Machine Vision Driven Adaptive Detection Method For Stripe Fabric Defects

Defect detection of striped fabrics has always been an important task for fabric manufacturers.Because the color matching and minimum cycle changes of different background stripes interfere with visual detection,the task of visual adaptive detection of defects on different background texture stripe fabrics has become a difficult engineering problem to solve.Based on the characteristics of background texture variation among different stripe fabrics,this paper studies machine vision-driven stripe fabric defect adaptive detection method.By introducing multi-scale detector and attention mechanism,a coordinate attentionenhanced stripe fabric defect multi-scale detection model is designed,which improves the detection accuracy of the detection model on stripe fabrics.A reliable pseudo label guided stripe fabric defect detection depth domain adaptive method is designed to achieve the migration of model defect detection performance.The main research work is as follows:To solve the problem that the stripe fabric defect detection task is easy to miss small area target,a coordinate attention-enhanced multi-scale stripe fabric defect detection method is studied.A multi-scale stripe fabric defect detection module with small-scale detector is designed.By fusing shallow feature information,the detection model is implemented to detect small-size defect features.A stripe fabric defect feature extraction module based on coordinate attention is designed.By integrating the coordinate information of the vertical and horizontal directions of the feature matrix,the spatial coordinate dependency of the detection model on the slim defect is extracted.The experimental results show that the detection method proposed in this paper improves the accuracy and recall index by 0.47% and 9.17%,respectively,compared with the benchmark method.It can effectively improve the detection rate of small size defects and large aspect ratio defects,and improve the detection performance of the model on stripe fabrics.To solve the problem that the detection knowledge of the detection model is difficult to reuse on the target stripe fabric without labels,an adaptive depth domain method for stripe fabric defect detection is studied.A cross-domain texture image pre-generation module is designed to generate the corresponding pseudo-texture image between the original and target stripe fabric images.A domain adaptive module based on mean teachers is designed.The training process of the original network is guided by introducing the pseudo-labels generated by the teacher network to detect the pseudo-texture images,and the detection model is aligned with the feature distribution on the original and target stripe fabric image sets.A reliable pseudo-label measurement strategy and model learning process are proposed.By measuring the stability of test results on the same image of models under different training cycles to evaluate the reliability of generated pseudo-labels,the filtering of reliable pseudo-labels and the model learning process based on reliable pseudo-labels are implemented.The experimental results show that the reliable pseudo-label guided stripe fabric defect detection depth domain adaptive method improves the accuracy by 16% to 91.23% and the recall rate by 21.68% to 91.79%compared with the benchmark model,which can effectively improve the detection performance of the model on the target stripe fabric and realize the migration of the detection model defect detection performance.Based on the above theory and according to the actual needs of stripe fabric detection,a cloud edge collaborative fabric defect detection system is built,which provides an effective solution tool and platform for stripe fabric adaptive defect detection.The research results in this paper provide theoretical methods and corresponding technical tools for streak fabric defect adaptive detection,and have important engineering value for improving the quality of streak fabric factory.