Long-tail distribution image classification method with noise label

Abstract

The invention discloses a long-tail distribution image classification method with a noise label, which is used for solving the problem of image classification with both long-tail features and noise labels by learning through relaxation interval loss depended on by a sample and assisted by an anti-noise data enhancement strategy. According to data noise features, sample-dependent slack variables are introduced to relax interval constraints when sample function intervals are calculated, and then sample-dependent smooth slack losses are calculated according to sample interval classification; and according to the data long tail features, a data enhancement strategy adjusted in stages is implemented, strong enhancement and weak enhancement are performed on samples respectively, and a sample screening mechanism based on relaxation loss is provided in a formal training stage for screening noise data. The method is simple and convenient to implement and flexible in means, and remarkable classification effect improvement is achieved on long-tail data, noise data and training data with the characteristics of the long-tail data and the noise data at the same time.

Description

technical field [0001] The invention relates to the field of image classification, in particular to a method for image classification under noise labels and long-tail distribution data. Background technique [0002] In recent years, convolutional neural networks (CNNs) have been widely used in the field of computer vision. When the number of training data is fixed, the increase in the number of parameters leads to an increasingly prominent phenomenon of overfitting. In order to improve the overall performance, the demand for accurately labeled data is also increasing. However, obtaining a large number of accurately labeled samples is usually expensive. For this, non-expert crowdsourcing or systematic labeling is a practical solution, however, it is prone to mislabeling of labels. Many benchmark datasets, such as ImageNet, CIFAR-10 / -100, MNIST, QuickDraw, etc., contain 3%~10% noisy label samples. Existing studies on noisy labels usually focus on splitting correctly and mis...

AI Technical Summary

Problems solved by technology

Existing research on noisy labels usually focuses on splitting correctly labeled and mislabeled samples, but ignores the distribution of data

Images