Boundary-sensitive Adaptive Decoupled Knowledge Distillation For Acne Grading

Abstract

Acne grading is a critical step in the treatment and customization of personalized therapeutic plans. Although the knowledge distillation architecture exhibits outstanding performance on acne grading task, the impact of non-label classes is not considered separately, resulting in low distillation efficiency for non-label classes. Such insufficiency will cause the misclassification of the acne images located on the edge of the decision boundary. To address this issue, a novel method named Adaptive Decoupled Knowledge Distillation (ADKD) which considers the uniqueness of the acne images is proposed. In order to explore the influence of non-label classes and enhance the model’s distillation efficiency on them, ADKD splits the traditional KD loss into two parts: non-label class knowledge distillation (NCKD), and label class knowledge distillation (LCKD). Additionally, it dynamically adjusts the NCKD based on the distance between the sample and each non-label class. This allows the model to allocate different learning intensities to various non-label classes, reducing the overrecognition of classes near the sample and the underrecognition of distant classes. The proposed method enables the model to better learn the fuzzy features between acne images, and more accurately classify the samples located on the decision boundary. To verify the proposed method, extensive experiments were carried out on ACNE04 dataset, ACNEHX dataset, and DermaMnist dataset. The experimental results demonstrate the effectiveness of this method, and its performance surpasses that of current state-of-the-art (SOTA) method.