResult
absdiff(tmpResult, listImages[i], diff1);
absdiff(tmpResult, result, diff2);
// If a pixel of the new image is closer to the mean, it replaces the old one
min(diff1, diff2, minDiff);
// Get the old pixels that are still ok
result = result & ~(minDiff - diff2);
// Get the new pixels
result += listImages[i] & (minDiff - diff2);
}
However the classic one should be also pretty fast. It is O(nb^2 * w * h) where nb is the number of images and w, h their width, height. The above is O(nb * w * h) with more operations on Mats.
The code for the classical one (almost all computations will be made in native):
Mat tmp;
// We will sorting pixels where the first mat will get the lowest pixels and the last one, the highest
for(int i = 0; i < listImages.size(); i++) {
for(int j = i + 1; j < listImages.size(); j++) {
listImages[i].copyTo(tmp);
min(listImages[i], listImages[j], listImages[i]);
max(listImages[j], tmp, listImages[j]);
}
}
// We get the median
Mat result = listImages[listImages.size() / 2];
Related
I want to create a genetic algorithm that recreates images. I have created the program for this processing but the images that evolve are not anything close to the input image.
I believe that I have a problem with my fitness function. I have tried many things from changing the polygon types that are part of the DNA, I have tried to do both a crossover and a single parent, and I tried multiple fitness functions: histogram comparison across all channels, pixel comparison, brightness comparison(for black and white images).
public void calcFitness(PImage tar){
tar.loadPixels();
image.loadPixels();
int brightness = 0;
for(int i = 0; i < image.pixels.length;i++){
brightness += Math.abs(parent.brightness(tar.pixels[i])-parent.brightness(image.pixels[i]));
}
fitness = 1.0/ (Math.pow(1+brightness,2)/2);
}
public void calculateFitness(){
int[] rHist= new int[256], gHist= new int[256], bHist = new int[256];
image.loadPixels();
//Calculate Red Histogram
for(int i =0; i<image.pixels.length;i++) {
int red = image.pixels[i] >> 16 & 0xFF;
rHist[red]++;
}
//Calculate Green Histogram
for(int i =0; i<image.pixels.length;i++) {
int green = image.pixels[i] >> 8 & 0xFF;
gHist[green]++;
}
//Calculate Blue Histogram
for(int i =0; i<image.pixels.length;i++) {
int blue = image.pixels[i] & 0xFF;
bHist[blue]++;
}
//Compare the target histogram and the current one
for(int i = 0; i < 256; i++){
double totalDiff = 0;
totalDiff += Math.pow(main.rHist[i]-rHist[i],2)/2;
totalDiff += Math.pow(main.gHist[i]-gHist[i],2)/2;
totalDiff += Math.pow(main.bHist[i]-bHist[i],2)/2;
fitness+=Math.pow(1+totalDiff,-1);
}
}
public void evaluate(){
int totalFitness = 0;
for(int i = 0; i<POPULATION_SIZE;i++){
population[i].calcFitness(target);
//population[i].calculateFitness();
totalFitness+=population[i].fitness;
}
if(totalFitness>0) {
for (int i = 0; i < POPULATION_SIZE; i++) {
population[i].prob = population[i].fitness / totalFitness;
}
}
}
public void selection() {
SmartImage[] newPopulation = new SmartImage[POPULATION_SIZE];
for (int i = 0; i < POPULATION_SIZE; i++) {
DNA child;
DNA parentA = pickOne();
DNA parentB = pickOne();
child = parentA.crossover(parentB);
child.mutate(mutationRate);
newPopulation[i] = new SmartImage(parent, child, target.width, target.height);
}
population = newPopulation;
generation++;
}
What I expect from this is to get a general shape and color that is similar to my target image but all I get is random polygons with random colors and alphas.
The code looks fine at first glance. You should first check that your code is capable of converging to a target at all , for example by feeding a target image that is either generated by your algorithm with a random genome (or a very simple image that it should be easily recreated by your algorithm).
You are using the SAD (sum of absolute differences) metric between pixels to calculate fitness. You can try using SSD (sum of squared differences) like you are doing in the histogram difference method but between pixels or blocks, that will heavily penalize large differences so the remaining images won't be too different from the target. You can try using a more perceptual image space like HSV so the images will be closer visually even if they are farther in RGB space.
I think comparing the histogram of the entire image may be too lax, as there are many different images that will result in the same histogram. Comparing individual pixels may be too strict, the image needs to be aligned very precisely to get low differences, so everything gets low fitness values unless you are very lucky so the convergence will be too slow. I would recommend that you compare the histogram between overlapping blocks, and don't use all the 256 levels, use only about 16 levels or so (or use some kind of overlapping).
