Java Code Examples for org.opencv.core.Core#subtract()

public static Mat transEstimate(Mat img, int patchSz, double[] airlight, double lambda, double fTrans, 
		int r, double eps, double gamma) {
	int rows = img.rows();
	int cols = img.cols();
	List<Mat> bgr = new ArrayList<>();
	Core.split(img, bgr);
	int type = bgr.get(0).type();
	// calculate the transmission map
	Mat T = computeTrans(img, patchSz, rows, cols, type, airlight, lambda, fTrans);
	// refine the transmission map
	img.convertTo(img, CvType.CV_8UC1);
	Mat gray = new Mat();
	Imgproc.cvtColor(img, gray, Imgproc.COLOR_BGR2GRAY);
	gray.convertTo(gray, CvType.CV_32F);
	Core.divide(gray, new Scalar(255.0), gray);
	T = Filters.GuidedImageFilter(gray, T, r, eps);
	Mat Tsmooth = new Mat();
	Imgproc.GaussianBlur(T, Tsmooth, new Size(81, 81), 40);
	Mat Tdetails = new Mat();
	Core.subtract(T, Tsmooth, Tdetails);
	Core.multiply(Tdetails, new Scalar(gamma), Tdetails);
	Core.add(Tsmooth, Tdetails, T);
	return T;
}
 

Mat buildSkeleton(Mat img){
    Mat morph = Imgproc.getStructuringElement(Imgproc.CV_SHAPE_CROSS, new Size(3, 3));
    Mat skel = new Mat(img.size(), CvType.CV_8UC1, Scalar.all(0));
    Mat eroded = new Mat();
    Mat temp = new Mat();

    boolean done = false;

    do{
        Imgproc.morphologyEx(img, eroded, Imgproc.MORPH_ERODE, morph);
        Imgproc.morphologyEx(eroded, temp, Imgproc.MORPH_DILATE, morph);
        Core.subtract(img, temp, temp);
        Core.bitwise_or(skel, temp, skel);
        eroded.copyTo(img);

        done = Core.countNonZero(img) == 0;
    }while (!done);

    return skel;
}
 

public static Mat LocalContrast(Mat img) {
	double[] h = { 1.0 / 16.0, 4.0 / 16.0, 6.0 / 16.0, 4.0 / 16.0, 1.0 / 16.0 };
	Mat mask = new Mat(h.length, h.length, img.type());
	for (int i = 0; i < h.length; i++) {
		for (int j = 0; j < h.length; j++) {
			mask.put(i, j, h[i] * h[j]);
		}
	}
	Mat localContrast = new Mat();
	Imgproc.filter2D(img, localContrast, img.depth(), mask);
	for (int i = 0; i < localContrast.rows(); i++) {
		for (int j = 0; j < localContrast.cols(); j++) {
			if (localContrast.get(i, j)[0] > Math.PI / 2.75) localContrast.put(i, j, Math.PI / 2.75);
		}
	}
	Core.subtract(img, localContrast, localContrast);
	return localContrast.mul(localContrast);
}
 

public String recognize(Mat img, String expectedLabel){
    // Ignore
    img = img.reshape(1,1);
    // Subtract mean
    img.convertTo(img, CvType.CV_32F);
    Core.subtract(img, Psi, img);
    // Project to subspace
    Mat projected = getFeatureVector(img);
    // Save all points of image for tSNE
    img.convertTo(img, CvType.CV_8U);
    addImage(projected, expectedLabel, true);
    //addImage(projected, expectedLabel);
    Mat distance = new Mat(Omega.rows(), 1, CvType.CV_64FC1);
    for (int i=0; i<Omega.rows(); i++){
        double dist = Core.norm(projected.row(0), Omega.row(i), Core.NORM_L2);
        distance.put(i, 0, dist);
    }
    Mat sortedDist = new Mat(Omega.rows(), 1, CvType.CV_8UC1);
    Core.sortIdx(distance, sortedDist, Core.SORT_EVERY_COLUMN + Core.SORT_ASCENDING);
    // Give back the name of the found person
    int index = (int)(sortedDist.get(0,0)[0]);
    return labelMap.getKey(labelList.get(index));
}
 

