How to improve image segmentation using the watershed?

Question

I am developing an application to detect the lesion area, for this I am using the grabcut to detect the ROI and remove the background from the image. However in some images it is not working well. He ends up not identifying the borders of the region of interest well. The watershed can better identify the edges for this type of work, however I am having difficulties making this transition from grabcut to watershed. Before processing the grabcut, the user uses touchevent to mark a rectangle around the image of interest (wound area) to facilitate the work of the algorithm. As the image below.

However, using other wound images, segmentation is not good, showing flaws in ROI detection.

Image using grabcut in app

Image using watershed in desktop

this is the code:

private fun extractForegroundFromBackground(coordinates: Coordinates, currentPhotoPath: String): String {
    // TODO: Provide complex object that has both path and extension

    val width = bitmap?.getWidth()!!
    val height = bitmap?.getHeight()!!
    val rgba = Mat()
    val gray_mat = Mat()
    val threeChannel = Mat()
    Utils.bitmapToMat(bitmap, gray_mat)
    cvtColor(gray_mat, rgba, COLOR_RGBA2RGB)
    cvtColor(rgba, threeChannel, COLOR_RGB2GRAY)
    threshold(threeChannel, threeChannel, 100.0, 255.0, THRESH_OTSU)

    val rect = Rect(coordinates.first, coordinates.second)
    val fg = Mat(rect.size(), CvType.CV_8U)
    erode(threeChannel, fg, Mat(), Point(-1.0, -1.0), 10)
    val bg = Mat(rect.size(), CvType.CV_8U)
    dilate(threeChannel, bg, Mat(), Point(-1.0, -1.0), 5)
    threshold(bg, bg, 1.0, 128.0, THRESH_BINARY_INV)
    val markers = Mat(rgba.size(), CvType.CV_8U, Scalar(0.0))
    Core.add(fg, bg, markers)

    val marker_tempo = Mat()
    markers.convertTo(marker_tempo, CvType.CV_32S)

    watershed(rgba, marker_tempo)
    marker_tempo.convertTo(markers, CvType.CV_8U)

    val imgBmpExit = Bitmap.createBitmap(width, height, Bitmap.Config.RGB_565)
    Utils.matToBitmap(markers, imgBmpExit)

    image.setImageBitmap(imgBmpExit)


    // Run the grab cut algorithm with a rectangle (for subsequent iterations with touch-up strokes,
    // flag should be Imgproc.GC_INIT_WITH_MASK)
    //Imgproc.grabCut(srcImage, firstMask, rect, bg, fg, iterations, Imgproc.GC_INIT_WITH_RECT)

    // Create a matrix of 0s and 1s, indicating whether individual pixels are equal
    // or different between "firstMask" and "source" objects
    // Result is stored back to "firstMask"
    //Core.compare(mark, source, mark, Core.CMP_EQ)

    // Create a matrix to represent the foreground, filled with white color
    val foreground = Mat(srcImage.size(), CvType.CV_8UC3, Scalar(255.0, 255.0, 255.0))

    // Copy the foreground matrix to the first mask
    srcImage.copyTo(foreground, mark)

    // Create a red color
    val color = Scalar(255.0, 0.0, 0.0, 255.0)
    // Draw a rectangle using the coordinates of the bounding box that surrounds the foreground
    rectangle(srcImage, coordinates.first, coordinates.second, color)

    // Create a new matrix to represent the background, filled with black color
    val background = Mat(srcImage.size(), CvType.CV_8UC3, Scalar(0.0, 0.0, 0.0))

    val mask = Mat(foreground.size(), CvType.CV_8UC1, Scalar(255.0, 255.0, 255.0))
    // Convert the foreground's color space from BGR to gray scale
    cvtColor(foreground, mask, Imgproc.COLOR_BGR2GRAY)

    // Separate out regions of the mask by comparing the pixel intensity with respect to a threshold value
    threshold(mask, mask, 254.0, 255.0, Imgproc.THRESH_BINARY_INV)

    // Create a matrix to hold the final image
    val dst = Mat()
    // copy the background matrix onto the matrix that represents the final result
    background.copyTo(dst)

    val vals = Mat(1, 1, CvType.CV_8UC3, Scalar(0.0))
    // Replace all 0 values in the background matrix given the foreground mask
    background.setTo(vals, mask)

    // Add the sum of the background and foreground matrices by applying the mask
    Core.add(background, foreground, dst, mask)

    // Save the final image to storage
    Imgcodecs.imwrite(currentPhotoPath + "_tmp.png", dst)

    // Clean up used resources
    firstMask.release()
    source.release()
    //bg.release()
    //fg.release()
    vals.release()
    dst.release()

    return currentPhotoPath
}

Exit:

How do I update the code to use watershed instead of grabcut?

Answer 1

A description of how to apply the watershed algorithm in OpenCV is here, although it is in Python. The documentation also contains some potentially useful examples. Since you already have a binary image, all that's left is to apply the Euclidean Distance Transform (EDT) and the watershed function. So instead of Imgproc.grabCut(srcImage, firstMask, rect, bg, fg, iterations, Imgproc.GC_INIT_WITH_RECT), you would have:

Mat dist = new Mat();
Imgproc.distanceTransform(srcImage, dist, Imgproc.DIST_L2, Imgproc.DIST_MASK_3); // use L2 for Euclidean Distance 
Mat markers = Mat.zeros(dist.size(), CvType.CV_32S);
Imgproc.watershed(dist, markers); # apply watershed to resultant image from EDT
Mat mark = Mat.zeros(markers.size(), CvType.CV_8U);
markers.convertTo(mark, CvType.CV_8UC1);
Imgproc.threshold(mark, firstMask, 0, 255, Imgproc.THRESH_BINARY + Imgproc.THRESH_OTSU); # threshold results to get binary image

The thresholding step is described here. Also, optionally, before you apply Imgproc.watershed, you may want to apply some morphological operations to the result of EDT i.e; dilation, erosion:

Imgproc.dilate(dist, dist, Mat.ones(3, 3, CvType.CV_8U));

If you're not familiar with morphological operations when it comes to processing binary images, the OpenCV documentation contains some good, quick examples.

Hope this helps!