Read about Histogram of oriented gradients (HOG) and other similar techniques to get ideas to improve your fitness function. I took an online course about object recognition in images, Coursera - Deteccion de Objetos by the University of Barcelona but it's in Spanish. I'm pretty sure you can find similar study materials in English.
Edit: before trying something more complex a good idea would be doing the SAD or SSD on the average of each overlapping block (which would have a similar effect to strongly blurring the reference and generated images and then comparing the pixels, but faster). The fitness function should be resilient against small changes. An image that it's shifted by a few pixels or that is very similar after discarding the low-level detail should have much better fitness than a very different image and I think blurring will have that effect.
I am now building a project based on the sample color blob tracking method. I used bounding rectangles around the contours to indicate the blobs. Now I want to improve this algorithm by using an error correction method. What I do now is simply summing up the pixels in the rect region using elemsum method and calculate the average intensity and set it as the new blob detection parameter in each frame. However, the problem is that it is not accurate since those pixels outside the contour but inside the bounding rect will be counted as well. And the result is poor.
In order to solve the problem, I used another a straightforward way to loop through each pixel in the rectangle region (which is a submat), and set all pixel values out of range to the desired (or previous) hsv scalar. Then sum up all the pixels again and calculate the average intensity. This would much more accurate and easily solves the problem. The problem is that the program runs too slow on the phone (with around 1 frame per sec), though the result is accurate.
I found some sources online on how to do it in c++ using mat.forEach. I do not want to do the ndk thing and I would like to know if there is a more efficient way to do it in Java (Android).
UPDATE:
It turned out I can solve the problem by simply reducing the sampling rate. Instead of calculating the average intensity of all pixels, just a few number of them would do the job. My code:
for (int i=0; i< bounding_rect_hsv.rows();i+=10){
for (int j=0; j<bounding_rect_hsv.cols();j+=10){
double[] data = bounding_rect_hsv.get(i, j);
for (int k = 0; k < 3; k++){
if (data[k] > new_hsvColor.val[k] + 30 || data[k] < new_hsvColor.val[k] - 30) {
data[k] = new_hsvColor.val[k];
}
}
bounding_rect_hsv.put(i, j, data); //Puts element back into matrix
}
}
My source code:
Rect rect = Imgproc.boundingRect(points);
// draw enclosing rectangle (all same color, but you could use variable i to make them unique)
Imgproc.rectangle(original_frame, new Point(rect.x, rect.y), new Point(rect.x + rect.width, rect.y + rect.height), new Scalar(255, 0, 0, 255), 3);
//Todo: use the bounding rectangular to calculate average intensity (turn the pixels out of the contour to new_hsvColor)
//Just change the boundary values would be enough
bounding_rect_rgb = original_frame.submat(rect);
Imgproc.cvtColor(bounding_rect_rgb, bounding_rect_hsv, Imgproc.COLOR_RGB2HSV_FULL);
//Todo: change the logic so that pixels outside the contour will be changed to new_hsvColor
for (int i=0; i< bounding_rect_hsv.rows();i++){
for (int j=0; j<bounding_rect_hsv.cols();j++){
double[] data = bounding_rect_hsv.get(i, j);
for (int k = 0; k < 3; k++){
if (data[k] > new_hsvColor.val[k] + 30 || data[k] < new_hsvColor.val[k] - 30)
data[k] = new_hsvColor.val[k];
}
bounding_rect_hsv.put(i, j, data); //Puts element back into matrix
}
}
If you want to compute the mean value of pixels inside a contour you can simply:
Create a mask, using drawContours with parameter CV_FILLED and color Scalar(255) on a black (Scalar(0)) initialized CV_8UC1 image with same size as the original image.
Use mean to compute the mean of pixels under the mask.
You also don't need to convert to HSV every region (Rect), but you can convert the whole image once, and then access the desired region directly on the HSV image.
In the general case you want to sum the pixel values of a lot of rectangular regions, you may prefer to compute the integral image and compute the sum as the difference of values at bottom-right and top-left rectangle positions.