public static Mat[] LaplacianPyramid(Mat img, int level) {
	Mat[] lapPyr = new Mat[level];
	//Mat mask = filterMask(img);
	lapPyr[0] = img.clone();
	Mat tmpImg = img.clone();
	for (int i = 1; i < level; i++) {
		// resize image
		Imgproc.resize(tmpImg, tmpImg, new Size(), 0.5, 0.5, Imgproc.INTER_LINEAR);
		lapPyr[i] = tmpImg.clone();
	}
	// calculate the DoG
	for (int i = 0; i < level - 1; i++) {
		Mat tmpPyr = new Mat();
		Imgproc.resize(lapPyr[i + 1], tmpPyr, lapPyr[i].size(), 0, 0, Imgproc.INTER_LINEAR);
		Core.subtract(lapPyr[i], tmpPyr, lapPyr[i]);
	}
	return lapPyr;
}
 

public static Mat LocalContrast(Mat img) {
	double[] h = { 1.0 / 16.0, 4.0 / 16.0, 6.0 / 16.0, 4.0 / 16.0, 1.0 / 16.0 };
	Mat mask = new Mat(h.length, h.length, img.type());
	for (int i = 0; i < h.length; i++) {
		for (int j = 0; j < h.length; j++) {
			mask.put(i, j, h[i] * h[j]);
		}
	}
	Mat localContrast = new Mat();
	Imgproc.filter2D(img, localContrast, img.depth(), mask);
	for (int i = 0; i < localContrast.rows(); i++) {
		for (int j = 0; j < localContrast.cols(); j++) {
			if (localContrast.get(i, j)[0] > Math.PI / 2.75)
				localContrast.put(i, j, Math.PI / 2.75);
		}
	}
	Core.subtract(img, localContrast, localContrast);
	return localContrast.mul(localContrast);
}
 

private static Mat preDehaze(Mat img, double a, double nTrans) {
	// nOut = ( (blkIm - a) * nTrans + 128 * a ) / 128;
	Core.subtract(img, new Scalar(a), img);
	Core.multiply(img, new Scalar(nTrans), img);
	Core.add(img, new Scalar(128.0 * a), img);
	Core.divide(img, new Scalar(128.0), img);
	return img;
}
 

/**
 * 其主要思路为：
	1、求取源图I的平均灰度，并记录rows和cols；
	2、按照一定大小，分为N*M个方块，求出每块的平均值，得到子块的亮度矩阵D；
	3、用矩阵D的每个元素减去源图的平均灰度，得到子块的亮度差值矩阵E；
	4、用双立方差值法，将矩阵E差值成与源图一样大小的亮度分布矩阵R；
	5、得到矫正后的图像result=I-R；
* @Title: unevenLightCompensate 
* @Description: 光线补偿 
* @param image
* @param blockSize
* void 
* @throws
 */
public static void unevenLightCompensate(Mat image, int blockSize) {
	if(image.channels() == 3) {
		Imgproc.cvtColor(image, image, 7);
	}
	double average = Core.mean(image).val[0];
	Scalar scalar = new Scalar(average);
	int rowsNew = (int) Math.ceil((double)image.rows() / (double)blockSize);
	int colsNew = (int) Math.ceil((double)image.cols() / (double)blockSize);
	Mat blockImage = new Mat();
	blockImage = Mat.zeros(rowsNew, colsNew, CvType.CV_32FC1);
	for(int i = 0; i < rowsNew; i ++) {
		for(int j = 0; j < colsNew; j ++) {
			int rowmin = i * blockSize;
			int rowmax = (i + 1) * blockSize;
			if(rowmax > image.rows()) rowmax = image.rows();
			int colmin = j * blockSize;
			int colmax = (j +1) * blockSize;
			if(colmax > image.cols()) colmax = image.cols();
			Range rangeRow = new Range(rowmin, rowmax);
			Range rangeCol = new Range(colmin, colmax);
			Mat imageROI = new Mat(image, rangeRow, rangeCol);
			double temaver = Core.mean(imageROI).val[0];
			blockImage.put(i, j, temaver);
		}
	}
	
