Research Of Fabric Defect Detection Algorithm Based On Weakly Supervision Learning

Fabric defect detection plays an important role in textile production.Due to the complex texture of the fabric image and the various defect types,the existing detection methods based on machine vision have poor adaptability and low accuracy.Deep learning technology has made significant progress in object detection,and provides a novel solution for detecting fabric defects.However,the detection performance depends on the large-scale training set of fabric images with pixel-level labels.In actual production,it is a very heavy work to construct a large-scale training set with pixel-level labels for each type of fabric.Weakly supervised learning technology only relies on image-level labels,which can achieve pixel-level defect location.Therefore,this paper studies the fabric defect detection algorithm based on weakly supervised learning to solve the dependence of the detection model on the pixel-level labeled training set.The specific research results are as follows:1)A weakly supervised fabric defect detection algorithm based on Simplified-LZFNet is proposed.Strongly supervised networks need a training set with pixel-level labels.In this paper,combined with Class Activation Map,a fabric defect detection network Simplified-LZFNet based on weakly supervised learning is proposed.Firstly,the maximum pooling layer and the fully connected layer in the strong supervision network are replaced with a global pooling layer to facilitate the maintenance of defect spatial information.Then,the weight of the corresponding feature map is calculated by the global pooling layer,and generates a class activation map that can effectively highlight the defect area after weighted fusion.Experimental results show that the proposed method can achieve pixel-level defect location by the training set with image-level labels..2)A weakly supervised fabric defect detection algorithm based on DLSE-Net is proposed.For fabric images with complex textures,the contrast between the normal background and the various defects is not obvious,which results in the unsatisfactory detection of weakly supervised fabric defects.Therefore,this paper proposes an improved weakly supervised learning network,called DLSE-Net,which combines L-SE module and Diation Up-Weight CAM for fabric defect detection.Firstly,in order to achieve feature fusion between different levels,multi-branch network technology is used to reduce the semantic gap caused by the connection between different levels and enhance the texture feature representation.Then,the L-SE module is proposed to optimize the weights of different channels globally,which is more conducive to the selection of important feature channels.Finally,the improved attention mechanism DUW-CAM can suppress complex texture backgrounds and highlight defect areas by combining dilated convolution and attention mechanisms.Experimental results show that the proposed method has significant positioning accuracy on two fabric data sets with different textures,and is better than the latest method.3)A weakly supervised fabric defect detection algorithm based on self-attention mechanism is proposed.In order to further reduce the semantic gap between strong supervision and weakly supervision,this paper proposes an improved weakly supervised fabric defect detection network,called T-SCDM-Net,which is based on the pixel correlation mechanism in self-supervision.The main network framework adopts a three-branch structure and L-SE module,which globally optimizes and integrates different channel weights to improve feature representation capabilities.Then,the receptive field is expanded by using the dilated convolution,and the obtained contextual semantic information is used to measure the prediction of the current pixel through the semantic similarity for improving the positioning result.The detection results on two different fabric data sets show that the proposed method further improves the detection performance.This paper studies the fabric defect detection algorithm based on weakly supervised learning,and proposes three effective detection methods.The pixel-level defect location of complex texture fabrics can be realized by relying on the training set of image-level labels.It provides support for the application of fabric defect detection algorithm in practice.