Research On Fabric Defect Detection Technology Based On EDSR And Improved Faster RCNN

Fabric defect detection is a very important part in the production of textile enterprises.The quantity of defects directly affects the quality of fabric.Therefore,how to further improve the accuracy of flaw detection is a hot issue in the field of image processing in the textile industry,which can help the textile industry reduce costs and increase efficiency and promote the development of relevant enterprises.Therefore,this paper carries out a detailed study from the common problems in the field of fabric defect detection.The main research work completed in this paper is as follows:（1）Research on improved Faster RCNN fabric defect detection technology.In view of the problems that the current leading algorithms in the field of fabric defect detection have insufficient recognition ability and weak generalization ability in defect detection,this paper,based on the mainstream deep neural network algorithm Faster RCNN,combined with the characteristics of common defects in the textile industry,carried out four research improvements.Firstly,k-means clustering algorithm is used to preset the proportion of anchor frames,and the proportion of 3 preset anchor frames in RPN module is increased to 9,so as to improve the size and shape of anchor frames generated in RPN and increase the ability to identify different defects.Then,aiming at the defect with the disproportion of aspect ratio,the trunk network of Faster RCNN,Res Net50,is improved and integrated into DCNv2 module in Res Net50,so that the network has the adaptive learning ability of the defect shape.Then,FPN network was introduced behind the backbone network to improve the multi-scale feature extraction ability of the network and enhance the accuracy of the detection network to identify small flaws.Finally,in view of the quantization operation error caused by the rounding of the position of the feature map by the Faster RCNN network itself,the ROIPooling in the ROI module is replaced with ROIAlign to correct the position of pixel points and reduce unnecessary errors.（2）Research on EDSR-based super-resolution reconstruction technology in the field of fabric defect detection.In order to solve the problem of insufficient defect feature information and high fusion degree of defect and background information in the current field of fabric defect detection,this paper proposes to apply super-resolution reconstruction technology to fabric defect detection.First,the samples are preprocessed to obtain low resolution images,and the DIV2 K data set is generated using the obtained low resolution images and the original high resolution images.Then,the EDSR network was used for super-resolution reconstruction of the dataset.The residual network was used to learn the high-frequency mapping information of high and low resolution sample pairs,and the sub-pixel convolution was used to obtain the twofold super-resolution reconstruction.In this way,the flaw feature information can be enriched,the fusion degree of flaw and background texture can be reduced,and the extraction ability of various flaw features can be enhanced.Finally,according to the characteristics of fabric defects,PSNR was selected as the evaluation standard of image reconstruction effect.When the PSNR threshold is reached,the super-resolution images generated by EDSR are sent to the improved Faster RCNN detection network for defect identification and detection.（3）Finally,three kinds of experiments are carried out to verify the detection effect of the proposed algorithm model.The experimental object is a large sample data set containing 20 kinds of common defects in the field of fabric defect detection.The first type of experiment was used to verify the improved detection effect brought by various improvements to Faster RCNN.The second type of experiment was used to verify the changes in detection accuracy of all defects before and after the introduction of EDSR.The third kind of experiment is used to verify the comparison between the algorithm proposed in this paper and other mainstream methods.Meanwhile,the Cascade RCNN network is emphatically selected to apply the above improvement to check the effect of the improved idea proposed in this paper on other mainstream convolutional networks.The results of three kinds of comparative experiments show that the overall recognition rate of 20 kinds of fabric defects based on EDSR and improved Faster RCNN reaches 85.0% m AP,which exceeds other current methods and can be used in the production and operation of textile enterprises.