	Core.subtract(blockImage, scalar, blockImage);
	Mat blockImage2 = new Mat();
	int INTER_CUBIC = 2;
	Imgproc.resize(blockImage, blockImage2, image.size(), 0, 0, INTER_CUBIC);
	Mat image2 = new Mat();
	image.convertTo(image2, CvType.CV_32FC1);
	Mat dst = new Mat();
	Core.subtract(image2, blockImage2, dst);
	dst.convertTo(image, CvType.CV_8UC1);
}
 

public Eyes getEyes(Mat img){
    double halfWidth = img.cols() / 2;
    double height = img.rows();
    double[] values = new double[4];
    values[0] = 0;
    values[1] = 0;
    values[2] = halfWidth;
    values[3] = height;
    Rect rightHalf = new Rect(values);
    values[0] = halfWidth;
    Rect leftHalf = new Rect(values);
    MatOfRect rightEyes = new MatOfRect();
    MatOfRect leftEyes = new MatOfRect();

    Mat rightHalfImg = img.submat(rightHalf);
    rightEyeDetector.detectMultiScale(rightHalfImg, rightEyes);
    Mat leftHalfImg = img.submat(leftHalf);
    leftEyeDetector.detectMultiScale(leftHalfImg, leftEyes);

    if (rightEyes.empty() || leftEyes.empty() || rightEyes.toArray().length > 1 || leftEyes.toArray().length > 1){
        return null;
    }

    Rect rightEye = rightEyes.toArray()[0];
    Rect leftEye = leftEyes.toArray()[0];

    MatOfFloat rightPoint = new MatOfFloat(rightEye.x + rightEye.width / 2, rightEye.y + rightEye.height / 2);
    MatOfFloat leftPoint = new MatOfFloat(img.cols() / 2 + leftEye.x + leftEye.width / 2, leftEye.y + leftEye.height / 2);

    MatOfFloat diff = new MatOfFloat();
    Core.subtract(leftPoint, rightPoint, diff);
    double angle = Core.fastAtan2(diff.toArray()[1], diff.toArray()[0]);
    double dist = Core.norm(leftPoint, rightPoint, Core.NORM_L2);
    Eyes eyes = new Eyes(dist, rightPoint, leftPoint, angle);
    return eyes;
}
 

public static Mat Saliency(Mat img) {
	// blur image with a 3x3 or 5x5 Gaussian filter
	Mat gfbgr = new Mat();
	Imgproc.GaussianBlur(img, gfbgr, new Size(3, 3), 3);
	// Perform sRGB to CIE Lab color space conversion
	Mat LabIm = new Mat();
	Imgproc.cvtColor(gfbgr, LabIm, Imgproc.COLOR_BGR2Lab);
	// Compute Lab average values (note that in the paper this average is found from the
	// un-blurred original image, but the results are quite similar)
	List<Mat> lab = new ArrayList<Mat>();
	Core.split(LabIm, lab);
	Mat l = lab.get(0);
	l.convertTo(l, CvType.CV_32F);
	Mat a = lab.get(1);
	a.convertTo(a, CvType.CV_32F);
	Mat b = lab.get(2);
	b.convertTo(b, CvType.CV_32F);
	double lm = Core.mean(l).val[0];
	double am = Core.mean(a).val[0];
	double bm = Core.mean(b).val[0];
	// Finally compute the saliency map
	Mat sm = Mat.zeros(l.rows(), l.cols(), l.type());
	Core.subtract(l, new Scalar(lm), l);
	Core.subtract(a, new Scalar(am), a);
	Core.subtract(b, new Scalar(bm), b);
	Core.add(sm, l.mul(l), sm);
	Core.add(sm, a.mul(a), sm);
	Core.add(sm, b.mul(b), sm);
	return sm;
}
 

