The link below use Matlab to remove non-text content from an image. I want to do the same thing with OpenCV in Java.
I don't have a Matlab to try with and I am new to OpenCV. Though I know some basics about the theory behind the process, but it's kind of difficult to make the translation from Matlab language into OpenCV 3.0. And preferably in Java.
ADD 1 - region detection with MSER (not resolved yet)
For MSER detection, I can use the following code to detect the MSER keypoints.
public static void MSERdetector(String imgName1, String suffix1) {
Mat imgMat1 = Imgcodecs.imread(picDir + imgName1 + "." + suffix1, Imgcodecs.CV_LOAD_IMAGE_GRAYSCALE);
String outImgName1 = picDir + "MSER" + "_keypoints_" + imgName1 + "_" + ".tif";
Mat outImg1 = new Mat();
FeatureDetector featureDetector = FeatureDetector.create(FeatureDetector.MSER); // create the feature detector
MatOfKeyPoint keypoints1 = new MatOfKeyPoint();
featureDetector.detect(imgMat1, keypoints1);
if (!keypoints1.empty()) {
Features2d.drawKeypoints(imgMat1, keypoints1, outImg1);
Imgcodecs.imwrite(outImgName1, outImg1);
System.out.println("done");
}
else {
System.out.println("No keypoints found for: " + imgName1);
}
}
And the output looks like this:
But I don't know how to convert keypoints into regions. What I need is below:
ADD 2 - Canny edges and intersection with MSER regions (not resolved yet)
Once I am able to find the MSER regions, I am supposed to intersect it with Canny edges. I can find some Canny edges as below. But I don't know how to do the intersection operation.
public static void CANNYedge(String imgName1, String suffix1) {
Mat imgMat1 = Imgcodecs.imread(picDir + imgName1 + "." + suffix1, Imgcodecs.CV_LOAD_IMAGE_GRAYSCALE);
//imgMat1 = ImageUtilities.Convert2BW(imgMat1);
String outImgName1 = picDir + "_CANNY_" + imgName1 + ".tif";
Mat outImg1 = new Mat();
Imgproc.Canny(imgMat1, outImg1, 0, 500);
Imgcodecs.imwrite(outImgName1, outImg1);
}
My canny edges output looks like this:
ADD 3 - Now I turned to use VS 2013 Community
For setting up OpenCV with VS2013, check here.
ADD 4 - Coding in VC++ 2013
Below is what I tried for now with reference to here.
//Step2: Detect MSER regions
Mat grayImage;
cvtColor(colorImage, grayImage, CV_BGR2GRAY);
imshow("Gray Image", grayImage);
waitKey(0);
Ptr<MSER> mserExtractor = MSER::create(); // create MSER extractor with default parameters. http://code.opencv.org/projects/opencv/wiki/MSER http://docs.opencv.org/master/d3/d28/classcv_1_1MSER.html#a49d72a1346413106516a7fc6d95c09bb
mserExtractor->setMinArea(150);
mserExtractor->setMaxArea(2000);
//Mat mserOutMask = Mat::zeros(grayImage.rows, grayImage.cols, CV_8UC3);
Mat vis;
//vis = Mat::zeros(grayImage.rows, grayImage.cols, CV_8UC3);
grayImage.copyTo(vis);
vector<vector<Point>> mserContours;
vector<Rect> mserBBox;//what's this?
mserExtractor->detectRegions(grayImage, mserContours, mserBBox);
for (int i = 0; i<mserContours.size(); i++)
{
drawContours(vis, mserContours, i, Scalar(255, 255, 255), 4);
}
imshow("MSER by contours", vis);
waitKey(0);
Mat vis2;
grayImage.copyTo(vis2);
for (vector<cv::Point> v : mserContours){
for (cv::Point p : v){
vis2.at<uchar>(p.y, p.x) = 255;
}
}
imshow("MSER by points", vis);
waitKey(0);
What I got are these:
vis1 - MSER by contours
vis2 - MSER by points
ADD 5
I just experimented with the text detection sample as suggested by Miki.
It requires some trained model files to run. And it took almost 2 minutes to finish but we can leave the performance later. My scenario is to OCR texts from complex screenshots (sorry to reveal that until now). Though the result is quite good for natural scenes. It is not so appealing for screenshots. Below is the result:
