HYBRID ATTENTION-GUIDED CAPSULE FRAMEWORK FOR AUTOMATED FACIAL ACNE SEVERITY CLASSIFICATION
Abstract
Automatic and precise evaluation of the severity of facial acne is an essential requirement of evidence-based clinical decision-making and to alleviate patient distress. Automatic and precise evaluation of the severity of facial acne is an essential requirement of evidence-based clinical decision-making and to alleviate patient distress. However, the modern deep learning methodologies are often challenged by constrained, massively imbalanced sets of data, thus lacking the ability to condense strong, perspective-free representations, which would provide credible severity stratification across heterogeneous imaging of patients. This paper presents the Hybrid Attention-Guided and Capsule-Based Pose Encoding (HAG-CAP) framework to the intelligent classification of acnes severity. VGG16 is combined with a sophisticated CBAM in this novel architecture to give priority to the salient lesions areas. Primarily, a Capsule Network is included to attain viewpoint-invariant feature encoding thus improving lesion discrimination. To enable generalization, the issues of data scarcity and class imbalance were alleviated using DCGAN-generated synthetic samples, thus increasing the number of annotated Roboflow samples to 2,500. The HAGCAP approach demonstrates better performance, which was confirmed by a thorough comparative analysis with the YOLOv6, YOLOv7 and YOLOv8 models. The overall accuracy was 99% with a precision and recall of 100%, and a strong kappa rate of 97.87%. The high reliability of the system justifies the fact that it can be implemented in light mobile health applications where scalable and automated dermatological decision support may be competently facilitated.
Keywords : HAG-CAP, acne severity classification, attention-guided feature selection, Capsule Networks, CBAM, DCGAN augmentation, deep learning, yolov6, yolov7, yolov8.
