Tiny Face Detection Based on Deep Learning

Abstract

This paper proposed an improved strategy of tiny-face detection applied to images, videos, and real-time scenarios. First, six feature layers which are more conducive to the detection of tiny faces are integrated on the basis of the Single Shot MultiBox Detector (SSD) framework, and the prediction box scale generated by the algorithm is adapted to the shape of the face. Then, a hierarchical training strategy is put forward. Different scales are set to divide the ground truth box (GT box) into different levels according to the area of pixels, and are sent to the network training in descending order. Finally, in the prediction stage, the new soft non maximum suppression (new-softnms) algorithm is proposed to accurately suppress the generated redundant prediction box. And the continuous Gaussian function is combined with the hard-threshold screening method of nms. The proposed solution of this paper obtains 93.7% and 82.6% average accuracy respectively in the FDDB and Wider Face datasets. In the detection of video and real-time scenarios, the minimum face size of 200 pixels can be detected while obtaining 45 frames per second (FPS).