private static Mat dehazeProcess(Mat img, Mat trans, double[] airlight) {
	Mat balancedImg = Filters.SimplestColorBalance(img, 5);
	Mat bCnl = new Mat();
	Core.extractChannel(balancedImg, bCnl, 0);
	Mat gCnl = new Mat();
	Core.extractChannel(balancedImg, gCnl, 1);
	Mat rCnl = new Mat();
	Core.extractChannel(balancedImg, rCnl, 2);
	// get mean value
	double bMean = Core.mean(bCnl).val[0];
	double gMean = Core.mean(gCnl).val[0];
	double rMean = Core.mean(rCnl).val[0];
	// get transmission map for each channel
	Mat Tb = trans.clone();
	Core.multiply(Tb, new Scalar(Math.max(bMean, Math.max(gMean, rMean)) / bMean * 0.8), Tb);
	Mat Tg = trans.clone();
	Core.multiply(Tg, new Scalar(Math.max(bMean, Math.max(gMean, rMean)) / gMean * 0.9), Tg);
	Mat Tr = trans.clone();
	Core.multiply(Tr, new Scalar(Math.max(bMean, Math.max(gMean, rMean)) / rMean * 0.8), Tr);
	// dehaze by formula
	// blue channel
	Mat bChannel = new Mat();
	Core.subtract(bCnl, new Scalar(airlight[0]), bChannel);
	Core.divide(bChannel, Tb, bChannel);
	Core.add(bChannel, new Scalar(airlight[0]), bChannel);
	// green channel
	Mat gChannel = new Mat();
	Core.subtract(gCnl, new Scalar(airlight[1]), gChannel);
	Core.divide(gChannel, Tg, gChannel);
	Core.add(gChannel, new Scalar(airlight[1]), gChannel);
	// red channel
	Mat rChannel = new Mat();
	Core.subtract(rCnl, new Scalar(airlight[2]), rChannel);
	Core.divide(rChannel, Tr, rChannel);
	Core.add(rChannel, new Scalar(airlight[2]), rChannel);
	Mat dehazed = new Mat();
	Core.merge(new ArrayList<>(Arrays.asList(bChannel, gChannel, rChannel)), dehazed);
	return dehazed;
}
 

private static Mat dehaze(Mat channel, Mat t, double minAtmosLight) {
	Mat t_ = new Mat();
	Core.subtract(t, new Scalar(1.0), t_);
	Core.multiply(t_, new Scalar(-1.0 * minAtmosLight), t_);
	Core.subtract(channel, t_, channel);
	Core.divide(channel, t, channel);
	return channel;
}
 

private static Mat dehaze(Mat img, Mat T, double airlight) {
	// J = (img - airlight) ./ T + airlight;
	Core.subtract(img, new Scalar(airlight), img);
	Core.divide(img, T, img);
	Core.add(img, new Scalar(airlight), img);
	return img;
}
 

@SuppressWarnings("unused")
public static Mat enhanceEachChannel(Mat image, int blkSize, int patchSize, double lambda, double eps, int krnlSize) {
	image.convertTo(image, CvType.CV_32F);
	// split image to three channels
	List<Mat> bgr = new ArrayList<>();
	Core.split(image, bgr);
	Mat bChannel = bgr.get(0);
	Mat gChannel = bgr.get(1);
	Mat rChannel = bgr.get(2);
	// obtain air-light
	double[] airlight = AirlightEstimate.estimate(image, blkSize);
	// obtain coarse transmission map and refine it for each channel
	double fTrans = 0.3;
	Mat T = TransmissionEstimate.transEstimateEachChannel(bChannel, patchSize, airlight[0], lambda, fTrans);
	Core.subtract(T, new Scalar(1.0), T);
	Core.multiply(T, new Scalar(-1.0), T);
	Mat Tb = Filters.GuidedImageFilter(bChannel, T, krnlSize, eps);
	T = TransmissionEstimate.transEstimateEachChannel(gChannel, patchSize, airlight[1], lambda, fTrans);
	Core.subtract(T, new Scalar(1.0), T);
	Core.multiply(T, new Scalar(-1.0), T);
	Mat Tg = Filters.GuidedImageFilter(gChannel, T, krnlSize, eps);
	T = TransmissionEstimate.transEstimateEachChannel(rChannel, patchSize, airlight[2], lambda, fTrans);
	Core.subtract(T, new Scalar(1.0), T);
	Core.multiply(T, new Scalar(-1.0), T);
	Mat Tr = Filters.GuidedImageFilter(rChannel, T, krnlSize, eps);
	// dehaze
	bChannel = dehaze(bChannel, Tb, airlight[0]);
	gChannel = dehaze(gChannel, Tg, airlight[1]);
	rChannel = dehaze(rChannel, Tr, airlight[2]);
	Mat outval = new Mat();
	Core.merge(new ArrayList<>(Arrays.asList(bChannel, gChannel, rChannel)), outval);
	return outval;
}
 

