An Image Anomaly Detection Method Based on Discrete-Continuous Feature Coupling

Abstract

The present invention relates to an image anomaly detection method based on discrete-continuous feature coupling, comprising the following steps: the first step: image feature extraction, including: data preprocessing; encoding network feature extraction; hash latent space and description latent space feature extraction : The obtained latent space features are expanded according to the space to obtain the block description features. Based on the description features, the corresponding hash features are obtained through the binarization mapping function, and the discrete statistical activation layer is added to make the reverse backpropagation can be distributed with derivatives. Discrete encoding is done in a discrete way; an additional discriminator is introduced to classify the extracted hash features and binomial distribution vectors; the second step: image feature fusion: solving the similarity matrix; constructing a similarity map; reconstructing the original image; Step 3: Anomaly detection.

Description

technical field [0001] The invention belongs to the field of deep learning, and mainly relates to image anomaly detection, self-supervised learning, representation learning and image restoration. Background technique [0002] At this stage, artificial intelligence technology has brought great changes to image analysis, and has a great impact on the manufacturing industry [1-2] and equipment condition monitoring [3] and other fields have a profound impact. In the process of intelligent analysis, affected by factors such as appearance defects and imaging sensor response distortion, abnormal samples are inevitable, and failure to screen out such samples will affect product quality assurance [4] and stable operation of the equipment [5] cause immeasurable losses. Abnormal samples have various appearances, inconsistent distribution and extremely small number, so traditional supervised classification algorithms and mature deep neural networks are difficult to apply to anomaly ...

AI Technical Summary

Problems solved by technology

[0005] The above methods have promoted the progress of anomaly detection algorithms to a certain extent, but both types of methods have limitations. The main reasons are: (1) Universal reconstruction problem: high-quality reconstruction methods require information-rich shallow features, while Shallow features are versatile, so there is a phenomenon of undersampling in the hidden space during the training phase, which leads to good reconstruction results for abnormal samples in some cases; (2) Low-quality interference problem: the limited expression mechanism limits the diversity of features , which reduces the model's reconstruction of differential content, which will lead to poor reconstruction quality of some normal samples and reduce the discrimination of the algorithm

Images