Image anomaly detection method based on high and low frequency reconstruction

Abstract

The invention discloses an image anomaly detection method based on high and low frequency reconstruction, which comprises a model training stage and a model inference stage, and in the model training stage, a generative adversarial network is trained for normal images in a training set, a network model is trained by using a server, network parameters are optimized by reducing a network loss function, and the model inference efficiency is improved. Obtaining a network model based on high and low frequency reconstruction until the network converges; in the model inference stage, the network model obtained in the model training stage is utilized to judge whether a new test image is an abnormal image or not according to the abnormal score. Information of different frequency bands of an image is obtained through a frequency domain decoupling module, reconstruction is carried out through different generators, interaction and selection of information of different frequency bands are achieved through a channel selection module in the coding stage of a network generator, the information of different frequency bands can be fully utilized by a network, the reconstruction capacity of the generators is improved, and the reconstruction efficiency is improved. And the performance of an image anomaly detection algorithm based on reconstruction is improved.

Description

Technical field [0001] The present invention belongs to the technical field of computer vision, and particularly relates to an abnormality detection method for an image based on high frequency reconstruction. Background technique [0002] Anomaly detection image task is to find the line between normal and abnormal samples samples to separate normal samples and abnormal samples as possible. At present a great difficulty encountered in actual anomaly detection is uneven real-world data, image data anomalies are often difficult to obtain, such as in the production line automatic monitoring product defects, defects that arise as general probability is very small . Simple classification task difficult to apply supervised learning anomaly detection in the image, so the vast majority of reliable methods of image abnormality detection are unsupervised anomaly detection is at the core of the training set to determine whether the input data and the training set data "similar." [0003] Abn...

AI Technical Summary

Problems solved by technology

Solved the problem that the image anomaly detection method based on reconstruction is not ideal in the scene of defect detection due to the limitation of generator generation ability

Images