public static Mat enhance(Mat image, double krnlRatio, double minAtmosLight, double eps) {
	image.convertTo(image, CvType.CV_32F);
	// extract each color channel
	List<Mat> rgb = new ArrayList<>();
	Core.split(image, rgb);
	Mat rChannel = rgb.get(0);
	Mat gChannel = rgb.get(1);
	Mat bChannel = rgb.get(2);
	int rows = rChannel.rows();
	int cols = rChannel.cols();
	// derive the dark channel from original image
	Mat dc = rChannel.clone();
	for (int i = 0; i < image.rows(); i++) {
		for (int j = 0; j < image.cols(); j++) {
			double min = Math.min(rChannel.get(i, j)[0], Math.min(gChannel.get(i, j)[0], bChannel.get(i, j)[0]));
			dc.put(i, j, min);
		}
	}
	// minimum filter
	int krnlSz = Double.valueOf(Math.max(Math.max(rows * krnlRatio, cols * krnlRatio), 3.0)).intValue();
	Mat kernel = Imgproc.getStructuringElement(Imgproc.MORPH_RECT, new Size(krnlSz, krnlSz), new Point(-1, -1));
	Imgproc.erode(dc, dc, kernel);
	// get coarse transmission map
	Mat t = dc.clone();
	Core.subtract(t, new Scalar(255.0), t);
	Core.multiply(t, new Scalar(-1.0), t);
	Core.divide(t, new Scalar(255.0), t);
	// obtain gray scale image
	Mat gray = new Mat();
	Imgproc.cvtColor(image, gray, Imgproc.COLOR_RGB2GRAY);
	Core.divide(gray, new Scalar(255.0), gray);
	// refine transmission map
	int r = krnlSz * 4;
	t = Filters.GuidedImageFilter(gray, t, r, eps);
	// get minimum atmospheric light
	minAtmosLight = Math.min(minAtmosLight, Core.minMaxLoc(dc).maxVal);
	// dehaze each color channel
	rChannel = dehaze(rChannel, t, minAtmosLight);
	gChannel = dehaze(gChannel, t, minAtmosLight);
	bChannel = dehaze(bChannel, t, minAtmosLight);
	// merge three color channels to a image
	Mat outval = new Mat();
	Core.merge(new ArrayList<>(Arrays.asList(rChannel, gChannel, bChannel)), outval);
	outval.convertTo(outval, CvType.CV_8UC1);
	return outval;
}
 

/**
 * Guided Image Filter for grayscale image, O(1) time implementation of guided filter
 *
 * @param I
 *            guidance image (should be a gray-scale/single channel image)
 * @param p
 *            filtering input image (should be a gray-scale/single channel
 *            image)
 * @param r
 *            local window radius
 * @param eps
 *            regularization parameter
 * @return filtered image
 */
public static Mat GuidedImageFilter(Mat I, Mat p, int r, double eps) {
	I.convertTo(I, CvType.CV_64FC1);
	p.convertTo(p, CvType.CV_64FC1);
	//[hei, wid] = size(I);
	int rows = I.rows();
	int cols = I.cols();
	// N = boxfilter(ones(hei, wid), r); % the size of each local patch; N=(2r+1)^2 except for boundary pixels.
	Mat N = new Mat();
	Imgproc.boxFilter(Mat.ones(rows, cols, I.type()), N, -1, new Size(r, r));
	// mean_I = boxfilter(I, r) ./ N;
	Mat mean_I = new Mat();
	Imgproc.boxFilter(I, mean_I, -1, new Size(r, r));
	// mean_p = boxfilter(p, r) ./ N
	Mat mean_p = new Mat();
	Imgproc.boxFilter(p, mean_p, -1, new Size(r, r));
	// mean_Ip = boxfilter(I.*p, r) ./ N;
	Mat mean_Ip = new Mat();
	Imgproc.boxFilter(I.mul(p), mean_Ip, -1, new Size(r, r));
	// cov_Ip = mean_Ip - mean_I .* mean_p; % this is the covariance of (I, p) in each local patch.
	Mat cov_Ip = new Mat();
	Core.subtract(mean_Ip, mean_I.mul(mean_p), cov_Ip);
	// mean_II = boxfilter(I.*I, r) ./ N;
	Mat mean_II = new Mat();
	Imgproc.boxFilter(I.mul(I), mean_II, -1, new Size(r, r));
	// var_I = mean_II - mean_I .* mean_I;
	Mat var_I = new Mat();
	Core.subtract(mean_II, mean_I.mul(mean_I), var_I);
	// a = cov_Ip ./ (var_I + eps); % Eqn. (5) in the paper;
	Mat a = new Mat();
	Core.add(var_I, new Scalar(eps), a);
	Core.divide(cov_Ip, a, a);
	//b = mean_p - a .* mean_I; % Eqn. (6) in the paper;
	Mat b = new Mat();
	Core.subtract(mean_p, a.mul(mean_I), b);
	// mean_a = boxfilter(a, r) ./ N;
	Mat mean_a = new Mat();
	Imgproc.boxFilter(a, mean_a, -1, new Size(r, r));
	Core.divide(mean_a, N, mean_a);
	// mean_b = boxfilter(b, r) ./ N;
	Mat mean_b = new Mat();
	Imgproc.boxFilter(b, mean_b, -1, new Size(r, r));
	Core.divide(mean_b, N, mean_b);
	// q = mean_a .* I + mean_b; % Eqn. (8) in the paper;
	Mat q = new Mat();
	Core.add(mean_a.mul(I), mean_b, q);
	//for (int i = 0; i < rows; i++) {
	//	for (int j = 0; j < cols; j++) {
	//		if (q.get(i, j)[0] <= 0)
	//			q.put(i, j, 1.0 / 255);
	//	}
	//}
	q.convertTo(q, CvType.CV_32F);
	return q;
}
 

