This was developed for a company looking to detect unhealthy skin using computer vision algorithms. This was one of the decided upon techniques, which can be used to detect pimples and wrinkles. No machine learning techniques are used for this technique because of a lack of available data at the time of creation.

Discussion

This is an algorithm designed for the detection of acne using techniques including thresholding, contours, and moments. The algorithm is tested using the AGES dataset and has an accuracy of 76.17%.

Like most computer vision processes, you start with a gaussian blur to remove smaller areas that are not of interest to the algorihthm. The image is then converted to a grayscale and a blob detection algorithm. Anything below a certain size is lacking detail for accurate detection, so they are removed.

//Cycles through different thresholds of the gray scaled image and locates blobs
          for (int i = 1; i < 25; i++) { //Cycles from 0 to 250
            //Saves gray scale
            for (int index = 0; index < frame.cols * frame.rows; index++) {
              if (gray.ptr<unsigned char>()[index] > i * 10) thresh.ptr<unsigned char>()[index] = 250;
            }
          }
          

A number of thresholds are applied to the grey-scalled image in intervals of 20, each is saved to a seperate image. Contours are found on each image and anything with a circular shape is saved. This is followed by measuring the colour of the area and detecting whether it is close to the colour red.

//Locates contours, this includes blobs that could be freckles/spots
          findContours(thresh, contours, hierarchy, RETR_TREE, CHAIN_APPROX_SIMPLE, Point(0, 0));
          
          vector<Moments> mu(contours.size());    //Area of shape
          vector<Point2f> mc(contours.size());    //Center points
          
          for (int j = 0; j < contours.size(); j++) {
          
            //Get center point of the contour and using "moments"
            mu[j] = moments(contours[j], true);
            mc[j] = Point2f(mu[j].m10 / mu[j].m00, mu[j].m01 / mu[j].m00);
          

The images below show the original image, contouring and the removal of areas until the final two pimples are located.