Automatic Acne Detection Model Based on Improved YOLOv7

Acne is a common skin disease typically diagnosed visually by dermatologists, involving manual localization and labeling. Instead of conventional manual marking, automated acne diagnosis can save medical time and avoid cross-contamination. This paper presents a new automatic detection model for acne based on improved YOLOv7. Specifically, the ELAN module in backbone is improved to enhance feature extraction functionality. Meanwhile, We added the EPSA module in Backbone to extract the multi-scale spatial feature information more effectively. And the concat node of Neck layer is adjusted to strengthen the feature fusion capability, thus improving the accuracy of target detection. Moreover, the activation function is replaced with ELU to improve the robustness of the model. Finally, the initial anchor boxes of YOLOv7 are adjusted by the K-Means algorithm to make the model more suitable for acne detection. We conducted our experiments using the publicly available ACNE04 dataset to train the model. The experimental results demonstrate the robust detection capabilities of the model, achieving an mean average accuracy(mAP) of 83.7% in acne detection. The mAP is 4.57% higher than the initial model and better than the general object detection model. This model holds promise for effectively managing patient health status and providing valuable assistance to physicians in the diagnostic process.