An image calibration error detection method and device

Abstract

The invention discloses an image calibration error detection method and device. The method includes: establishing an image data set, using the image data set to train an unsupervised anomaly detection neural network, and reassessing the hidden layer features generated during the training process. structure to obtain a hidden layer feature reconstruction layer; embedding the hidden layer feature reconstruction layer into the unsupervised anomaly detection neural network, then performing anomaly detection on the image data to be detected, and judging whether the image data to be detected is calibrated according to the anomaly score function Error; the device includes: a sequentially connected backbone network module, a hidden layer feature reconstruction module and an anomaly scoring module; the invention can expand the anomaly score gap between normal samples and abnormal samples, and improve anomaly detection performance.

Description

technical field [0001] The present invention relates to the technical field of machine learning, and in particular, to an image calibration error detection method and device. Background technique [0002] In the process of image calibration, it is inevitable that wrong calibration will occur due to factors such as the carelessness of calibration personnel. Manual detection of error calibration is time-consuming and labor-intensive, and unsupervised anomaly detection can reduce the investment of human resources. For unsupervised anomaly detection tasks, the method of representation learning is usually used at present. According to the type of training data, the corresponding unsupervised representation learning task is designed, and finally the difference between the training speed or training difficulty of abnormal and normal samples is used to detect abnormal samples. However, the current unsupervised anomaly detection methods have the following defects: On the one hand, ...

AI Technical Summary

Problems solved by technology

However, the current unsupervised anomaly detection methods have the following defects: On the one hand, due to the unsupervised setting and the difficulty in controlling the training degree of the deep learning network model, the network model under this setting cannot distinguish abnormal samples during the training process, and will simultaneously Try to fit the normal and abnormal samples as much as possible, so that the reconstruction error between the two types of samples will gradually decrease as the training process progresses, which will lead to a decrease in the performance of anomaly detection, such as figure 2 On the other hand, the current unsupervised anomaly detection is mainly based on autoencoder or geometric transformation. Both methods have advantages and disadvantages. The former is simple to implement but has relatively poor performance, while the latter has good performance but cannot For non-image data, the existing methods can only support a certain type of method, cannot support the former and the latter at the same time, and have relatively poor versatility

Images