Abstract: Automated facial acne detection is crucial for precise clinical diagnosis and treatment. However, existing methods still suffer from significant issues of missed detection and false detection of small acne targets. For more accurate acne detection, an enhanced YOLOv8s method with two key modifications is proposed in this paper. Firstly, the original backbone network of YOLOv8s is improved into a hybrid backbone network that integrates with Transformer. This improvement effectively combines the advantages of convolutional neural network capturing local detail information and Transformers maintaining global feature information, significantly enhancing the feature extraction and representation capabilities for small acne objects. Secondly, a multi-scale channel attention module is integrated into the neck network, enabling adaptive feature weight adjustment through cross-scale context aggregation, thereby mitigating semantic-scale inconsistency. Experiments on both public and self-built facial acne datasets demonstrate that, compared with the current state-of-the-art DSDH method, the proposed method achieves mAP improvements of 1.20% and 5.24% on respective datasets, with corresponding detection speed increases of 46.3 and 47.6 frame/s.

Key words: facial acne, YOLOv8s, small object detection, Transformer, multi-scale feature fusion