private static Mat filterSingleChannel(Mat p, double s, ArrayList<Mat> Isubchannels, ArrayList<Mat> Ichannels, 
		Mat mean_I_r, Mat mean_I_g, Mat mean_I_b, Mat invrr, Mat invrg, Mat invrb, Mat invgg, Mat invgb, 
		Mat invbb, double r_sub) {
	Mat p_sub = new Mat();
	Imgproc.resize(p, p_sub, new Size(p.cols() / s, p.rows() / s), 0.0, 0.0, Imgproc.INTER_NEAREST);

	Mat mean_p = boxfilter(p_sub, (int) r_sub);

	Mat mean_Ip_r = boxfilter(Isubchannels.get(0).mul(p_sub), (int) r_sub);
	Mat mean_Ip_g = boxfilter(Isubchannels.get(1).mul(p_sub), (int) r_sub);
	Mat mean_Ip_b = boxfilter(Isubchannels.get(2).mul(p_sub), (int) r_sub);

	// convariance of (I, p) in each local patch
	Mat cov_Ip_r = new Mat();
	Mat cov_Ip_g = new Mat();
	Mat cov_Ip_b = new Mat();
	Core.subtract(mean_Ip_r, mean_I_r.mul(mean_p), cov_Ip_r);
	Core.subtract(mean_Ip_g, mean_I_g.mul(mean_p), cov_Ip_g);
	Core.subtract(mean_Ip_b, mean_I_b.mul(mean_p), cov_Ip_b);

	Mat temp1 = new Mat();
	Mat a_r = new Mat();
	Mat a_g = new Mat();
	Mat a_b = new Mat();
	Core.add(invrr.mul(cov_Ip_r), invrg.mul(cov_Ip_g), temp1);
	Core.add(temp1, invrb.mul(cov_Ip_b), a_r);
	Core.add(invrg.mul(cov_Ip_r), invgg.mul(cov_Ip_g), temp1);
	Core.add(temp1, invgb.mul(cov_Ip_b), a_g);
	Core.add(invrb.mul(cov_Ip_r), invgb.mul(cov_Ip_g), temp1);
	Core.add(temp1, invbb.mul(cov_Ip_b), a_b);

	Mat b = new Mat();
	Core.subtract(mean_p, a_r.mul(mean_I_r), b);
	Core.subtract(b, a_g.mul(mean_I_g), b);
	Core.subtract(b, a_b.mul(mean_I_b), b);

	Mat mean_a_r = boxfilter(a_r, (int) r_sub);
	Mat mean_a_g = boxfilter(a_g, (int) r_sub);
	Mat mean_a_b = boxfilter(a_b, (int) r_sub);
	Mat mean_b = boxfilter(b, (int) r_sub);

