Research On Fabric Defect Detection Algorithm Based On Visual Saliency Of Multi-feature Aggregation

In the process of textile production,various defects inevitably appear on the surface of the fabric,but these defects have a great impact on the quality of the textile.Nowadays,artificial vision detection methods are still widely used in industrial scenes.Due to subjective influence,they are inefficient and prone to false detection and missed detection,resulting in low accuracy,which is difficult to meet the current requirements of fabric defect detection.There is an urgent need for effective automatic detection methods of fabric defects to manage and control the quality of textiles.The visual saliency method based on a biological vision system can quickly search and locate the region of interest.Convolutional neural networks(CNNs)can effectively learn robust feature and have the ability to represent semantic information and texture details at the same time.With the rapid development of full convolution neural network,it has been introduced into visual saliency and modeled by more and more research.However,due to the complex texture of fabric images,how to learn powerful features and fuse features to obtain saliency is still a challenging task.(1)A salient fabric defect detection algorithm based on multi-scale feature fusion is proposed.In view of the low accuracy of fabric defect detection caused by the unclear comparison between fabric defects and fabric background.Res Net-50 is utilized as the backbone network to extract rich low-level spatial information and high-level semantic information.Inspired by the visual perception field,a multi-scale feature learning module is designed to further extract effective feature information.Aiming at the information redundancy caused by the direct integration of feature information,a feedback attention refinement fusion module is designed.The attention refinement module is added between each fusion module.At the same time,the features after the first fusion are fed back to the next fusion process,which improves the detection accuracy of fabric defects.The experimental results show that it still has good detection accuracy when the contrast between fabric background and defects is low.(2)A salient fabric defect detection algorithm based on a bi-directional inference network is proposed.Aiming at the problem of difficult detection and poor universality of small fabric defects,a two-way significance inference method combining top-down and bottom-up is proposed in this thesis.The top-down process is to gradually use higher-level and richer semantic features to infer high-level saliency based on Res Net-50 as the backbone network.Single-scale features can not capture the multi-scale context information of different objects.According to the complex and diverse characteristics of fabric texture and defects,a self-fusion enhanced representation module is designed to generate effective features to solve the problem of scale variation.Residual connection is mainly introduced in the module,which breaks the symmetrical structure of the network and improves the feature representation ability of the network.In order to avoid weakening the effective information of the multi-scale features from side output,the interactive feature fusion module is used to iteratively fuse the multi-level salient estimates from coarse to fine in a bottom-up manner to generate a better prediction map.Experimental results show that the proposed method can accurately locate and segment small object defects,and improve the robustness of the algorithm.(3)A fabric defect detection algorithm based on the improved VGG-16 lightweight saliency model is proposed.First,a lightweight backbone network is constructed.The convolution in the two high-level convolution blocks of VGG-16 is replaced by depthwise separable convolution,which can reduce the parameters without weakening the feature representation.High-level features have rich semantic information.The cross-level relation module is introduced to combine the outputs of the last two layers of the backbone network to obtain a rough initial salient prediction and realize the fabric defect location.Then the refinement fusion module is used to refine and merge the low-level feature information.This method achieves the detection effect equivalent to other advanced saliency detection methods with less parameters,reduces the requirements for system memory and computing ability,and provides a choice for the real-time application of the algorithm in industrial field.