Research On Fabric Defect Detection Based On Visual Saliency Of Fully Convolutional Network

In most fabric production lines,the fabric defects caused by mechanical action or broken yarn seriously affect the quality of products and the benefit of enterprises.Therefore,fabric defect detection is an important link of quality assurance in the textile manufacturing industry.Recently,fully convolutional networks(FCNs)have achieved excellent performance on pixelwise prediction task,particularly salient object detection.Compared with background texture,fabric defects are usually salient,the fabric defect detection can deservedly be regarded as a salient object detection task.Therefore,this thesis innovates in saliency model by combining the FCN and the texture characteristics of fabric defect,and researches fabric defect detection algorithms based on visual saliency of fully convolutional neural networks.The innovative research results are summarized as follows:1)A fabric defect detection algorithm based on saliency model of discriminative feature and bi-directional message interaction was proposed.In view of the low detection accuracy of traditional fabric defect detection methods,multi-scale attention-guided feature extraction modules were firstly used to obtain more discriminative features that respond to the correct scale of the defect.Then,a bi-directional message interaction module was adopted to obtain multi-level features with rich contextual information through message interaction along two directions.Finally,a cross-level contrast feature extraction module was added to enhance the local contrast of features along each resolution axis,while gradually fusing multi-level features to predict the final saliency result.The experimental results prove that the performance of the proposed method is better than other state-of-the-art saliency detection methods,and it has a better detection effect on fabric defects of different scales.2)A fabric defect detection algorithm based on dual-branch balance saliency model of discriminative feature was proposed.Firstly,a saliency branch was realized through multiscale discriminative feature extraction modules and a bi-directional stage-wise integration module,which solves the problems of the large amount of parameters and calculations of the multi-scale extraction module in the previous work,as well as the insufficient information fusion of the traditional significance model.In order to balance the network structure,a bootstrap refinement module was trained as another branch to guide the restoration of feature details.The experimental results show that the proposed method is superior to other advanced saliency detection methods in seven evaluation metrics.3)A fabric defect detection algorithm based on efficient lightweight saliency model was proposed.To alleviate the contradiction between the computational cost and performance of the saliency model,a flexible multi-scale octave convolution module was proposed to replace the traditional convolution operation.The multi-scale octave convolution module stores the low-frequency part of the feature in a low-resolution tensor,which reduces spatial redundancy and can more completely extract in-layer multi-scale features.In addition,combining the attentional feature fusion module to integrate multi-scale features within a stage and multi-level features across stages.An extremely lightweight saliency detection model is constructed by using the multi-scale octave convolution module and the attention feature fusion module.And a more superior performance is obtained with ?6%parameters of the current advanced saliency detection model.