	Imgproc.resize(mean_a_r, mean_a_r, 
			new Size(Ichannels.get(0).cols(), Ichannels.get(0).rows()), 0.0, 0.0, Imgproc.INTER_LINEAR);
	Imgproc.resize(mean_a_g, mean_a_g, 
			new Size(Ichannels.get(0).cols(), Ichannels.get(0).rows()), 0.0, 0.0, Imgproc.INTER_LINEAR);
	Imgproc.resize(mean_a_b, mean_a_b, 
			new Size(Ichannels.get(0).cols(), Ichannels.get(0).rows()), 0.0, 0.0, Imgproc.INTER_LINEAR);
	Imgproc.resize(mean_b, mean_b, 
			new Size(Ichannels.get(0).cols(), Ichannels.get(0).rows()), 0.0, 0.0, Imgproc.INTER_LINEAR);

	Mat result = new Mat();
	Core.add(mean_a_r.mul(Ichannels.get(0)), mean_a_g.mul(Ichannels.get(1)), temp1);
	Core.add(temp1, mean_a_b.mul(Ichannels.get(2)), temp1);
	Core.add(temp1, mean_b, result);
	return result;
}
 

private static Mat filterSingleChannel(Mat p, double s, ArrayList<Mat> Isubchannels, ArrayList<Mat> Ichannels, 
		Mat mean_I_r, Mat mean_I_g, Mat mean_I_b, Mat invrr, Mat invrg, Mat invrb, Mat invgg, Mat invgb, 
		Mat invbb, double r_sub) {
	Mat p_sub = new Mat();
	Imgproc.resize(p, p_sub, new Size(p.cols() / s, p.rows() / s), 0.0, 0.0, Imgproc.INTER_NEAREST);

	Mat mean_p = boxfilter(p_sub, (int) r_sub);

	Mat mean_Ip_r = boxfilter(Isubchannels.get(0).mul(p_sub), (int) r_sub);
	Mat mean_Ip_g = boxfilter(Isubchannels.get(1).mul(p_sub), (int) r_sub);
	Mat mean_Ip_b = boxfilter(Isubchannels.get(2).mul(p_sub), (int) r_sub);

	// convariance of (I, p) in each local patch
	Mat cov_Ip_r = new Mat();
	Mat cov_Ip_g = new Mat();
	Mat cov_Ip_b = new Mat();
	Core.subtract(mean_Ip_r, mean_I_r.mul(mean_p), cov_Ip_r);
	Core.subtract(mean_Ip_g, mean_I_g.mul(mean_p), cov_Ip_g);
	Core.subtract(mean_Ip_b, mean_I_b.mul(mean_p), cov_Ip_b);

	Mat temp1 = new Mat();
	Mat a_r = new Mat();
	Mat a_g = new Mat();
	Mat a_b = new Mat();
	Core.add(invrr.mul(cov_Ip_r), invrg.mul(cov_Ip_g), temp1);
	Core.add(temp1, invrb.mul(cov_Ip_b), a_r);
	Core.add(invrg.mul(cov_Ip_r), invgg.mul(cov_Ip_g), temp1);
	Core.add(temp1, invgb.mul(cov_Ip_b), a_g);
	Core.add(invrb.mul(cov_Ip_r), invgb.mul(cov_Ip_g), temp1);
	Core.add(temp1, invbb.mul(cov_Ip_b), a_b);

	Mat b = new Mat();
	Core.subtract(mean_p, a_r.mul(mean_I_r), b);
	Core.subtract(b, a_g.mul(mean_I_g), b);
	Core.subtract(b, a_b.mul(mean_I_b), b);

	Mat mean_a_r = boxfilter(a_r, (int) r_sub);
	Mat mean_a_g = boxfilter(a_g, (int) r_sub);
	Mat mean_a_b = boxfilter(a_b, (int) r_sub);
	Mat mean_b = boxfilter(b, (int) r_sub);

