Research On Fabric Defect Detection Technology Based On Deep Learning

Defect detection refers to marking the location,size and category of defects in fab-ric,which is an essential link in the production process of fabric.Traditional fabric defect detection is conducted by manual detection,but there are some problems in manual de-tection,such as missing or misdetection caused by eye fatigue,difficulty for human eyes to identify small defects,subjective judgment errors,and long detection time.The real-ization of automatic defect detection instead of manual detection is beneficial to improve the quality of cloth production,reduce labor cost and shorten production time.The main work of this paper is as follows:1.Proposed a fabric defect clustering strategy based on Kmeans++.For data sets that only mark the location and size of defects without specific categories,it is not effective to directly build models for flaw detection,so it is necessary to use clustering preprocess-ing.In this paper,the scale and area of defects were selected as clustering features,and Kmeans algorithm,Mini Batch KMeans algorithm and Agative algorithm were used for experiments.According to the experimental results,Kmeans++ algorithm was selected to perform Fabric defect clustering on Tian-Fabric data sets.It is convenient to construct the fabric defect detection model.2.Proposed an improved YOLOv5 s model YOLOF for fabric defect detection.To solve the problem of unbalanced category samples in fabric defect data set,the weight sampling method was introduced to optimize the model,and the samples with fewer oc-currence times were over-sampled during model training to achieve the effect of category sample balance.A feature fusion adaptive mechanism optimization model was introduced to solve the extremely small target defect problem,and the weight of each scale feature map contribution was considered in feature fusion.An inner volume convolution opti-mization model is introduced to solve the problem of extreme aspect ratio target defects.The spatial specificity of inner volume convolution improves the long distance pixel de-pendence problem.Through experimental comparison and analysis,YOLOF model with combinational optimization has a good effect.In the Fabric defect detection data sets Tian-Fabric and Fu-Fabric,YOLOF model proposed in this paper improves m AP by 0.061 and0.026 respectively compared with YOLOv5 s benchmark model.This model will serve as a model for subsequent prototype system deployment.3.Proposed a fabric defect detection prototype system EFDS based on edge comput-ing.EFDS mainly include center server and several edge server,center server is respon-sible for maintaining the edge server status,control the edge server task,and provide the client with a number of interfaces,the edge server deployment,on each inspecting ma-chine for industrial real-time detection,image acquisition of cloth camera to write test results into the center server database,Compared with the single-center server architec-ture,EFDS has the advantages of low communication cost and fast model inference.In this paper,the fabric defect clustering strategy,the fabric defect detection model YOLOF and the fabric defect detection prototype system EFDS have achieved good results on the two data sets,which is beneficial to realize the intellectualization and automation of fabric defect detection.