Eyes Localization Algorithm Based on Prior MTCNN Face Detection

摘要

Aiming at the inaccuracy of the regression position of MTCNN human eyes when the image resolution is low, a method for eyes localization algorithm which is based on prior MTCNN face detection is proposed in this paper. Firstly, the eyebrow area is segmented according to the face area and pupil position obtained by MTCNN network. Then, the gray gradient integral projection in horizontal direction and the gray integral projection in vertical direction of eyebrow area are calculated. Finally, The MTCNN key points positioning results and integral projection results are combined to position the human pupil position accurately. The experimental results show that the proposed algorithm has a detection accuracy of 95.02%, and has good robustness for eye image detection with different quality.