Fabric defect detection is critical for ensuring textile quality in industrial manufacturing. However, existing methods face significant challenges when dealing with complex texture backgrounds, high similarity between defects and normal patterns, extreme aspect ratios, and small defect sizes. To address these issues, we propose ASCR-DETR, an enhanced Detection Transformer specifically designed for fabric defect detection. First, we propose an Adaptive Spatial Geometric Convolution (ASGC) module, which dynamically adjusts sampling locations and modulation parameters through a Tri-Dimensional Cooperative Attention (TDCA) mechanism, enabling effective modeling and accurate capture of irregular defect geometries. Second, we design a Context-Guided Spatial Reconstruction Feature Pyramid Fusion Network (CSRFPN) that implements a progressive three-stage "calibration-reconstruction-fusion" framework. This network establishes cross-scale semantic correlations by aggregating multi-resolution contextual information, performs fine-grained spatial reconstruction to enhance local discriminability at each pyramid level, and achieves progressive top-down feature aggregation through dynamic interpolation and gated fusion mechanisms, effectively preserving small defect details while suppressing background interference across multiple scales. Third, we incorporate Coordinate Attention (CA) mechanisms to enhance global contextual representation, and introduce Dynamic Range Histogram Self-Attention (DRHS-AIFI) for intra-scale feature interaction, effectively suppressing background interference through value-domain statistical modeling. Extensive experiments on public and industrial datasets demonstrate that ASCR-DETR achieves 85.3% mAP@50(+7.1%) and 45.9% mAP@50:95(4.9%), outperforming state-of-the-art CNN-based and Transformer-based detectors while maintaining real-time performance, validating its effectiveness and deployment potential for industrial fabric inspection systems. The source code is available at:https://github.com/justDo000/ASCR-DETR.
Liu et al. (2026) studied this question.