DefinePK

Abstract

This paper presents the comparison between two face detection methods and their parts, which for this case are the two eyes and the mouth, which are the Viola-Jones (VJ) algorithm and the semantic segmentation based on convolutional neural networks (SegNet). To make the comparison, the training of the proposed SegNet is carried out using a database of previously labeled faces, to be later tested to verify its operation, where 97.55% of average accuracy and a mIoU of 76.64% are obtained. As for the VJ algorithm, an improved version for Matlab is used, which is able to detect the parts of the face even when it has an inclination of up to 20°. The tests are carried out with 10 images of the CelebA dataset, in such a way that each algorithm identifies the complete face, the right and left eye independently, and finally the mouth. In the event that any part of the face has been removed, the algorithm should not detect that section, since if it does it is counted as a false positive. In the tests, VJ obtained an overall accuracy of 79.38%, a low percentage compared to that obtained by SegNet, which was 97.97%. This allows to see the capacity of the proposed network to identify the parts of the face and estimate when there is no information of any of the parts to be detected.