Baseline for lightweight defect detection of industrial fabrics: frequency domain enhancement and synthetic defect generation

To address challenges such as high annotation costs, strong texture periodicity, and model deployment difficulties in textile defect detection tasks, this paper proposes a lightweight and reproducible baseline method—WeaveADLite. This approach eliminates the need for manual annotation by leveraging an open-source texture dataset (DTD) and automatically generating multi-type defects (stains, scratches, holes, patches) to produce pixel-level supervision signals, enabling "zero-threshold" data acquisition. To enhance the model's robustness to fabric periodic textures, a frequency-domain prior is introduced at the input layer: high-pass residuals and bidirectional Gabor responses are added to the original RGB channels, forming a 6-channel input that effectively suppresses lighting interference and highlights texture discontinuities. The backbone network adopts a lightweight UNet structure based on depthwise separable convolutions, combined with a hybrid loss function of BCEWithLogits and Soft Dice for efficient training. On the DTD synthetic validation set, WeaveADLite achieves outstanding results in both pixel-level IoU and AUROC, while maintaining industrial real-time inference speeds. Experiments demonstrate that the proposed method can accurately locate and segment multiple types of fabric defects while preserving structural simplicity, exhibiting strong reproducibility and transfer potential. This study provides a minimal yet deployable baseline for textile visual inspection.