Research On Few-shot Learning Method For Fine-grained Image Classification

Fine-grained image classification is a key research problem in computer vision,which is much more difficult than generic image classification due to the high cost of collecting and annotating fine-grained image data,and the high similarity among finegrained classes.Although most existing fine-grained image classification methods alleviate small inter-class variance problem to a certain extent,they tend to rely on a large number of data to feed the model,and further result in the biased and inaccurate classifiers in the situation of small sample size.To address the above problems,some methods have proposed few-shot learning methods for fine-grained image classification.These methods are based on the few-shot learning framework while mining the local information of the image,so that the model can lose their attachment to the amount of data as much as possible.However,these methods ignore the influence of the relationship between features on the classification effect,which leads to the classification result is not ideal.Based on the existing few-shot learning framework,combined with the relationship of the features,this paper proposes two few-shot learning methods for fine-grained image classification.On the basis of fully considering the relationship between features,this paper proposes two novel few-shot learning models to solve the problem of few-shot fine-grained image classification.The main work of this paper is as follows:(1)A few-shot fine-grained image classification method based on fusion of global features and local features is proposed.The core of this method is to construct the attention map with the aid of the relationship between global features and local features,and enhance the discrimination of local features with the attention mechanism,so as to improve the classification ability of the model.The model applies the idea of the prototype learning to obtain the global features of each category,and utilizes these features to construct a class-aware attention map,which can make the model pay more attention to key local areas.At the same time,the model introduces a metric learning method based on the relationship between local features to calculate similarity,that improves the accuracy of few-shot learning by expanding the features and introduces local features to make it more suitable for fine-grained image classification.In the meanwhile,the model designs a joint loss function to adjust the influence of global features,so that the model can achieve the optimal state.Finally,the effectiveness of the model is verified by comparative experiments on multiple fine-grained image data sets.At the same time,visualization and ablation experiments are used to verify the influence of each part of the model.(2)A few-shot fine-grained image classification method based on local feature combination is proposed.The core of this method is to design a local feature combination mechanism and a local region attention mechanism to improve the representation ability of features,so as to improve the classification ability of the model.The direct use of local features for classification will introduce noise features,and the scattered features can not well represent the image and category.Therefore,the model introduces the local feature combination mechanism to integrate the local features so that new representative features can describe categories and images more completely.Based on the new regional representative features,this paper proposes a regional attention mechanism to find the importance of different regions,and use this attention mechanism to guide the prediction results.Generally speaking,important areas have a greater impact on the classification results,and vice versa.Finally,the performance of the model is verified by comparative experiments on fine-grained images,and the effectiveness of each part is verified by ablation experiments.