Cloth Defect Detection System Based On Xavier

In the textile industry,fabric defects affect the quality of products to a large extent.Cloth defect detection plays an important role in industrial production.In the past,Chinese enterprises hired experienced workers to detect fabric defects,which was inefficient and labor-intensive.In recent years,an alternative of using computer vision technology is rising up,being a feasible way to the demanualization,high efficiency and low cost of fabric defect detection.Therefore,the research and development of defect detection system has high engineering application value for the increase of production capacity in the industry.In this thesis,we design a cloth defect detection system based on Xavier using modular design ideas.The system is divided into four modules,namely image acquisition module,encoder module,algorithm processing module and intercommunication module.The image acquisition module collects cloth images through a line-scan camera and performs image preprocessing.The encoder module is responsible for encoding the RGB image into JPEG format.The algorithm processing module,as the core module of the system,realizes the defect object recognition and classification of cloth images.The open source solution based on the two-stage target detection convolutional neural network,Cascade R-CNN,is used as the benchmark model.Aiming at the characteristics of cloth defect detection,the detection algorithm is improved in five aspects: frequency domain filtering that suppresses background,selection of lightweight backbone network Mobile Net,adding attention mechanism,adjusting loss function,and mosaic data enhancement.In terms of training and testing,we use open source deep learning development toolkit,MMDetection,to simplify the process.The intercommunication module,based on the TCP/IP,realizes the data and command communication via private protocol implementation between the system software module and the PC client.The test results show that the system has realized the functions of cloth defect image collection and coding,defect detection processing,and data communication with PC client.The functions of each module meet the requirements;the defect detection algorithm m AP in the performance test reaches 0.701,and the detection speed can reach108m/ min,showing good engineering application value.