Foreground-Aware Knowledge Distillation for Enhanced Damage Detection

Abstract

Damage detection remains a critical challenge, especially within the industrial automation sector, necessitating the development of advanced inspection technologies and their potential applications. Conventional industrial inspection methods are hindered by high costs and operational disruptions, motivating the development of innovative and efficient solutions. This paper introduces a novel, architecture-agnostic deep neural network (DNN) knowledge distillation (KD) method able to enhance vision-based damage detection performance even in challenging industrial environments. Our proposed method integrates foreground knowledge with feature KD to enhance data feature utilization in detection models, effectively minimizing background clutter. The results demonstrate the efficiency of our method in consistently enhancing the student’s training process, including up to a 12% increase in mean Average Precision (mAP), across various DNN architectures. Our approach bridges the gap between academic research and real-world industrial current applications, offering a robust solution for damage detection in insulated pipelines.