Illuminance-Invariant Defect Detection for Multiform Fabrics Using Multiscale Feature Fusion Model

Abstract

Illuminance variations and the diversity of fabric defect types present significant challenges for defect detection in multiform fabrics. To address these issues, this paper proposes a novel multiscale feature fusion model for illuminance-invariant defect detection, incorporating feature extraction, feature selection, and feature fusion. First, to mitigate the substantial illuminance differences caused by variations in the material, texture, color, and knitting density under supplementary lighting, an adaptive local feature extractor is introduced. This extractor alleviates the limitations of traditional Detail Processing Modules under varying illuminance conditions. Second, to handle the challenges arising from fabric diversity in terms of material composition and knitting techniques, and the consequent variability in defect types (like size, shape), an elastic-parameterized feature selection module (EFSM) is proposed. Leveraging B-spline parameterization, the EFSM significantly reduces the parameter burden for feature selection. Finally, a multiscale information and attention-integrated defect classification module, called enhanced defect classification module, is developed to accurately fuse and classify diverse defect features. These enhancements eliminate the model’s dependency on illuminance conditions and significantly improve defect detection performance for multiform fabrics. The effectiveness of the proposed approach is validated in public and homemade datasets.