Research on small target face detection algorithm based on improved YOLOv8

Abstract

Small target face detection has a wide range of applications in fields such as security monitoring, intelligent access control, and human flow statistics. However, due to the low proportion of face pixels in the image and insufficient feature expression, the performance of the model detection is significantly affected. Although YOLOv8 is suitable for general object detection, its application in low-resolution scenes is limited by the loss of high-resolution features and insufficient extraction of key region information in small object face detection. To solve the above problems, this paper proposes an improved YOLOv8 algorithm based on multi-branch feature fusion, and optimizes the YOLOv8 model from the following aspects:1) Add high-resolution feature branches to retain small face detail information; 2) Embed the CBAM attention module in the core small target feature layer to enhance the features of key areas of the face; 3) Optimize the multi-branch feature fusion path of the neck network to achieve effective complementarity of features at different levels. 4) Expand the dimension of the detection head to cover small face detection scenarios. The experimental results show that the precision, recall and mAP of the improved model on the WiderFace dataset have increased by about 2.17, 2.04 and 1.78 percentage points respectively compared with the YOLOv8 model. On the subset of small target face images with a maximum face size of less than 40 pixels, the precision, recall and mAP are about increased by 4.07, 2.7 and 3.08 percentage points respectively, which effectively improves the detection precision of small target faces.