Enhanced RT-DETR for real-time transparent tape detection on fabric surfaces with reflection-resistant attention module

Abstract

In textile production, transparent tape is used for fabric fixing and quality marking, and its identification directly affects the accuracy of automated sorting. Because the tape’s color is similar to the fabric, with low contrast and strong reflectivity, traditional visual methods have difficulty achieving reliable recognition. To solve this, this paper introduces a high-precision transparent tape detection method based on the RT-DETR (Real-Time Detection Transformer) network. The efficient cross-stage partial Darknet53 (ECSPDarknet) is adopted as the backbone network, compressing model parameters significantly while enhancing feature extraction capabilities. The reflection-resistant attention module (RRAM) is integrated during the feature fusion stage to strengthen multiscale feature fusion, effectively solving the recognition challenges caused by the similarity between transparent tape and the background, as well as high reflection. The dynamic group shuffle transformer (DGST) replaces the reparameterization convolutional C3 (RepC3), resolving the high computational load and real-time bottlenecks introduced by the latter’s multibranch structure. In addition, the bounding box regression loss function is replaced with a weighted sum of SmoothL1 and FocalEIoU loss functions, optimizing the model’s convergence efficiency and improving detection accuracy. Ablation experiments were conducted with three sets of random seeds. The results show that the improved model reduces parameters by 43.8%, floating-point operations by 39.6%, and increases FPS by 13.2% compared with the baseline. Precision, recall, F1-score, mAP@0.5, and mAP@[0.5:0.95] improve by 2.5%, 4.2%, 3.3%, 2.2%, and 1.0%, respectively. Meanwhile, the algorithm outperforms mainstream methods in terms of detection accuracy, providing a foundation for high-precision, transparent tape identification on fabrics.