<![CDATA[Facial skin diseases are common health conditions that can significantly affect both physical and psychological well-being. Early identification is essential to minimize the risk of disease progression. However, in many areas, there is still a lack of access to dermatological care. Although deep learning algorithms have been widely used in medical image categorization, few studies offer a direct comparison between convolutional neural networks (CNN) and transformer-based architectures within a cohesive experimental framework, especially concerning the classification of facial skin diseases. This study compares the effectiveness of ResNet-50 with Swin Transformer V2 and develops a deep learning system to classify six different types of skin problems on the face. The models were evaluated using accuracy, precision, recall, and F1-score after the dataset was divided into subsets for testing, validation, and training. According to the trial results, Swin Transformer V2 achieves an astounding accuracy of 97.54%, outperforming ResNet-50, which achieves 94.44%. The training curves indicate stable learning behavior with minimal overfitting. Grad-CAM visualization is applied to improve interpretability by highlighting relevant regions in the images. The best-performing model is implemented in a Flask-based web application as a prototype system for early detection. These results demonstrate how transformer-based architectures can improve classification performance and highlight their potential applications in practical diagnostic support systems]]>
