An improved approach for face detection using superpixels, moment-based matching, and isosceles triangle matching

Face detection serves as a crucial step for a wide range of applications in computer vision. In this paper, we delve into the task of face detection. An algorithm is proposed for colour images to be robust to varied illumination, background setting, head pose and skin colour. Taking advantage of the superpixel segmentation followed by our trained SVM classifier, we are able to identify different skin-tone faces and generate face candidates. The moment-based elliptic shape matching is performed to remove invalid facial regions. Based on chroma and luma components, our scheme establishes the Eyemap and the Mouthmap to yield a pool of candidates for facial features. A delicate examination procedure considering the texture, colour and spatial relations with respect to the eyes-mouth pair is employed to verify each face candidate. Experimental results demonstrate better detection on the Caltech database in terms of F-measure. Results also show that our proposed algorithm more effectively rules out non-human faces than state-of-the-art algorithms.