Deep learning algorithm for detecting anomaly and segmenting abnormal region in image

Abstract

The invention relates to the technical field of unsupervised image anomaly detection and segmentation, in particular to a deep learning algorithm for detecting anomaly and segmenting an abnormal region in an image. The method comprises the following steps: S1, firstly, performing positive sample image feature extraction, designing a deep learning algorithm for detecting anomalies and segmenting abnormal regions in an image, enabling the algorithm to be high in accuracy, and obtaining top1 classification precision on a public data set test list. Reasoning time is short, only model parameters need to be stored, and positive sample features or statistics thereof do not need to be stored; compared with an algorithm for performing mahalanobis distance measurement according to feature statistics, the algorithm is wider in applicability, and the image does not need to meet the constraint condition of alignment; the algorithm process is simple, only the optimal single-level feature representation needs to be selected, and multi-level feature fusion is not needed. The segmentation threshold value can be selected in different modes according to actual conditions. And the algorithm is good in generalization ability and easy to reuse.

Description

technical field [0001] The invention relates to the technical field of unsupervised image anomaly detection and segmentation, in particular to a deep learning algorithm for detecting anomalies and segmenting anomalous regions in images. Background technique [0002] There are two main existing unsupervised anomaly segmentation algorithms: 1. Based on the reconstruction method, the positive sample image is reconstructed through the autoencoder (AutoEncoder). Due to the large difference between the abnormal sample before and after reconstruction, the abnormal region is segmented. This method model The amount of parameters is large, the training time is long, the reconstruction effect of complex patterns is not good, and the segmentation accuracy is not high; 2. Based on the pre-trained network to extract feature representation, and use knn and other methods for image retrieval, the image to be tested is the closest to the feature space The positive sample reference picture is ...

AI Technical Summary

Problems solved by technology

[0002] There are two main existing unsupervised anomaly segmentation algorithms: 1. Based on the reconstruction method, the positive sample image is reconstructed through the autoencoder (AutoEncoder). Due to the large difference between the abnormal sample before and after reconstruction, the abnormal region is segmented. This method model The amount of parameters is large, the training time is long, the reconstruction effect of complex patterns is not good, and the segmentation accuracy is not high; 2. Based on the pre-trained network to extract feature representation, and use knn and other methods for image retrieval, the image to be tested is the closest to the feature space The positive sample reference picture is compared to segment the abnormal area. This method needs to construct a positive sample feature dictionary. The memory usage of the dictionary and the knn search time are proportional to the number of positive sample reference pictures. When the amount of data is very large, the algorithm It is very time-consuming. In addition, the feature extractor is usually a convolutional neural network. In order to obtain the feature representation of different receptive fields, multi-level feature fusion is required.

Images