Research On Anomaly Detection And Location Technology Of Industrial Image

In recent years,with the rapid development of computer vision technology,it has been applied to industrial quality inspection industry and achieved preliminary results.As an important research content in the field of computer vision,industrial image anomaly detection can detect abnormal data deviating from the normal expected behavior and provide guarantee for the normal operation of various systems.Therefore,it is of great significance to apply anomaly detection technology to anomaly detection and location of industrial images.Firstly,this dissertation introduces the knowledge of traditional machine learning methods and deep learning and points out their problems in anomaly detection.Then deep learing and traditional meachine learning industrial image anomaly detection and location algorithms based on local outlier factor and based on one class support vector machine are proposed.The main research contents and achievements of this disserrtation are as follows:1.An industrial image anomaly detection and location algorithm based on local outlier factor is proposed.The image level features are extracted by using the residual network pre trained on Imagenet data set,and the anomaly score of the measured image is calculated by using local outlier factor to realize anomaly detection.In order to accurately locate the abnormal pixels in the image,output features of the first three layers of the residual network are fused to extract the embedded vector features.The abnormal score of each super pixel in the feature map is calculated and its corresponding abnormal score map is obtained.The abnormal score map is upsampled and visualized to achieve abnormal pixel mask.Experiments on MVTec AD data set show the proposed is higher than other excellent algorithms with the average AUC(Area Under Curve)value of anomaly detection 92.5% and the average AUC value of anomaly location 91.1%.2.An industrial image anomaly detection and location algorithm based on one class support vector machine is proposed.The features are extracted from the pre trained residual network,and the minimum hypersphere center of each super pixel of the feature map is found by using support vector data description(SVDD)algorithm.Calculate the Euclidean distance from each super pixel of the feature map to the center of the hypersphere as its anomaly score,and obtain the super pixel level anomaly score map,Upsample and visualize the anomaly score map the anomaly location mask map.The maximum score of the abnormal pixel in the abnormal image is taken as its abnormal score.Experiments on MVTec AD data set show that the proposed is better than the algorithm in Chapter 3with the average AUC value of anomaly detection 89.1% and the average AUC value of anomaly location 94.1%,and the computational time is shortened by half.