	Imgproc.resize(mean_a_r, mean_a_r, 
			new Size(Ichannels.get(0).cols(), Ichannels.get(0).rows()), 0.0, 0.0, Imgproc.INTER_LINEAR);
	Imgproc.resize(mean_a_g, mean_a_g, 
			new Size(Ichannels.get(0).cols(), Ichannels.get(0).rows()), 0.0, 0.0, Imgproc.INTER_LINEAR);
	Imgproc.resize(mean_a_b, mean_a_b, 
			new Size(Ichannels.get(0).cols(), Ichannels.get(0).rows()), 0.0, 0.0, Imgproc.INTER_LINEAR);
	Imgproc.resize(mean_b, mean_b, 
			new Size(Ichannels.get(0).cols(), Ichannels.get(0).rows()), 0.0, 0.0, Imgproc.INTER_LINEAR);

	Mat result = new Mat();
	Core.add(mean_a_r.mul(Ichannels.get(0)), mean_a_g.mul(Ichannels.get(1)), temp1);
	Core.add(temp1, mean_a_b.mul(Ichannels.get(2)), temp1);
	Core.add(temp1, mean_b, result);
	return result;
}
 

/**
 * Guided Image Filter for grayscale image, O(1) time implementation of guided filter
 *
 * @param I guidance image (should be a gray-scale/single channel image)
 * @param p filtering input image (should be a gray-scale/single channel image)
 * @param r local window radius
 * @param eps regularization parameter
 * @return filtered image
 */
public static Mat GuidedImageFilter(Mat I, Mat p, int r, double eps) {
	I.convertTo(I, CvType.CV_64FC1);
	p.convertTo(p, CvType.CV_64FC1);
	//[hei, wid] = size(I);
	int rows = I.rows();
	int cols = I.cols();
	// N = boxfilter(ones(hei, wid), r); % the size of each local patch; N=(2r+1)^2 except for boundary pixels.
	Mat N = new Mat();
	Imgproc.boxFilter(Mat.ones(rows, cols, I.type()), N, -1, new Size(r, r));
	// mean_I = boxfilter(I, r) ./ N;
	Mat mean_I = new Mat();
	Imgproc.boxFilter(I, mean_I, -1, new Size(r, r));
	// mean_p = boxfilter(p, r) ./ N
	Mat mean_p = new Mat();
	Imgproc.boxFilter(p, mean_p, -1, new Size(r, r));
	// mean_Ip = boxfilter(I.*p, r) ./ N;
	Mat mean_Ip = new Mat();
	Imgproc.boxFilter(I.mul(p), mean_Ip, -1, new Size(r, r));
	// cov_Ip = mean_Ip - mean_I .* mean_p; % this is the covariance of (I, p) in each local patch.
	Mat cov_Ip = new Mat();
	Core.subtract(mean_Ip, mean_I.mul(mean_p), cov_Ip);
	// mean_II = boxfilter(I.*I, r) ./ N;
	Mat mean_II = new Mat();
	Imgproc.boxFilter(I.mul(I), mean_II, -1, new Size(r, r));
	// var_I = mean_II - mean_I .* mean_I;
	Mat var_I = new Mat();
	Core.subtract(mean_II, mean_I.mul(mean_I), var_I);
	// a = cov_Ip ./ (var_I + eps); % Eqn. (5) in the paper;
	Mat a = new Mat();
	Core.add(var_I, new Scalar(eps), a);
	Core.divide(cov_Ip, a, a);
	//b = mean_p - a .* mean_I; % Eqn. (6) in the paper;
	Mat b = new Mat();
	Core.subtract(mean_p, a.mul(mean_I), b);
	// mean_a = boxfilter(a, r) ./ N;
	Mat mean_a = new Mat();
	Imgproc.boxFilter(a, mean_a, -1, new Size(r, r));
	Core.divide(mean_a, N, mean_a);
	// mean_b = boxfilter(b, r) ./ N;
	Mat mean_b = new Mat();
	Imgproc.boxFilter(b, mean_b, -1, new Size(r, r));
	Core.divide(mean_b, N, mean_b);
	// q = mean_a .* I + mean_b; % Eqn. (8) in the paper;
	Mat q = new Mat();
	Core.add(mean_a.mul(I), mean_b, q);
	q.convertTo(q, CvType.CV_32F);
	return q;
}
 

private void computePhi(){
    Mat Psi_repeated = new Mat();
    Core.repeat(Psi, Gamma.rows(), 1, Psi_repeated);
    Core.subtract(Gamma, Psi_repeated, Phi);
}