Research On Industrial Visual Inspection Algorithm By Integrating Some Prior Information

With the gradual deepening of the transformation and upgrading of China’s manufacturing industry,traditional methods for industrial visual inspection are increasingly showing their limitations.At the same time,deep learning has achieved a great deal of success in the field of computer vision and is beginning to make its way into the field of industrial visual inspection.Although industrial visual inspection based on deep learning can solve more complex problems than traditional methods,there are still several bottlenecks that limit the development of industrial visual inspection at a faster pace.For example,the high omission rate will increase the cost of industrial manufacturing,the cost of obtaining more than tens of thousands of level label data is high,and industrial visual measurement requires high stability and accuracy of object contours.How to solve these problems and make industrial visual inspection better serve the industrial manufacturing industry is still an urgent and challenging research topic.Aiming at different research problems,the main research contents of this dissertation include:(1)Defect classification based on multi-position bi-branch siamese networks;(2)Defect classification based on self-training contrastive learning;(3)Industrial visual inspection based on second-order statistical features;(4)Vectorized segmentation based on periodic B-splines.The specific research contents are as follows:Defect classification based on multi-position bi-branch siamese networks:Position in industry refers to information composed of substrate model,component type or location.By using position information,the partitioning of industrial data can be realised to reduce the difficulty of network learning.Inspired by this,we propose a defect classification algorithm based on a multi-position bi-branch Siamese network.First,we design a multi-position weighted resampling method to extract paired data with the same position to solve the problem of high omission rate caused by imbalanced defect types in industrial visual inspection.Then,to fully learn from pairwise data,we propose a multi-position bi-branch siamese network architecture to perform similarity measurement and multi-classification simultaneously.Extensive experimental results demonstrate the effectiveness of this method and its generalization ability to unknown position data,which effectively reduces the omission rate of defect samples.Defect classification based on self-training contrastive learning:To address the problem that the existing industrial visual inspection methods rely on a large amount of expensive labeled data,we propose a self-training contrastive learning method.This method achieves excellent model performance by utilizing a large number of unlabeled samples.Specifically,we first propose a self-training architecture based on the first work and use a multi-position based mix strategy for data processing.The mix strategy effectively combines labeled samples and unlabeled samples,so that the network can learn more discriminative defect features.Then we weight the pseudo-labels of the unlabeled data,a nd gradually increase the weight of the pseudo-labels to reduce the influence of the pseudo-label noise on the training.The experimental results show that the proposed method can improve the accuracy of the model with fewer labeled samples,thereby reducing the burden of data labeling in industrial production.Industrial visual inspection based on second-order statistical features:Although the second work can effectively reduce the workload of data annotation,it is still necessary to rely on manual annotation of data.Therefore,we propose to use the prior information of 3D models of products for industrial visual inspection.First,we use graphics technology to render the model,and then we solve the domain shift problem between the synthetic dataset and the real data by using the domain generalization of second-order statistical features.Finally,we verify the effectiveness of our method on several synthetic industrial datasets and publicly available datasets.Vectorized segmentation based on periodic B-splines:Aiming at the problem that industrial visual measurement requires high stability and accuracy of object contours,we propose a cascade framework combined with periodic B-splines for instance segmentation.Specifically,we consider both global and local features of objects and introduce two types of graph structures to better exploit object geometry and appearance information.Then,we regress the spline control points to the object boundary via a mix network,perform spline sampling to obtain an initial prediction of the contour,and finally regress the predicted contour to the real contour of the object.In addition,we add a fairness regularization to further constrain the contour and prevent instability phenomena.Several experiments show that our method can produce stable and accurate vectorized contours.We carry out an in-depth and systematic study of the current difficult problems in industrial visual inspection.By analysing and extracting prior knowledge in real scenes,combined with deep learning,industrial visual inspection can be performed quickly and effectively.We conduct detailed and comprehensive experiments on industrial data collected in real scenarios and public benchmark datasets to verify the feasibility and effectiveness of our method.