An Acne Detector for Skin Image Based On Attention Enhanced Feature Pyramid Networks

Automatic acne detection can help people more quickly detect acne disease. At present, most of the researches are based on two-stage detectors, and the researches based on single-stage detectors are very few. However, smaller and faster models are required in practical applications. In this paper, we build a faster and more accurate acne detection model based on YOLOX. First, we propose an Attention Enhanced Feature Pyramid Network (AEFPN) to improve the efficiency and quality of feature fusion. It works by controlling the delivery of feature information and using attention. Then, we propose a stronger data augmentation strategy named MixAug, which uses Copy-Paste and Mixup by turning with a certain probability to make the model converging faster and the training curve smoother. In addition, to assign higher quality samples for small objects, we use Normalized Wasserstein Distance to improve the label assignment strategy called Simplified Optimal Transport Assignment (SimOTA). Finally, we conducted a number of experiments on the AcneSCU dataset. The results demonstrate the proposed method could improve the performance of YOLOX in acne detection, allowing YOLOX to increase 2.62% average precision (AP). At the same time, it can effectively improve the convergence speed of the model. Our model is superior to the current research in model size, detection speed and accuracy.