ارائه روشي مبتني بر يادگيري ماشين به صورت سبك‌وزن براي تشخيص بي‌درنگ عيوب پارچه

In modern textile industries, quality assurance an‎d the timely detection of weaving defects are of critical importance for reducing waste an‎d increasing productivity. However, the principal challenge lies in the gap between the high accuracy of computationally intensive deep learning models an‎d the limited computational resources available in industrial hardware. This constraint underscores the necessity of developing methods that establish an optimal balance among detection accuracy, processing speed, an‎d computational efficiency. In this context, the present study introduces a semi-supervised framework aimed at proposing a lightweight an‎d real-time method for fabric defect detection, which is computationally designed to be suitable for deployment on industrial embedded hardware. In the proposed approach, the transformer-based RF-DETR model is employed as a stronger model for generating pseudo-labels from unlabeled data, while the YOLOv12s model is utilized as the final lightweight model for real-time inference. To eva‎luate the effectiveness of the semi-supervised approach, two models with the YOLOv12s architecture were trained under identical conditions. The results indicated that the baseline model, trained solely on 850 native images of fabric defects, achieved an mAP50 of 91.78%. In contrast, the final model trained on a combined dataset (comprising 850 native images along with 1,784 pseudo-labeled an‎d refined images) attained an mAP50 of 98.8% on the validation dataset an‎d 90.1% on the test dataset, demonstrating a significant performance improvement resulting from the adoption of semi-supervised learning. To assess practical deployability in resource-constrained environments, the final model was optimized within the TensorRT framework, an‎d eight-bit integer weight quantization was applied. The quantized version, while maintaining an acceptable level of accuracy, achieved an mAP50 of 88.7% an‎d simultaneously increased the processing speed to 57.77 frames per second. The results indicate that the accuracy degradation caused by quantization is acceptable for real-time industrial applications when weighed against the substantial gains in speed an‎d computational efficiency. These findings demonstrate that combining the lightweight YOLOv12s architecture with a pseudo-label generation strategy an‎d inference optimization provides a balanced framework in terms of accuracy, speed, an‎d resource consumption. From a computational perspective, this framework aligns with the requirements of industrial embedded hardware an‎d offers an efficient, accurate, an‎d scalable solution for automated fabric defect inspection in textile production lines.

6

157186