Research Of Fabric Image Defect Detection And Location Algorithm Based On Convolutional Neural Network

The detection and localization of fabric defects is one of the important steps in the quality monitoring of textiles.However,the existing defect detection methods are not highly adaptive and the detection speed is slow due to the various types of fabric defects and the complex image texture background,so it is difficult to satisfy the industrial requirements.For complete feature information of the image can be extracted by deep learning technology,it has a good effect in the application of object detection.Combined with the characteristics of fabric images,in this thesis,we focus on a research of fabric image defect detection and localization algorithm based on convolutional neural network.The specific research results are as follows:(1)First fabric image acquisition system is built by constituent part the light source,lens,camera,image processing card and actuator,and then collecting a certain size of uneven hole,oil pollution,burr,drop needle,mark,slub fabric defect image based on this system,and through the transpose,gaussian filtering,image enhancement operations to enlarge datasets of the fabric image,then build the fabric image datasets,support for subsequent deep learning sample.(2)A fabric defect detection and location algorithm based on SSD(Single Shot MultiBox Detector)network model is proposed in this paper.SSD model achieves excellent detection effect through multi-feature layer fusion in target detection.Fabric images are characterized by texture details,which are generally at the bottom and middle of the depth network.The existing SSD detection model is directly used to detect fabric defects,which often missed small defect.In this paper,the fabric image features were combined to fuse the middle and bottom layers of the convolutional neural network for image representation.Then,the candidate areas were detected by extracting the default box.Finally,the detection accuracy was enhanced by Non-maximum Suppression(NMS).Experimental results show that the proposed method has good detection performance for small defects.(3)A fabric defect Detection algorithm based on ODFTNet(Object Detection based on Feature Transfer Network)is proposed.The above method has certain detection effect,but when the training datasets is small,the detection performance is not ideal.Based on the above problems,first improve the basic network VGGNet-16,use 3×3 convolution kernel in different feature mapping layers,use 2×2 core pooling operation on feature graph,the size of feature layer is gradually reduced,get deeper feature layer,and realize multi-scale feature layer fusion and prediction defect.In addition,the transfer learning technology is combined to obtain useful information for the target model from the trained source model.Finally,the Soft-NMS method is adopted to obtain the final defect box to improve the detection rate.This method improves the detection rate and has good detection performance in small database,at the same time,has strong robustness.