Robust Image Classification Algorithms With Weakly Supervised Learning

Along with the birth and development of deep learning,the field of computer vision has begun to develop by leaps and bounds.Due to their powerful learning ability,deep learning models have been breaking the historical records of various vision tasks.In reality,these breakthroughs rely on a large amount of labeled data for model training.However,collecting such a large labeled dataset is time-consuming and labor-intensive.To reduce the cost of annotating data,some scholars automatically annotate the data by models.However,the accuracy of labels annotated by models is difficult to be guaranteed,so there are incorrect labels in the collected data,which affects the training of deep learning models.Therefore,there are often two conditions in real-world datasets.The first is that only a small portion of the data is correctly labeled,and the remaining large amount of data has not yet been labeled.The second is that a large number of labels are noisy,i.e.,the data is mislabeled.To address these problems,researchers have proposed weakly-supervised learning to reduce the need for clean labels for model training and thus reduce the cost of annotating data and improve the robustness of the model.Therefore,the study of weakly-supervised learning has important practical significance and application value.Furthermore,image classification,the basic task of computer vision,has a wide range of application scenarios.However,the lack and noise of labels degrade the performance of image classification models.Therefore,this paper takes the weakly-supervised learning method as the research core,studies the weakly-supervised learning methods for the datasets with few labeled data as well as the datasets with noisy labels,and applies the weaklysupervised learning methods to the task of image classification to improve the robustness of classification models.The main research content and the contributions of this paper are summarized as follows:(1)Firstly,a weakly-supervised image classification method based on dual learning is proposed for the dataset with a small number of labeled data.The method proposes a dual learning strategy to help the model learn more effective information and accelerate the convergence of the model by training with different augmented data at the same time.Specifically,during the training process,soft and hard labels are generated by the predictions of the model from weakly-augmented data.Then,we train the model by strongly-augmented data with soft and hard labels.In addition,to guide the training of the model,we propose a simple and effective method for generating sample weights.Assigning different weights to the samples helps the model pay more attention to the samples with high confidence but also learn from the other samples,which improves the model performance.In order to promote the selflearning of the model,we train the model by minimizing the cosine distance between the features extracted from different layers.Finally,experimental results on different image datasets verify that the proposed method can effectively improve the performance of the model with few labeled data.(2)Secondly,a weakly-supervised image classification method based on an anticurriculum learning strategy is proposed to deal with the problem of the lack of annotated medical images.The method utilizes an anti-curriculum learning strategy to select the samples for training.Specifically,at each training stage,we generate pseudo labels for a certain percentage number of samples selected by the prediction of the model for training.In order to further improve the accuracy and stability of the generated pseudo labels,the averaged model predictions of the data with different data augmentations are used as the pseudo labels.In this paper,the pseudo labels are generated by the model of previous training stage.However,in general,the model performance becomes better after training for more epochs before overfitting.Thus,we propose the temporal refinement by linearly combining the pseudo labels with the current model prediction to make the pseudo labels more accurate and further improve the performance of the model.We conducted the experiments on two medical image datasets and the experimental results validate the effectiveness of the proposed method.(3)Thirdly,a teacher-student architecture based on feature space renormalization is proposed for learning with noisy labels.The method simultaneously trains two models named the teacher model and the student model.The teacher model is trained by the semi-supervised learning method and then provides the prediction to generate pseudo labels for the samples.The student model is trained by supervised learning method with the pseudo labels.In order to improve the effectiveness of the semi-supervised learning method,we improve the partition method for the dataset.In addition,to further improve the performance of the student model,the feature space renormalization mechanism based on the theory of group representation is proposed to guide the self-learning of the student model.A better teacher model can provide more accurate predictions,which contributes to the learning of the student model.Meanwhile,a better student model is conducive to the division of the dataset.Therefore,in our proposed architecture,the two models can improve each other’s performance.The experimental results demonstrate that the teacher-student architecture proposed in this paper has better robustness to noisy labels.(4)Fourthly,a weakly-supervised image classification method based on negative learning and feature space renormalization is proposed to further improve the performance of the model on datasets with a high noise rate.Inspired by the effectiveness of the model ensemble,the method trains two models by semi-supervised learning method and supervised learning method respectively,and the ensemble prediction is used as the final prediction.In order to make full use of the available data on the datasets with a high noise rate,we propose a novel sample selection method.We further select reliable samples by the memory of the model and add them into the labeled subset to improve the effectiveness of the semi-supervised learning method.In addition,we combine negative learning and feature space renormalization to train the unlabeled data to further improve the performance of the semi-supervised learning method.For the supervised learning method,we utilize the predictions of the two models to refine the original labels and train the model with refined labels.The experimental results show that the method proposed in this paper can be well adapted to the datasets with a high noise rate.In conclusion,this paper conducts a series of research around the problem of the lack of labels and incorrect labels in real-world datasets,proposes weakly-supervised image classification algorithms applicable to a small number of labels and noisy labels,and verifies that the proposed methods are effective in reducing the need for accurate labels and can improve the robustness of image classification algorithms to noisy labels by the experiments on several different datasets.