A New Descriptor for Acne Detection and Skin Care Framework Using AI Algorithms

Abstract

Accurately detecting and assessing the severity of acne is crucial for effective patient treatment. However, dermatologists often encounter challenges in grading acne precisely due to the similar appearance of lesions with varying severity. This study proposes a robust framework using the YOLOv9 algorithm, achieving a mean average precision (mAP) of 0.540. Advanced image preprocessing techniques, including noise reduction, hue normalization, and contrast enhancement, were employed to mitigate the effects of low resolution, poor lighting, and noise, enhancing detection accuracy by approximately 15% under challenging conditions. The model incorporates a Multi-level Fusion Layer, combining spatial and semantic features from various scales, which improved precision and recall for small and overlapping acne lesions by 12%. Trained on a diverse dataset of over 41,000 images, the framework accurately identifies different acne types—comedones, papules, nodules, and scars—across varying skin tones. This approach facilitates early detection, timely intervention, and enhanced lesion visualization, paving the way for practical applications in dermatological diagnostics

Keywords

Acne detection; Dermatology; Deep learning; Object detection; YOLOv9