Im trying to decompile the convolution matrix for the filters on the Motorola Gallery App
Im using the following code to read the pixel data:
public static void main(String[] foo) {
new JavaWalkBufferedImageTest1();
}
public void printPixelARGB(int pixel) {
int alpha = (pixel >> 24) & 0xff;
int red = (pixel >> 16) & 0xff;
int green = (pixel >> 8) & 0xff;
int blue = (pixel) & 0xff;
System.out.println("argb: " + alpha + ", " + red + ", " + green + ", " + blue);
}
private void marchThroughImage(BufferedImage image) {
int w = image.getWidth();
int h = image.getHeight();
System.out.println("width, height: " + w + ", " + h);
for (int i = 0; i < h; i++) {
for (int j = 0; j < w; j++) {
System.out.println("x,y: " + j + ", " + i);
int pixel = image.getRGB(j, i);
printPixelARGB(pixel);
System.out.println("");
}
}
}
public JavaWalkBufferedImageTest1() {
try {
// get the BufferedImage, using the ImageIO class
BufferedImage image =
ImageIO.read(this.getClass().getResource("WhiteSpot.jpg"));
marchThroughImage(image);
} catch (IOException e) {
System.err.println(e.getMessage());
}
}
It gives the desired output.
But this is not leading me anywhere into finding the matrix.
How do I modify the code so that I can input 2 image files, Original & Filtered. And get the convolution matrix. Or is there an online tool that I can use, where I upload multiple Original & Filtered samples, and get the convolution matrix?
I would approach this problem (will work if only convolution was applied and only once!) like this:
first you need to find matrix size.
So you can loop through all possible matrix sizes or use a big one expecting zero values inside. That can be slow but you can try to estimate matrix size from the bluryness of sharp intensity edges (to how many pixels is the color bleeded).
for tested (big) matrix size try to find zero values
So for each tested matrix value form 2 arrays. One with few samples of pixels from convoluted image and second with pixels (shifted by tested matrix position) from original image. Now compute correlation coefficient between the two and if no significant correlation present you can assume the value in matrix is zero.
solve the remaining nonzero values in matrix
You can do this algebraically (form as many equation as many nonzero values you have (do not forget to choose pixels with different color). Or you can do search to minimize error/distance of convoluted output by your matrix and the real convoluted output. You can use something like mine approx class in C++ for this but if Your matrix is big then this will take a lot of time.
[Notes]
If more then just single convolution is applied to image then this will most likely not work.
Convolution matrices are mostly square and symmetric around mid value so you can compute the quarter of values and mirror the rest ... speeding up the computation.
Also see Dealing with filters and colour's which is similar question.
Bullet #2 can be used to detect the matrix size. As usually the max nonzero values are in the central cross of matrix.
0 0 2 0 0
0 1 2 1 0
2 2 3 2 2
0 1 2 1 0
0 0 2 0 0
Values 2 are the central cross and value 3 is the mid value of convolution matrix. So start computing the #2 correlations from mid point in x and y direction. And when hit zero value you are most likely at the edge of convolution matrix. So you can use that as matrix size (unless the matrix is some weird filter ...).
Also each color channel can have its own convolution matrix so may be you should do this separately per channel (and convolution can be done in any color space not just in RGB).
There might be other approaches for this (my guts tells me this should be solvable by PCA)
If you got black area in the original image then you can use that to find the values of matrix more easily (it will significantly ease up the algebraic approach)
I have code
public static void program() throws Exception{
BufferedImage input = null;
long start = System.currentTimeMillis();
while((System.currentTimeMillis() - start)/1000 < 220){
for (int i = 1; i < 13; i++){
for (int j = 1; j < 7; j++){
input = robot.createScreenCapture(new Rectangle(3+i*40, 127+j*40, 40, 40));
if ((input.getRGB(6, 3) > -7000000) && (input.getRGB(6, 3)<-5000000)){
robot.mouseMove(10+i*40, 137+j*40);
robot.mousePress(InputEvent.BUTTON1_MASK);
robot.mouseRelease(InputEvent.BUTTON1_MASK);
}
}
}
}
}
On a webpage there's a matrix (12*6) and there will randomly spawn some images. Some are bad, some are good.
I'm looking for a better way to check for good images. At the moment, on good images on location (6,3) the RGB color is different from bad images.
I'm making screenshot from every box (40 * 40) and looking at pixel in location (6,3)
Don't know how to explain my code any better
EDIT:
Picture of the webpage. External links ok?
http://i.imgur.com/B5Ev1Y0.png
I'm not sure what exactly the bottleneck is in your code, but I have a hunch it might be the repeated calls to robot.createScreenCapture.
You could try calling robot.createScreenCapture on the entire matrix (i.e. a large rectangle that covers all the smaller rectangles you are interested in) outside your nested loops, and then look up the pixel values at the points you are interested in using offsets for the x and y coordinates for the sub rectangles you are inspecting.
I have 2 Mat objects, overlay and background.
How do I put my overlay Mat on top of my background Mat such that only the non-transparent pixels of the overlay Mat completely obscures the background Mat?
I have tried addWeighted() which combines the 2 Mat but both "layers" are still visible.
