FAC-DETR: a lightweight fabric defect detection model integrating Fourier convolution and multiscale linear attention

To tackle localization inaccuracies and false positives caused by complex fabric textures, significant morphological variations within predefined defect categories, and large variations in defect scales, this paper proposes FAC-DETR, a lightweight defect detection model based on an enhanced RT-DETR framework. First, the backbone network is redesigned using a Cross-Stage Partial (CSP) structure. By integrating FourierConv, we introduce an efficient feature extraction module named FConv-Block, which improves the network’s capability to capture intricate fabric textures and defect details while reducing the number of parameters. Second, a Multi-Scale Linear Attention mechanism is incorporated into the RT-DETR encoder. This module adaptively highlights critical defect regions and suppresses redundant background information, thereby enhancing feature discriminability and reducing computational cost. Finally, we propose a Context-Guided Spatial Feature Reconstruction Pyramid Network, which leverages multi-scale contextual information to refine object feature representation, improving both robustness and localization accuracy in complex industrial environments. Experimental results on the TILDA fabric dataset show that FAC-DETR achieves a 4.2% increase in mAP@0.5 and a 7.6% gain in mAP@0.5:0.95 over the baseline RT-DETR, while using 1.37 M fewer parameters. Additionally, the model demonstrates strong cross-dataset generalization on the MVTec and an industrial dataset, attaining mAP@0.5 scores of 94.2% and 97.1%, respectively. While our supervised approach relies on annotated data for specific defect classes, these results confirm its high efficiency and accuracy in targeted fabric defect detection.