A parameterizable interactive application for human hand segmentation in RGB images

Abstract

Segmentation is a fundamental technique in image analysis that allows objects, including human hands, to be detected and classified. This work presents the results of developing and testing a parameterizable interactive application, Hands Segmentation, for the segmentation of human hands in RGB images. The results show that the application efficiently segmented 100% of the regions of the RGB images where human hands existed. However, 2% of the segmentations included other human body parts, such as the arms. Excluding that 2%, the average of the Jaccard Index effectiveness metric was 0.73, which is considered an acceptable value for this type of application. A sample of twenty-two (22) human hands distributed in ten (10) images obtained by systematic sampling from a free-use image bank was used. Image processing was carried out with the Hands Segmentation application developed in C++ using artificial vision techniques and the specialized OpenCV library. The processing method included: a) transformation of the color space from RGB to YCrCb, b) image segmentation by simulation of hydrographic analysis (Watershed), c) delimitation of the regions based on contours and d) calculation of the effectiveness using the Jaccard Index (IoU) and precision metric, using an ad hoc qualification metric. The systematic use of the application shows that an adequate choice of the parameters a) brightness, b) contrast, c) dimension of the pixel-by-pixel operator, d) dimension of the morphological operator, and e) dimension thresholds of the regions, are essential to obtain an adequate segmentation of a human hand in an RGB image.