Research On Visual Feature Representation Method For Surface Defect Detection

In recent years,the rapid development of artificial intelligence technology has already had a profound impact on traditional manufacturing,and countries around the world have also carried out industrial upgrading of manufacturing.China is a large manufacturing country.Manufacturing is an important pillar of national economic development and a strategic deployment to implement the deep integration of industrialization and informatization,as well as build a strong manufacturing country.The "Made in China 2025" issued by the State Council in 2015 and the "New Generation Artificial Intelligence Development Plan"issued in 2017 aim to accelerate the transformation of manufacturing enterprises from traditional labor-intensive to intelligent and sophisticated.Vigorously developing the intelligent manufacturing industry can not only promote the rapid upgrading of innovation and entrepreneurship in China,but also promote Chinese manufacturing to the world.Classification and detection of surface defects based on visual imaging as one of the largest application points in smart manufacturing,has received extensive attention from academia and industry.How to accurately represent,precisely classify and effectively detect surface defects with complex backgrounds,and realize efficient processing of surface defect features in various scenarios is a frontier challenge and technical commanding height in the field of machine vision.Therefore,researching efficient and generalized deep learning-based surface defect classification and detection methods has become a challenging problem to be solved in the current manufacturing field.The classification and detection of surface defects based on machine vision refers to the use of machine learning and deep learning technologies to judge and locate the category and location of surface defects from the collected images,so as to assist in the control of product quality.However,machine vision-based surface defect detection is oriented to a wide variety of defects,which are formed by factors such as raw material quality,production environment,resulting in large differences in shape and size,and these problems pose a great challenge to surface defect detection.The use of supervised learning paradigm results in significant head category bias when dealing with classification tasks with limited defect samples and uneven class distribution,and cannot accurately classify defect samples.In addition,the pre-trained models obtained using supervised and self-supervised learning are not spatially aware enough,resulting in poor performance in defect detection tasks.To address these issues,this paper improves the model structure to alleviate the problem of poor shape and size adaptation in surface defect detection.For the defect classification problem with limited samples and uneven class distribution,this paper proposes two selfsupervised learning methods to obtain a model without feature bias from the perspective of data and feature utilization,so as to improve the classification performance.For the problem of insufficient spatial awareness of the model in the defect detection task,this paper deeply draws on the principle mechanism of object detection and proposes a single-branch self-supervised learning framework combined with spatial feature sampling to improve the feature generalization ability of the model.The specific contributions of this paper are as follows.1)Aiming at the problem of insufficient adaptability of models to diversified and differentiated features in surface defect detection tasks,this paper elucidates the characterization mechanisms of convolutional networks in the depth and width directions,proposes to fuse convolution operations of different depths and widths in a single convolution module,and constructs a shape and size aware backbone network to improve the visual feature representation capability of the model.The experimental results show that the method designed in this paper can effectively classify and detect objects with large appearance differences,enhance the feature representation ability based on the residual network,and significantly improve the performance of surface defect classification and detection.The proposed method outperforms the baseline model by 1.5%and 2.0%in the ImageNet classification dataset and the COCO object detection dataset,respectively;meanwhile,the proposed method can consistently improve the performance by more than 2%in multiple defect detection datasets.2)Aiming at the category bias and overfitting problems in the surface defect classification task when the model is dealing with limited dataset with category imbalance,this paper reveals the representation mechanism of data and model for feature extractor optimization,proposes data augmentation strategies and feature extraction methods based on stitched images and hybrid-task learning,and constructs an efficient self-supervised learning framework based on the Siamese networks to obtain a generalization model without feature bias.In addition,a novel sampling strategy is designed to further reduce the computational consumption of the model and significantly improve the generalization performance of the pre-trained model.Under the premise of only consuming 50%of the calculation,it also achieves state-of-the-art performance in various limited classification datasets with category imbalances.The two proposed methods outperform the baseline model by 2.4%and 2.7%,respectively,on the ImageNet classification dataset;while both outperform supervised learning by more than 1.2%on different defect classification datasets.3)Aiming at the problem that the existing pre-training model has insufficient perception ability of defect spatial features,this paper deeply borrows the mechanism of object detection and proposes global and local sampling strategy to construct an efficient self-supervised learning framework for object detection.A hybrid optimization method containing localization,clustering and contrastive tasks is embedded into a unified single-branch network to achieve self-supervised feature extraction for object detection.Compared with the Siamese network-based methods,the method designed in this paper consumes less computation and achieves state-of-the-art performance on a variety of surface defect detection datasets.In the COCO object detection dataset,the proposed method outperforms the supervised learning pre-trained model by 1.7%;meanwhile,it outperforms the supervised learning by more than 1.6%in different defect detection datasets.