Abstract: Aiming at the problems of inefficiency of dense multi-face detection, faces are easy to be occluded and difficult to be detected with small size, a YOLOv8-IFD dense multi-face detection algorithm is proposed. Based on YOLOv8, firstly, the DCNv2 convolutional variant module is embedded in the backbone network to increase the modulation mechanism and use the weights to adjust dynamically, and enhance the model learning ability to improve the detection accuracy; then, the CA attention mechanism module is introduced in the tail and neck networks of the YOLOv8 backbone network to strengthen the ability of image feature extraction, and inhibit the influence of irrelevant features on the accuracy; finally, a loss WIoD algorithm is introduced based on the dynamic non monotonic focusing mechanism based loss WIoU Loss is introduced as the bounding box regression loss function to accelerate network convergence and improve the detection accuracy of the model. Experiments are conducted using the Wider Face public dataset, and the results show that the improved YOLOv8-IFD algorithm improves the mean average accuracy by 2.4% compared with the original algorithm, and the amount of computation is relatively reduced. In addition, compared with other algorithms, there is a significant improvement in the effect of dense multi-face detection, which verifies the effectiveness and applicability of this improved method.
辛鹏, 娄铖. 基于改进YOLOv8的密集多人脸检测[J]. 吉林化工学院学报, 2024, 41(11): 24-30.
XIN Peng, LOU Cheng. Intensive multi-face detection based on improved YOLOv8. Journal of Jilin Institute of Chemical Technology, 2024, 41(11): 24-30.
