YOLO-M: An Improved YOLOv5 Method for Occlusion Face

With the rapid development of deep learning, face detection has made significant progress. However, in the face of occlusion problems caused by lighting, height, facial features, and obstructions, the accuracy of face detection algorithms has significantly decreased. In response to the problem of low accuracy in face detection under complex local occlusion, we establish a dataset of occluded faces and propose the YOLO-M occluded face detection algorithm based on the state-of-the-art detector YOLOv5. Firstly, we improved the backbone feature extraction network and designed a new Improved-Path Aggregation Network (I-PANet) for multi-scale feature fusion of images to obtain richer contextual information and suppress occluded facial information. Secondly, drawing on the idea of Swin Transformer, it is embedded into the C3 convolution block to form Swin Transformer Prediction Head (STPH). Finally, we conducted experiments on WIDER FACE which is the most challenging face dataset, the result show that the algorithm proposed in this paper can effectively improve the result of occluded face detection.