Inner Surface Defect Detection Of Stainless Steel Bellows Based On Deep Learning

Surface defects of industrial products can affect the normal use of the products,cause safety hazards and damage the aesthetics of the products.Affected by equipment and technology,in most cases,only human eye detection can be used in actual production.This manual detection method is not efficient and accurate,and seriously reduces industrial production capacity.In order to alleviate this problem,practitioners use computer vision algorithms based on machine learning to promote industrial defect detection,which design some feature extraction operators,extract defect image features,and realize automatic defect detection and classification to a certain extent.However,these methods are not robust and have poor adaptability to some different scenarios,making it difficult to deal with some product defects which has variable shapes in actual production.In this thesis,deep learning technology is used into the field of industrial defect detection,and we research on the location and classification of bellows inner surface defects by using self-collected bellows inner surface defect data,which are regarded as a breakthrough.We first analyzes the sample type and quantity of the bellow inner surface defect dataset,using the more effective method which named multi-task composite network in the field of industrial defect detection,and proposes a defect detection model that combines image classification and objection detection algorithm.In this algorithm,intuitive classifier realizes the preliminary classification of defect images,and then uses the objection detector to realize further detection of defect images based on the features combined with the features extracted in classification.We use the data augmentation strategy to alleviate the impact of insufficient data volume.And we prove the feasibility of automatic defect detection for stainless steel bellows through our network performance indicators and feature maps’ analysis.Since there is a gap between the defect objection detection algorithm and the expected.With reference to the semantic segmentation algorithm,a region segmentation algorithm is designed to obtain a certain region category in the defect image.The region segmentation algorithm trained by the labeling rulers besigned by ourselves.We take the distribution of defect targets and analyze the entire defect dataset,then design an artificial weight module to enhance the network model’s ability to perceive the frequent locations of defects;the multi-head attention and scale attention mechanism are used to further pay attention to the defect characteristics of the image and effectively alleviate the block effect.And we validate the feature extraction effect of various texture enhancement modules,so that the network can make inferences combined with background texture information.Finally,through the fusion experiment,the effect of the algorithm is analyzed and it is determined that the texture enhancement module has scene limitations,and the location and classification of stainless steel bellows inner wall defects dataset and some public defect dataset are realized.Finally,in view of the poor effect of multi-category segmentation of stainless steel bellows defect segments,this paper further designs a fine-grained classification algorithm that is more in line with the defect detection scene.With the features extracted by the classification network and the local parts extracted by the basic network in the region segmentation network combining according to specific rules,and the defect classification is realized.