Improvement of YOLO26s for Detecting Foreign Objects on Non-Woven Polypropylene Fabric

Abstract

To address the small size, low contrast, and strong visual similarity to fibrous backgrounds of metal foreign objects and chipping on nonwoven-fabric surfaces, this study uses YOLO26s as the baseline and develops a detection model tailored to this scenario through improvements in feature fusion and the loss function. First, a residual coordinate attention module is introduced after high-resolution P3 feature fusion, allowing the model to focus more precisely on critical regions. Second, the CIoU loss and normalized Wasserstein distance are combined during training at a ratio of 0.90:0.10, with the NWD scale constant set to 12.8 pixels, to improve tiny-object matching. Ablation experiments on a nonwoven-fabric foreign-object dataset show that the improved model achieves a precision of 0.723, a recall of 0.442, and an mAP@0.5 of 0.465, representing gains of 3.0, 2.6, and 0.6 percentage points, respectively, over the YOLO26s baseline. Among four alternative attention-module configurations, the SimAM configuration obtains the highest overall mAP@0.5 and mAP@0.5:0.95, at 0.454 and 0.194, respectively, but both remain below those of the complete model proposed here. These results indicate that combining coordinate attention with NWD can enhance the detection capability for low-contrast tiny objects, providing an effective solution.