Differentiable NMS via Sinkhorn Matching for End-to-End Fabric Defect Detection
The paper "Differentiable NMS via Sinkhorn Matching for End-to-End Fabric Defect Detection" addresses significant challenges in fabric defect detection using deep learning methods. Conventional non-maximum suppression (NMS) disrupts gradient flow, preventing end-to-end learning. Moreover, obtaining pixel-level annotations in industrial settings is costly, which further complicates fabric defect detection due to the complex nature of textile background and defect morphology.
In this paper, the authors propose a differentiable NMS framework reformulated as a bipartite matching problem, solved via the Sinkhorn-Knopp algorithm. This approach preserves gradient flow throughout the detection network, paving the way for uninterrupted end-to-end optimization. The reformulation allows the integration of proposal quality, feature similarity, and spatial relationships, targeting irregular morphologies and ambiguous boundaries present in fabric defects.
A key component of the proposed framework is the entropy-constrained mask refinement mechanism. By modeling uncertainty in a principled manner, the approach enhances localization precision. The paper reports extensive experiments on the Tianchi fabric defect dataset, demonstrating notable improvements over existing methods and achieving real-time speeds suitable for industrial deployment. Additionally, the framework showcases adaptability across different architectures and generalization to broader object detection tasks, evidenced by competitive results on the COCO benchmark.
The authors highlight three main contributions in their work:
Differentiable NMS via Sinkhorn-Knopp Algorithm: The paper reformulates the non-maximum suppression task as a differentiable bipartite matching problem, which is solved using the Sinkhorn-Knopp algorithm. This allows for end-to-end training with continuous gradient flow through the detection pipeline.
Entropy-Constrained Mask Refinement Mechanism: The introduction of an entropy-constrained mask refinement process progressively improves proposal masks by uncertainty modeling, enhancing detection precision through theoretical convergence guarantees.
Competitive Framework for Fabric Defect Detection: The unified approach achieves competitive performance comparable with fully supervised methods, while offering superior efficiency-accuracy trade-offs beneficial for industrial applications.
The implications of this research are profound, indicating potential pathways for future developments in AI, particularly in automating defect detection processes in various industrial applications. The reformulation of traditional post-processing tasks offers avenues for integrated and holistic AI solutions where seamless end-to-end optimization may significantly improve performance without compromising computational efficiency.
The ability to effectively model complex defect morphologies while maintaining adaptability across detection architectures underscores the importance of differentiable components in designing modern AI systems. Future research may explore advanced feature aggregation techniques to improve small object detection and self-supervised pretraining strategies to reduce annotation demands, broadening the applicability of such frameworks to diverse industrial domains beyond fabric inspection.