Few-Shot Image Classification Via Feature Optimization Learning

In recent years,with the development and progress of computer vision,image recognition algorithms based on neural networks have achieved astonishing results in accuracy,which is inseparable from the training of large-scale datasets.However,in some special fields such as military,medical,and aerospace,it is not possible to provide a large number of datasets to support neural network training.When the data volume is far less than the training requirements,the knowledge learned by the model is insufficient to solve the current task,often resulting in serious overfitting and lack of generalization ability.The emergence of few-shot learning aims to solve the above problems,hoping that the model can utilize a small number of labeled samples for learning to complete the corresponding task of classification or recognition.Due to the scarcity of samples,current few-shot learning methods have borrowed the idea of transfer learning,whose core is to use an auxiliary set that does not intersect with the target task to learn transferable knowledge or capabilities,and use the learned prior knowledge to help solve new unknown tasks.This article adopts a few-shot learning method based on fine-tuning to alleviate the problem of sample scarcity.How to efficiently construct powerful prior knowledge is the focus of this article’s research.Compared to the learning ability of building models through meta learning or situational learning,an embedded model learning method with strong feature representation can also provide intuitive and effective assistance for the generalization ability of small samples.Moreover,in the face of different classification settings for few-shot problems,there is no need to retrain the entire network model.The main research work of the paper is as follows:(1)In order to improve the feature representation and generalization ability of the model,a dual spatial constraints-based few-shot image classification method is proposed.On the basis of fine-tuning,feature extraction is carried out for auxiliary data by means of full classification pre training.In order to avoid the phenomenon of feature collapse,self supervised learning tasks are added at the same time,and the discriminability of individual features is enhanced by learning the semantic structure information of samples with an auxiliary task with rotation prediction.In addition,in order to enhance the robustness of the feature extractor,distance constraints are applied to the relatively loose high-dimensional feature space using sample label information,aiming to improve the model’s ability to judge unknown samples.(2)To further enhance the generalization performance of the model,few-shot image classification method based on clustering optimization is proposed on the basis of(1).During the training process,a memory block is added to store various types of feature information,which is used to compensate for the lack of reference information during distance constraints.This helps to form strong anchor points in high-dimensional feature spaces,effectively increasing feature differences between different categories,and optimizing clustering results for each category.At the same time,in order to further improve the classification performance of small samples,a K-means clustering method based on dynamic weight adjustment is proposed,which adjusts the initialization and update strategies of traditional clustering and improves the fine-tuning effect of classifiers in a direct push environment.