JDCBL: A Joint Dynamic Contexts and Balanced Learning Framework for Real-World Fabric Defect Detection

Fabric defect detection is crucially essential for maintaining fabric quality. Existing intelligent detection systems, however, often fall short in real scenarios due to significant variations in defect sizes, which typically manifest as small and elongated structures. In addition, insufficient samples in real scenarios result in a long-tail distribution in training data, posing challenges to the balanced data prerequisites of having collecting balanced data for deep learning methods. This paper introduces the Joint Dynamic Contexts and Balanced Learning (JDCBL) framework to cope with these challenges. Within this framework, Contextual Transformer Block (CoTBlock) is designed to capture both global and fine-grained features of fabric defects simultaneously, which forms CoTNet. This modification boosts the interactivity of contextual information between input keys and effectively enhances self-attention learning. Additionally, class average and class complement strategies (BCL) amplify the exposure of tail-class defects so they can be accurately identified. Extensive validation on the Tianchi fabric dataset demonstrates that the JDCBL framework surpasses other leading methods in multiple defect detection tasks, achieving a MAP50 of 39.3%—over 5% higher than competing methods.