Research On Domain Generalization Classification Tasks Based On Meta-learning

Recently,deep learning has achieved remarkable results on classification tasks.The classic deep learning model assumes that the training and test data follow the independent and identically distributed rule.However,it is difficult for real application tasks to achieve this assumption and result in a decrease in model performance.In order to solve the problems of the scarcity of labeled samples and insufficient model generalization ability,domain generalization was proposed and received extensive attention.Among them,the domain generalization method based on meta-learning improves the generalization ability of the model by simulating domain differences,and has achieved good performance.However,the current domain generalization model based on meta-learning is limited to traditional episodic training methods,and the metatraining tasks are limited,which affects the improvement effect of meta-learning on classification tasks.Moreover,the existing feature optimization methods based on semantic alignment fail to fully mine the label information of all source domain samples,which limits the ability of the model to extract domain-invariant but class-dependent features.Focusing on the classification task of domain generalization scenarios,this paper proposes a novel domain generalization classification model based on meta-learning to solve the problems of limited meta-task set and insufficient constraint of feature function in the existing meta-learning generalization method.In order to improve the generalization performance of the model,the mixed-episodic training method was adopted to enrich the meta-task set,and the classification network based on learnable cross-domain generalization centroid-based classification network,feature extractor independent parameter updating method and semantic loss were used to optimize the feature distribution.Experimental results show that the proposed method has a stable performance improvement on natural image dataset,histopathological image dataset and switchgear PD audio detection red phase audio dataset.The main research contents and innovations of this paper are as follows:Firstly,a simple and novel mixed episodic training method is proposed,which exposes the classification model to more diverse generalization scenarios,so that the model can be exposed to more varied generalization task knowledge in the training process;Second,introduce learnable cross-domain generalization centroids for classification networks,and design a meta-learning parameter update method that is independent of feature extractors based on cross-domain generalization centroids,and give full play to the guiding role of generalization centroids;Thirdly,the semantic alignment loss function based on cross-domain generalization centroid is introduced to normalize the feature space with domain-independent and domain-dependent semantic alignment to achieve fine-grained feature alignment and improve the generalization ability of the model.