关键词: 视觉感知, 缺陷检测, 显式特征提取, 自动特征提取, 深层神经网络

Abstract: With the advantage of high efficiency and preventing secondary injury,inspection of surface defects based on visual perception has been widely used in various industry scenarios.The recent development of deep-learning technology contributes further to the progress and the application of vision-based defect inspection.The existing vision inspection methods under the track of explicit feature extraction and automatic feature extraction were reviewed.The basic structure of vision-based automatic inspection system was briefly introduced,and the visual perception involved in defect inspection was classified into three levels of classification,object detection and segmentation.Subsequently,existing defect inspection methods were divided into two categories that were the methods based on explicit feature extraction and the ones based on automatic feature extraction.Further,the former was categorized into statistical,spectrum and model-based methods while the later was categorized into classification-based,object detection-based and segmentation-based inspection networks.The representative methods of each class were reviewed,and the characteristics and the application scenarios of the methods were summarized.Aiming at the high cost of data collection and annotation,weakly-supervised inspection methods and anomaly inspection methods were introduced in recent years as well.Several well influential data sets for industrial surface defects were mentioned.Two key issues were discussed,which were how to alleviate the dependency on large amounts of annotated data and how to improve the adaptability of inspection methods in industrial applications.On this basis,several potentially development topics to be further investigated were suggested.

Key words: vision perception, defect inspection, explicit feature extraction, automatic feature extraction, deep convolutional neural network