Learn From Incomplete Labels In Image Classification

In recent years,there has been significant development in the field of deep learning image classification.However,the success of deep neural networks often depends on a large amount of manually labeled data,which results in a significant cost in terms of money and time,especially when labeling has to be done by professionals.In contrast,automated methods can obtain noisy or unlabeled samples at a low cost in a short time,which is known as the problem of incomplete label learning.However,traditional classification methods are difficult to achieve ideal performance when dealing with image data that has noisy labels or no labels.Therefore,the use of incomplete label images to train models has received increasing attention from researchers in the computer vision field.Among them,research on noisy label learning focuses on how to use noisy labels to train models,while semi-supervised learning focuses on how to train models with a small number of labeled samples and a large number of unlabeled samples.Currently,both noisy label learning and semi-supervised learning face some challenges.For noisy labels,the best existing methods still lack observation of the model changes during training with noisy samples.Instead,they strengthen the robustness of the training process through regularization-based methods such as changing the model structure,teacher-student models,and Mixup mixing strategies.Although these methods improve the performance compared to ordinary supervised methods,they still result in the model being continuously affected by noisy labels during training,especially in the case of high noise rates.As for semisupervised learning,recently proposed FixMatch and FlexMatch have achieved significant results in the field of semi-supervised learning based on threshold screening strategies.However,these two methods use extreme approaches of pre-defined constant thresholds and adaptive thresholds for each category.With FixMatch,the high confidence unlabeled samples screened by using a high fixed threshold ratio are relatively small,resulting in a longer training process.Although FlexMatch alleviates this issue by introducing dynamic thresholds for each category,it still has problems with unstable results and low discriminative feature representation,especially when there are few labeled samples.In this article,we propose solutions to the existing problems in noisy label learning and semi-supervised learning,promoting the further development of incomplete label problems.For noisy label learning,we propose the BMPS method,which models the probability distribution of samples in the training set using a beta mixture distribution based on the observation of sample category probability distribution changes during training from a global perspective.Then,based on the differences in learning progress of individual clean or noisy samples during training,we introduce smoothed category probabilities to model the learning quantity of samples from a local perspective.Finally,both modeling methods are used to select clean samples from the training set,allowing noisy samples to be treated as unlabeled samples for training.For semi-supervised learning,we propose the CHMatch method,which trains the model using adaptive threshold screening strategies based on a memory-based library and contrastive learning based on a hierarchical matching graph.This method first proposes a robust threshold learning strategy based on the memory library to select high confidence samples.Then,we introduce the rarely used prior information of the hierarchical structure as an additional supervisory signal and construct a matching graph for contrastive learning.BMPS and CHMatch both achieve stable and superior results in their respective fields in commonly used benchmark tests.In the noisy learning scenario,when using an 80%noise rate,the BMPS method improves the best previous results by 12%and 22%,respectively.In the semisupervised learning scenario,using the CIFAR-100 dataset,the WRN-28-2 model,and 4 and 25 labeled samples per category,CHMatch reduces the error rate by 8.44%and 9.02%,respectively,compared to FlexMatch.