The overlay Mat has a transparent channel while the background Mat does not.
The pixel in the overlay Mat is either completely transparent or fully obscure.
Both Mats are of the same size.
The function addWeighted won't work since it will use the same alpha value to all the pixels. To do exactly what you are saying, to only replace the non transparent values in the background, you can create a small function for that, like this:
cv::Mat blending(cv::Mat& overlay, cv::Mat& background){
//must have same size for this to work
assert(overlay.cols == background.cols && overlay.rows == background.rows);
cv::Mat result = background.clone();
for (int i = 0; i < result.rows; i++){
for (int j = 0; j < result.cols; j++){
cv::Vec4b pix = overlay.at<cv::Vec4b>(i,j);
if (pix[3] == 0){
result.at<cv::Vec3b>(i,j) = cv::Vec3b(pix[0], pix[1], pix[2]);
}
}
}
return result;
}
I am not sure if the transparent value in opencv is 0 or 255, so change it accordingly.... I think it is 0 for non-transparent adn 255 for fully transparent.
If you want to use the value of the alpha channel as a rate to blend, then change it a little to this:
cv::Mat blending(cv::Mat& overlay, cv::Mat& background){
//must have same size for this to work
assert(overlay.cols == background.cols && overlay.rows == background.rows);
cv::Mat result = background.clone();
for (int i = 0; i < result.rows; i++){
for (int j = 0; j < result.cols; j++){
cv::Vec4b pix = overlay.at<cv::Vec4b>(i,j);
double alphaRate = 1.0 - pix[3]/255.0;
result.at<cv::Vec3b>(i,j) = (1.0 - alphaRate) * cv::Vec3b(pix[0], pix[1], pix[2]) + result.at<cv::Vec3b>(i,j) * alphaRate;
}
}
return result;
}
Sorry for the code being in C++ and not in JAVA, but I think you can get an idea. Basically is just a loop in the pixels and changing the pixels in the copy of background to those of the overlay if they are not transparent.
* EDIT *
I will answer your comment with this edit, since it may take space. The problem is how OpenCV matrix works. For an image with alpha, the data is organized as an array like BGRA BGRA .... BGRA, and the basic operations like add, multiply and so on work in matrices with the same dimensions..... you can always try to separate the matrix with split (this will re write the matrix so it may be slow), then change the alpha channel to double (again, rewrite) and then do the multiplication and adding of the matrices. It should be faster since OpenCV optimizes these functions.... also you can do this in GPU....
Something like this:
cv::Mat blending(cv::Mat& overlay, cv::Mat& background){
std::vector<cv::Mat> channels;
cv::split(overlay, channels);
channels[3].convertTo(channels[3], CV_64F, 1.0/255.0);
cv::Mat newOverlay, result;
cv::merge(channels, newOverlay);
result = newOverlay * channels[3] + ((1 - channels[3]) * background);
return result;
}
Not sure if OpenCV allows a CV_8U to multiply a CV_64F, or if this will be faster or not.... but it may be.
Also, the ones with loops has no problem in threads, so it can be optimized... running this in release mode will greatly increase the speed too since the .at function of OpenCV does several asserts.... that in release mode are not done. Not sure if this can be change in JAVA though...
I was able to port api55's edited answer for java:
private void merge(Mat background, Mat overlay) {
List<Mat> backgroundChannels = new ArrayList<>();
Core.split(background, backgroundChannels);
List<Mat> overlayChannels = new ArrayList<>();
Core.split(overlay, overlayChannels);
// compute "alphaRate = 1 - overlayAlpha / 255"
Mat overlayAlphaChannel = overlayChannels.get(3);
Mat alphaRate = new Mat(overlayAlphaChannel.size(), overlayAlphaChannel.type());
Core.divide(overlayAlphaChannel, new Scalar(255), alphaRate);
Core.absdiff(alphaRate, new Scalar(1), alphaRate);
for (int i = 0; i < 3; i++) {
// compute "(1 - alphaRate) * overlay"
Mat overlayChannel = overlayChannels.get(i);
Mat temp = new Mat(alphaRate.size(), alphaRate.type());
Core.absdiff(alphaRate, new Scalar(1), temp);
Core.multiply(temp, overlayChannel, overlayChannel);
temp.release();
// compute "background * alphaRate"
Mat backgroundChannel = backgroundChannels.get(i);
Core.multiply(backgroundChannel, alphaRate, backgroundChannel);
// compute the merged channel
Core.add(backgroundChannel, overlayChannel, backgroundChannel);
}
alphaRate.release();
Core.merge(backgroundChannels, background);
}
it is a lot faster compared to the double nested loop calculation.