Research On Self-Supervised Few-Shot Image Classification With Uncertainty Information

Images are an important form of information preservation and transmission for humans.Image recognition models based on deep learning have made significant progress and played important roles in fields such as autonomous driving,intelligent agriculture,and surveillance.However,machine learning models that based on deep neural networks typically require a large amount of accurately labeled data for supervised training,which is often difficult to meet in practical applications.In scenarios with insufficient training samples,supervised image recognition models may exhibit slow convergence,poor accuracy,and weak generalization ability.Therefore,the high cost of collecting and labeling large amounts of data in specific fields has limited the further application of deep image recognition models in academia and industry.Few-shot learning aims to learn information about relevant object categories from a small number of samples and is considered an effective approach for image recognition in data-limited scenarios.The major challenge facing few-shot learning is that the accuracy of current few-shot image classification models is still low,leaving significant room for improvement.To further improve the classification ability of few-shot image models,this thesis introduces the theories of self-supervised learning and uncertainty modeling to study few-shot image classification models.The main research work is summarized as follows:(1)To address the problem of insufficient feature extraction in current metric-learning-based fewshot classification models,a self-supervised few-shot image classification model that based on feature fusion methods is proposed.This model uses the Mixup technique to fuse features at the sample feature level,generating augmented features with latent category information,which alleviates the problem of insufficient training samples for supervised tasks.Additionally,this model incorporates self-supervised learning techniques by constructing consistency-oriented and diversity-oriented selfsupervised pretext tasks.By leveraging the highly complementary properties of contrastive selfsupervised pretext tasks and few-shot learning tasks,the model proposes a framework that combines self-supervised learning and few-shot learning.This approach enables the feature extraction network to learn rich semantic features,leading to improved classification performance in few-shot image classification models.Finally,extensive experiments on commonly used few-shot learning datasets were conducted,and the proposed method was compared with multiple classical methods,demonstrating superior performance over other methods.(2)Furthermore,considering the inherent data uncertainty in training samples and generated samples in few-shot image classification tasks,a self-supervised few-shot image classification method that incorporates uncertainty information is proposed based on the previously mentioned model.This method embeds uncertainty evaluation and distance measurement with uncertainty information into the self-supervised few-shot image classification pipeline and trains them simultaneously under a unified framework.Addtionally,the approach realizes the selection of optimal few-shot training tasks and generated features,thereby improving the classification performance of the proposed model.Finally,experiments on commonly used public datasets validate the effectiveness and superiority of this method.(3)This thesis proposes and implements a prototype system for few-shot image retrieval to address the challenge of difficult data annotation for specific scenarios such as rare species identification.The system uses two self-supervised few-shot image classification models proposed in this thesis to achieve image classification and retrieval functions in data-limited scenarios,providing support for practical image retrieval needs.