FAD-Enhanced Visual Tracking for Textile Defect Detection: A Multispectral Feature Fusion Approach

Huanxin Wei, Zhaodi Hu

Article
2026 / Volume 9 / Pages 587-615
Received 2 July 2025; Accepted 23 September 2025; Published 11 March 2026
https://doi.org/10.31881/TLR.2026.587

Abstract
The textile industry urgently requires automated defect detection systems that can operate reliably under diverse production conditions while prioritizing accuracy. Existing vision-based solutions struggle with spectral variability, dynamic defect evolution during high-speed manufacturing, and computational constraints. This paper presents a novel Feature Alignment and Distillation (FAD)-enhanced visual tracking framework, incorporating a FAD module to refine multispectral features and a dynamic Siamese tracker to ensure precise localization. Comprehensive ex-periments on the TextileDefect-3K dataset demonstrate state-of-the-art performance, achieving 94.8% precision, 93.1% recall, and 93.9% F1-score. The proposed method significantly outperforms the RGB baseline and the NIR-UV baseline by 13.4 and 6.7 absolute F1-score points, respectively, effectively resolving spectral limitations. It showed particularly strong results for micro-defects, achieving 87.5% F1 in normal conditions (far surpassing the attention baseline’s 69.1%) and maintaining robust detection (84.6% F1) even in adverse scenarios. Furthermore, the system demonstrated exceptional stability under industrial speeds, maintaining 91.9% F1 at 5 m/s, and exhib-ited high environmental robustness with an average performance degradation of only 2.0 percentage points. These results establish that our approach not only bridges the gap between industrial-grade real-time processing and laboratory-level accuracy but also sets a new benchmark for robust quality control in Industry 4.0 smart manufacturing, proving effective where existing single-modality solutions falter.

Keywords
textile defect detection, multispectral feature fusion, dynamic Siamese tracking, real-time quality control, industrial computer vision