Out-of-distribution image detection method based on attention enhancement and input disturbance

Abstract

The invention provides an out-of-distribution image detection method based on attention enhancement and input disturbance, the method adopts an input disturbance technique, the influence on samples in distribution is greater than that of samples out of distribution, so that the confidence score of the samples in distribution is higher, and meanwhile, a temperature scaling technique is used, the prediction probability distribution of the samples in the distribution is more sharp, the prediction probability of the samples out of the distribution is smoother, and the confidence score difference between the samples inside and outside the distribution is further increased; compared with the mode that a generative model is directly used for carrying out a distributed external sample detection task, the method has the advantages that the method does not need to introduce additional hyper-parameters, the model is relatively simple, and the training time can be saved; compared with a generative adversarial method for performing an out-of-distribution sample detection task, the method has the advantages that the method is not excessively limited to training data, misjudgment is not easy to generate for edge samples, and a better detection effect can be obtained.

Description

technical field [0001] The invention relates to the field of out-of-distribution sample detection methods, and more specifically, to an out-of-distribution image detection method based on attention enhancement and input disturbance. Background technique [0002] Out-of-distribution images are data distributed far from the data used for training, often significantly different from in-distribution data. The method of detecting out-of-distribution data from a large number of in-distribution data is called out-of-distribution sample detection or outlier detection method. In contrast to anomaly detection tasks, out-of-distribution detection is usually applied to labeled data, i.e., considers identifiable categorical features of in-distribution samples of a dataset. Compared with unsupervised anomaly detection methods, the out-of-distribution data coverage in this task is too large to be considered as the same class. [0003] At present, deep learning has become an important mea...

AI Technical Summary

Problems solved by technology

Compared with unsupervised anomaly detection methods, the out-of-distribution data coverage in this task is too large to be considered as the same class

[0003] At present, deep learning has become an important means to solve problems, but the problem is that many models can only be successfully applied when the test data and training data are similar, and often cannot be given when encountering some samples that are far from the distribution of training data. reliable results, and may even give a very confident but erroneous prediction

This method is relatively simple and direct, but the effect is not very good compared to other methods

Images