Research On Fabric Defect Detection Method Based On Deep Learning

With the improvement of people's life,the demand for fabric is increasing,and the production process of fabric is developing towards the direction of intelligent manufacturing to satisfy the demand.Among it,a very important link is the defect detection of fabric.The speed and the accuracy of traditional detection methods can not meet the requirements of the intelligent production lines.The rapid development of computer vision technology provides a more intelligent and efficient technical approach for the fabric defect detection.Fabric defects have three characteristics: different scales,large proportion of small defects,large aspect ratio span,and unbalanced categories.The method of manually designing the feature expression using traditional image processing methods cannot adapt to its complex defect characteristics.Therefore,a fabric defect detection method based on deep learning technology is designed.By analyzing the object detection algorithm based on deep learning,combining the characteristics of fabric defects and the needs of test results,A two-stage object detection method is used as the basic detection framework.Aiming at the characteristics of fabric defects with different scales and large proportion of small defects,a multi-scale feature extraction method based on residual structure is proposed.The feature pyramid structure is used to extract the multi-scale features of the defect image,and the feature re-calibration is carried out within the structure of the backbone network to focus on the key channel features.At the same time,the deformable convolution is integrated to enhance the network's ability to capture the defect shape.To further improve the feature utilization efficiency,the ROI Pooling calculation method is redesigned and the global features of the image are superimposed in the process.At the same time,a detection method based on improved Faster R-CNN is designed for the characteristics of fabric defects.First a cluster analysis algorithm is used to statistically analyze the fabric defect data set,and key parameter anchor is designed for the characteristics of the large length-width ratio span of the fabric defect data set.At the same time,a multi-detector structure using a non-single threshold is used to modify the bounding box multiple several times.Aiming at the imbalance of the fabric defect data set,the data set is expanded by using classic data enhancement method,and the loss function that can focus more on difficult-to-classify samples is redesigned based on the Focal loss idea to further improve the detection accuracy.Finally,experimental verification and result analysis is conducted.The training of the deep learning model is completed.Then visually displayed and analyzed the results of multi-scale feature extraction,and tested the detection accuracy and detection time of the designed method,which verified the effectiveness of the designed method.