| With the advent of big data,image classification has provided many convenient changes to people’s lives,and the deep learning algorithms used for image classification have been in full swing recently.However,in some cases,the traditional large-scale deep learning image classification algorithms are no longer applicable because of the large amount of training data required for some data that requires a large amount of effort or is inherently small.This leads to the concept of Few-shot,and image classification for Few-shot tasks is one of the important branches.Traditional image classification mainly uses the residual network Res Net as the backbone network to extract features,and due to the different public datasets,there may be cases where the same class may be misclassified due to large differences in appearance or background,or different classes may be misclassified as the same class by the model in similar situations.In the previous work,better feature extraction for these cases could not be achieved due to the limitations of the network.The accuracy and efficiency of few-shot image classification can be significantly improved by improving the performance of backbone network and strengthening the attention to and extraction ability of key parts of data.Optimizing the backbone network is a good solution to the above-mentioned problems in the traditional approach,which can improve the characterization ability from the basic feature extraction stage,so as to better solve the problems of large background differences and intra-class differences in images,and greatly improve the overall effect of the network.Therefore,this paper proposes a novel approach that optimizes the backbone network and integrates an attention module,to investigate the Few-shot image classification problem,and the main work of this paper is as follows:(1)The few-shot image classification algorithm based on the D-Res Ne Xt backbone network model is proposed.Aiming at the problem that most of the existing few-shot image classification methods use ordinary residual networks to extract features without paying attention to the limited capability of the backbone network itself,and considering the ability to improve the performance of the backbone network by Res Ne Xt,an upgraded version of Res Net that is better used in the field of image classification,this paper combines the few-shot task characteristics,and proposes an optimized network applicable to few-shot image classification.Improving the overall feature extraction capability of the network,while combining self-attention and cross-correlation,makes the final obtained image features further focus on what is helpful for classification.The proposed method is evaluated through multiple comparative experiments on three publicly available datasets,showcasing the overall improvement in model performance.(2)The few-shot image classification algorithm based on optimized channel attention backbone network model is proposed.In order to extract more information about similar or different features between different images,and to address the problem that the backbone network lacks effective interaction between channels,features cannot be communicated between channels,and inaccurate recognition of images that are blurred or have backgrounds with large intra-class differences,the Res Ne St network is improved by combining previous work and proposing an F-Res Ne St network model suitable for Few-shot tasks.The model first introduces split attention into the backbone network,exploiting its ability to obtain interaction information between features and its utility in different feature representations,and makes optimization changes based on the previous model to improve the network extraction ability again,while combining with the FCA(Few-shot channel attention)module to do some appropriate cross-channel interactions without changing the channel dimensionality to achieve better network performance without dimensionality reduction.This method is compared with the four kinds of public data sets,and the improved model performance is improved,and the effect of this model on the classification task of few-shot image is verified. |
