Research On The Application Of Manifold Metric And Causal Inference In Few-Shot Image Classification

Few-shot image classification,as a fundamental task in computer vision,requires computers to predict and classify test images with a small number of trained image samples.In this thesis,we focus on three types of few-shot image classification tasks from the perspective of data patterns in training and testing sets: supervised few-shot learning,unsupervised few-shot learning,and supervised few-shot domain adaptation.In supervised and unsupervised few-shot image classification,the training set and test set categories do not intersect,with a focus on learning to learn; In supervised few shot domain adaptation,the training set and test set have the same category but different domains,with a focus on domain adaptation under few-shot conditions.Manifold metric and causal inference are two parallel methods.In this thesis,by deeply exploring these two technologies,three kinds of few-shot image classification tasks are solved,and the classification accuracy and application value of the corresponding algorithms are improved.Firstly,in supervised few-shot image classification tasks,we use manifold metric and propose an oblique manifold structure suitable for few-shot scenarios and a solution for manifold metric to address the challenges of the representation ability,classification modeling,and optimization process of oblique manifold structure.In addition,in unsupervised few-shot image classification tasks,we use causal inference to analyze,and propose an unsupervised metalearning framework.Finally,for supervised few-shot domain adaptation tasks,we introduce ideas such as hypothesis adaptation and causal inference,and propose a front-door adjustment scheme.The main innovative work of this thesis includes:1.we propose a transductive few-shot image classification algorithm on oblique manifold.Firstly,we introduce an oblique manifold suitable for few-shot scenarios,which can consider the geometric structure of the manifold when measuring to solve the issue of distance measurement; Secondly,to solve the problem of feature modeling in manifold geometry,we propose a non-parametric region self-attention and spatial pyramid module to enhance the discriminative and generalization abilities of oblique manifold.At the same time,we use the characteristics and advantages of transductive learning to take the entire test set as the input of the model,improving the model’s expressive ability; Finally,we propose a loss function to optimize the manifold space using tangent space as the medium to solve the manifold space optimization.Evaluation experiments on publicly available benchmark datasets demonstrate the effectiveness of the algorithm.2.We propose a causal intervention unsupervised meta-learning algorithm.Firstly,to address the issue of context bias,we abstract the unsupervised few-shot image classification into a causal structure model,analyze the reasons for context bias,and propose a novel adjustment formula combining Variational Auto Encoders; Secondly,to solve the prior factor independence problem,we propose a prior factor conditional independence hypothesis,and use a directed acyclic graph to learn the conditional independence relationship; Finally,we propose Causal Gaussian Mixture models,unsupervised and supervised causal Expectation Maximum algorithms for meta-training and meta-testing to solve the problem of dual-optimization problem.Experiments on publicly available benchmark datasets demonstrate the effectiveness of the algorithm.It enhances the interpretability of the model by generating counterfactual images through intervention operations.3.We propose a few-shot hypothesis adaptation algorithm based on causal intervention.First,to address the issues of data transmission and privacy leakage when accessing source domain data,we introduce the hypothesis adaptation to eliminate the access to source domain when training in the target domain; Secondly,a causal structure model is constructed to analyze the impact of domain-specific information,and a front-door adjustment strategy including feature modeling and hybrid classification modules is proposed to eliminate the negative effects of domain related information.Finally,we solve an optimization method for feature mixing,which makes the model converge quickly and avoids overfitting.The effectiveness of the algorithm is demonstrated through domain adaptation task experiments on available benchmark datasets.