<p>With escalating global demands for the quality and safety of textile products, fabric flaw inspection is now considered an indispensable procedure in quality assurance, which guarantees product integrity and high standards. Compared with traditional manual detection, which has strong subjectivity and low efficiency, and traditional machine vision has poor adaptability to specific scenes, it has obvious advantages in complex scenes and large-scale data processing. We developed an efficient network, DSFR-DETR, to detect flaws in fabric. It achieves high accuracy and rapid convergence compared to existing methods. Firstly, we design an Efficient Reparameterization (EffiRep) module to simplify the backbone network structure. This module adapts to different feature sizes by dynamically adjusting the convolution kernel parameters. Secondly, building upon the EffiRep module, we further propose an Efficient Representation Cross Stage Aggregation (EffiRepCSA), as the core module of the backbone network. By employing a flexible channel splitting and aggregation strategy, it utilizes multiple convolutional branches and residual connections to boost the performance and quality of feature acquisition, concurrently maintaining the network’s lightweight architecture. Thirdly, we propose the Adaptive High-Low Frequency Feature Fusion (AHLF) module, which adopts a frequency domain structure combined with a hierarchical feature fusion strategy to enhance the model’s generalization ability and efficiency. The experiments demonstrate that the DSFR-DETR proposed in this paper improves 11.5% on mAP over two-stage models such as Faster R-CNN and 1.9% over single-stage detection models such as YOLOv8s. Furthermore, the convergence speed is significantly faster than other comparative models. Additional experiments confirm the favorable generalization capability and robustness of the proposed DSFR-DETR algorithm, demonstrating its potential for industrial applications.</p>