Noise identification method for image data

Abstract

A noise identification method for image data comprises the following steps: S1, in noise learning based on a neural network, using a Gaussian mixture distribution model, and taking a loss value of a to-be-detected sample and a maximum non-target probability as training dynamics for joint modeling; wherein the maximum non-target probability refers to the maximum probability, except the target class probability, in each classification identification probability output after the sample passes through the neural network; s2, calculating a loss value and a maximum non-target probability of a to-be-detected sample by using the trained neural network model, fitting a Gaussian mixture distribution model by using the calculated loss value and the maximum non-target probability, and outputting a probability that the to-be-detected sample belongs to a clean sample and a noise sample; and dividing the to-be-detected sample into a clean sample or a noise sample according to the probability. According to the method, the distance information between the sample and the center of the class and the confusion information between the sample and the centers of other classes are considered at the same time, the difficult sample and the noise sample are effectively separated, and the noise detection accuracy is improved.

Description

technical field [0001] The present invention relates to image recognition, in particular to a noise recognition method in image classification tasks. Background technique [0002] Abbreviations and terms: [0003] Learning with Noise: This task refers to how to learn a high-performance model on a dataset containing noisy labels. The training of deep neural networks usually requires a large number of clean samples, but in practical application scenarios, obtaining large-scale, high-quality clean labels faces problems such as high labeling cost, high time overhead, and low labeling quality. For example, in the crowdsourcing scenario, the company usually invites several annotators to annotate a large number of unlabeled samples. However, due to the uneven annotation capabilities of different annotators and mislabeling, etc., a large amount of annotation costs and time costs are incurred. Get a dataset with noisy labels. In addition, it is a low-cost way to obtain samples by ...

AI Technical Summary

Problems solved by technology

However, due to the existence of difficult samples in clean samples, some clean samples will also have a large loss value and thus be judged as noise samples. The accuracy of detecting noise samples with a simple loss value threshold method is low

In other words, some clean samples are too different from the category center, which leads to being mistaken for noise samples

Images