Facial Grid Environment for Vitiligo Detection: A Deep Reinforcement Learning Perspective

This research addresses the challenging task of vitiligo detection within facial images through the lens of Deep Reinforcement Learning (DRL). Leveraging the Skin Disease dataset from Kaggle, our study focuses on the integration of three distinct convolutional neural network architectures—ResNet, VGG, and EfficientNet—within the DRL framework. Facial landmarks, crucial for spatial awareness, were extracted using OpenCV as a preprocessing step. The DRL agent navigates through a grid overlaying facial images, with states corresponding to grid cells, to reach the vitiligo-affected region and maximize cumulative rewards. Analysis of the models revealed ResNet's superior stability in the loss curve, indicative of its effective feature learning. The classification accuracy further reinforced ResNet's prominence, reaching an impressive 97%, outperforming VGG (94.5%) and EfficientNet (95.3%). While VGG and EfficientNet demonstrated competitive performances, ResNet's robustness and accuracy position it as the preferred architecture for vitiligo detection in facial images. This study contributes to the burgeoning field of dermatological diagnostics by showcasing the efficacy of DRL and different CNN architectures in automating vitiligo detection. The findings not only advance our understanding of the interplay between deep learning and dermatology but also lay the foundation for more accurate and efficient models in the broader landscape of skin disorder identification.