Research And Application Of Deep Learning Based Small Object Defect Detection

In the textile field,quality inspection of fabric is an important part of the product production chain.Due to the limitation of production equipment,stains,holes and other defects often appear on the surface of the fabric.The traditional detection method is manual detection,but this method has problems such as high labor intensity,low detection accuracy and slow detection speed.With the development of artificial intelligence,it is an inevitable development trend to use computer vision instead of human eyes in fabric defect detection to free labor and improve production efficiency.However,the current object detection algorithm is not ideal for detecting small targets such as defects.Therefore,this paper proposes a small target defect detection technology based on deep learning,which can improve the accuracy of fabric defect detection and realize the real-time detection of fabric defects.The main research contents of this paper are as follows.(1)In this paper,a figured fabric with a complex background pattern is used as the research object,and the figured fabric with defects is photographed and collected,and the image annotation tool is used to annotate and establish the figured fabric defect dataset.Then the public fabric defect dataset is used to supplement the fabric defect types.In case of an insufficient amount of data,geometric transformation is used for data argumentation to expand the size of the figured fabric defect dataset.(2)The composition structure of convolutional neural network is studied,the network structure of two-stage and one-stage object detection algorithms is analyzed,and the performance of the two types of algorithms is compared,and YOLOv5 is selected as the basic model of this paper by analyzing the causes of low accuracy problems in small object detection tasks and the status of related technology research.(3)In order to improve the accuracy of small object defect detection,a defect detection algorithm based on improved YOLOv5 is proposed.To enhance the feature extraction capability of the network,the attention mechanism Squeeze-and-Excitation(SE)is introduced in the backbone of YOLOv5,and the traditional Leaky ReLU activation function is replaced by the ActivateOrNot(ACON)activation function.The experiments show that the optimized model has higher detection accuracy.(4)A fabric defect detection system is implemented,and the system is designed based on the defect detection algorithm model to realize the functions of image acquisition and defect detection,and the system is tested,which shows that the system has certain application value.