Research On Visual Inspection Methods For Industrial Surface Defects

After decades of rapid development,China’s manufacturing scale has leaped to the first place in the world,and a technologically advanced manufacturing system is gradually established in China,which has become an important cornerstone of the economic and social development.In the process of production and utilization,it is inevitable that a variety of surface defects will appear on industrial products.These defects not only affect the beauty and comfort of the products,but also bring negative effects to their usability,and in some severe cases,they may cause serious safety accidents.Therefore,surface defect detection technologies have received more and more attention in industrial scenarios.With the continuous development of artificial intelligence,computer vision based surface inspection has made some progress in both theoretical research and practical applications.However,due to the factors like uneven image qualities,sparse and unbalanced defective samples,and difficulties in feature representation learning,some existing visual inspection methods still suffer from performance bottlenecks such as high false alarm rate and poor robustness.To address these problems,an in-depth study on varigrained industrial visual inspection tasks is presented in this thesis by considering sample distribution characteristics,defect prior information and class-level feature learning,respectively.The main contributions of this thesis are summarized as follows.(1)A defective image determination model has been proposed based on deep convolutional autoencoder with structural similarityIn the scenario of industrial visual inspection,determining whether an input image is defective or not is the primary task.Considering the relative scarcity of defective images and the difficulty of labelling,a defective image determination model has benn proposed based on deep convolutional autoencoder with structural similarity.The model follows an encoder-decoder architecture and can be trained without the usage of defective samples.Therefore,it differs from traditional supervised models and can avoid being dependent on large-scale,high-quality annotated samples.When dealing with defective images with complex textures,the model can be flexibly embedded with a direct feature connection structure,which improves its ability in reconstructing complex textured backgrounds of images.In order to maximize the differences in reconstruction errors of defective and defect-free images,the structural similarity index is employed in both the training and testing process of the model,which is able to measure the difference between the original and reconstructed images by comprehensively considering the luminance,contrast and structural information.Experiments on three real-world industrial surface defect datasets show that compared with the conventional pixel-wise indices,using the structural similarity can significantly improve the performance of the model in defective image determination.In addition,experimental results demonstrate the superiority of the proposed method among existing unsupervised approaches.(2)A defect target segmentation model has been proposed based on entity sparsity pursuitDefect target segmentation aims to find the accurate contours and locations of defects from the input image,and it is the basis for quantitative analysis of defects in industrial inspection.However,most of previous approaches are task-specific,weak in the generalization ability and lack of a relatively universal theoretical framework.Based on the fact that surface textures of industrial images tend to form a low-rank structure in the original image space or feature space,and the presence of surface defects can damage such a low-rank structure,a defect target segmentation model based on entity sparse pursuit has been proposed.At the feature level,considering that most industrial images are grayscale with limited availability of features,a new feature extractor has been designed based on the idea of local binary pattern,which shows a better discriminative ability in distinguishing image background and defective targets.On this basis,grayscale features as well as texton features are fused to compose the feature extraction module.In designing the model,the regularization constraints commonly used in the traditional low-rank and sparse frameworks are discarded,but the intrinsic priors of entity sparsity and local saliency are explored to guide the process of low-rank and sparse decomposition of the feature matrix.Experimental results on several industrial surface defect datasets show that the proposed model has a good generalization capability with a low false negative rate and a low false positive rate.(3)A defect type recognition model has been proposed based on supervised contrastive learningDefect type recognition refers to the qualitative analysis of defects from the perspectives of their morphologies,causes or damage degrees,which can be generally formalized as a multi-classification problem.In real-world industrial visual inspection scenarios,defect targets are with diverse shapes,sizes and locations,and the annotated samples are not able to cover all possible appearances of defects within the same class.At the same time,there are also situations where the appearances of defective samples are with intra-class diversity and inter-class similarity,which further deepens the difficulty for compact feature representation learning.To solve these issues,a defect type recognition model has been proposed based on the supervised contrastive learning.It implements class-level contrastive learning and defect classification under a unified framework,which helps to build compact descriptions of samples belonging to different classes.Under the impact of the contrastive learning branch,the backbone network is able to learn a more compact feature space,which in return improves the recognition ability of the classification network.In addition,an effective strategy for generating features is instantiated,which improves the ability of the model in identifying defective samples with intra-class diversity and inter-class similarity.Experimental results on a hot-rolled steel strip surface defect classification dataset and a rail fastener anomaly classification dataset demonstrate the effectiveness of the proposed model.