Abstract:
Facial Acne is a prevalent skin condition, can have significant economic and psychological effects on those affected, making accurate grading essential for effective treatment. In this study, we introduced a new grading system for acne that accounts for varios types of acne and a metric for precise severity assessment. Given the challenge of distinguishing different acne with similar appearances and counting them accurately, we cropped facial skin images into patches and compared a network, YOLOv8, YOLOv7, YOLOv6 and CNN, for analysis. YOLOv8 improves image quality through a median filter and histogram equalization, enhances feature representation with a channel attention mechanism, and addresses class imbalance using region-based focal loss. Additionally, model pruning and feature-based knowledge distillation were employed to reduce the model’s size. After processing through our model, the bounding boxes is being made around the skin acne, and the grading is refined using patient metadata. The entire process was implemented on a mobile device, resulting in an app that allows users to assess acne severity with dermatologist-level. We compared these four models and selected the best model based on confusion matrix, F1 curve, precision-recall curve, recall curve and Precision curve. Upon comparing YOLOv8 , YOLOv7, YOLOv6 and CNN we found that YOLOv8 has outperfoms other models and achieved better results.
