Lightweight real-time face detection method based on improved YOLOv4
Meng'An Shi, Yang Gao · 2021 International Conference on Computer Information Science and Artificial Intelligence (CISAI) · 2021
Aiming at the problem that YOLO, a general target detection algorithm, has a large number of network parameters, and it is difficult to get real-time feedback when it is directly embedded in face detection tasks, a lightweight target detection method improved by YOLOv4 is proposed. Changed backbone feature extraction network CSPDarknet53 to pre-select MobileNet and GhostNet as backbone instead. The best backbone feature extraction network was selected from WIDER FACE, a large open data set of human faces. At the same time, considering that the scale direction of the face is different when it is detected, the k-means++ algorithm is used to generate a new face prediction bounding box to accelerate the convergence speed and improve the detection accuracy. The experimental results show that compared with the original algorithm, the network parameters and computational effort are greatly reduced, and the original accuracy is retained to meet the needs of human face detection.