DWEE-YOLO: a frequency–edge–geometry enhanced YOLOv10 network for low-contrast fabric defect detection

Reliable fabric defect detection is still difficult in industrial inspection, as many defects are hidden in repetitive textures and appear with weak contrast, small size, slender morphology, or unclear boundaries. This paper presents DWEE-YOLO, a YOLOv10n-based detector enhanced from frequency, edge, and geometric perspectives. In the shallow backbone, a dual-channel wavelet decoupling module separates low-frequency texture backgrounds from high-frequency defect responses. In the neck, Sobel gradient priors are combined with lightweight convolutional features to strengthen boundary-sensitive semantic fusion. Focaler-MPDIoU is further used to improve box regression for small and irregular defects. On the Tianchi dataset, DWEE-YOLO reaches 82.0% mAP@0.5, 49.5% mAP@0.5:0.95, and 36.8 FPS, exceeding YOLOv10n by 5.3 and 3.2 percentage points in the two mAP metrics. Tests on TILDA also confirm its cross-dataset adaptability.