Research On The Fabric Defect Detection Algorithm Based On The Visual Saliency Of Context Coordination

Fabric defect detection is a key step in quality control in the textile manufacturing industry.However,due to the diversity and complexity of defects and environmental factors,traditional manual and automatic fabric defect detection methods often have problems of low efficiency or accuracy.Therefore,it remains a highly challenging task.Visual saliency can quickly locate the most salient object area in an image by imitating human perception mechanism,which has attracted more and more research attention and achieved breakthrough success in many fields.Therefore,fabric defect detection algorithm based on visual neural network saliency model has very considerable research value.However,existing fabric defect detection algorithms based on convolutional neural network visual saliency often do not fully consider the characteristics of fabric defects and the coordination of contextual information,which often result in insufficient defect feature representation and poor generalization ability.Therefore,this thesis combines the characteristics of visual neural networks and fabric defects to construct a new saliency detection model from three aspects: contextual information coordination,attention mechanism,and long-term dependency relationship.The main research content is as follows:(1)A contextual attention cascade feedback saliency model for fabric defect detection is proposed.This method utilizes attention mechanism to dynamically select multi-scale features,which can effectively solve the problem of scale variation in defects.Through cross-level and cross-scale attention feedback design,important multi-scale complementary information is transmitted from deeper layers to adjacent shallow layers,promoting information exchange of multi-scale features in U-shaped detection networks,improving model feature representation ability and performance,and reducing computational complexity.The experimental results on two different fabric datasets demonstrate that the proposed method can more accurately and completely detect fabric defects of different sizes compared to existing state-of-the-art methods,especially have excellent performance for small defects.(2)A saliency model based on context progressive attention aggregation is proposed for fabric defect detection.This method can solve the problem of insufficient actual receptive field in the U-shaped structure-based salient object detection model by accessing cross level feature pyramid blocks after different stages of the existing classification network.In addition,by using the progressive aggregation method of parallel cross level bidirectional attention features,this model can solve the problem that the high-level semantic information will be gradually diluted by the shallow details in the U-shaped structure-based salient object detection method can be solved,resulting in the absence of subjects.The experimental results show that compared with existing methods,the proposed method has clear boundaries,accurate positioning,and significantly improved detection performance,especially for linear defects.(3)A saliency model based on Transformer for adjacent context coordination is proposed for fabric defect detection.This method constructs a neighboring contextual collaborative block by introducing a local branch and two adjacent branches,while utilizing local and global saliency information to highlight defect targets,effectively enhancing the discriminative power of defect features,and solving the problems of insufficient feature representation and low contrast between defects and background in defect detection based on U-shaped structure.In addition,by stacking four Transformer blocks to capture global contextual information with long-term dependencies and using it as a prior to guide local information,the saliency of local information is further enhanced to generate more accurate detection results.The experimental results illustrate that the proposed method still achieves good detection performance in the case of low contrast between fabric defects and background textures,outperforming existing state-of-the-art methods in most evaluation indicators,especially improving the defect detection efficiency in pattern dataset.