Research On Neural Network Structure Of Few-shot Image Classification Based On Metric Learning

Image classification is a basic and critical task in the field of computer vision.Under the condition of large samples,the image classification method based on deep neural network has excellent classification performance,which has exceeded the human recognition level.However,when the sample size is scarce,the traditional large sample method is prone to overfitting,resulting in a sharp decrease in the generalization of the model.Based on this,most of the existing few-shot image classification methods use data augmentation,self-supervision and other methods to make up for the lack of training samples,and introduce transfer learning,meta-learning and other methods to pre-train model or accelerate model optimization,thereby improving model classification performance.However,due to the small amount of labeled sample data,these methods cannot meet the requirements of the model for the representation ability of sample features,which makes the model prone to overfitting problems.Therefore,on the basis of existing research,this paper designs a neural network structure for fewshot image classification based on metric learning to improve the representation ability of sample features and effectively alleviate the over-fitting problem of the model.The main work of this paper is as follows:(1)A multi-branch noise-augmentation based relation network for few-shot image classification is proposed.Existing few-shot image classification methods only use conventional data augmentation strategies(such as geometric transformation and color transformation)to expand the data,without considering the real scene in the presence of noise in real life,and use the max pooling layer in the metric module and most of the information in the feature map is compressed,and it is easy to ignore a lot of information that is potentially valuable for classification.In order to solve the above problems,this paper proposes multi-branch noise augmentation relation network.The network expands the traditional single branch into multiple branches.First,each branch generates different types of noise pollution data,and then learns through the embedding module and the metric module.Among them,the average pooling layer is used in the metric module to ban the maximum pooling layer,focusing on the overall information of the feature map,avoiding ignoring potentially important information from the feature map and finally improve the classification capability of the model.(2)A double feature reconstruction network for few-shot image classification is suggested.Most existing few-shot image classification methods only average the information of different positions in the image and then convert the features of multiple positions into a single vector.Although this method integrates the image information of multiple positions,it does not take into account that the simple averaging operation will make the feature vector vulnerable to the irrelevant background of the image during the conversion process,leading to poor classification ability of the model.In order to solve the above problems,a double feature reconstruction network is proposed in this paper.The network constructs double feature vectors through two modules,in which the primary feature module generates an attention mask based on the image,so that the feature vector can more represent the target area;the secondary feature module interferes with the feature vector to enhance its generalization performance.In conjunction,the network also considers the contextual information of the support classes through an auxiliary loss function to enhance the classification performance of the model.(3)A feature relevant relational network for few-shot image classification is presented.As the existing few-shot image classification only consider the intra-class relation the information from the support and query samples was not fully utilised between the same categories in each task,resulting in limited generalisation performance of the model.In order to solve the above problems,this paper proposes feature relevant relation networks.The network makes best use of the information from the support and query samples to find the most relevant target region for each task,making the learned features more discriminative and recognizable.In addition,a support sample feature traversal loss function is proposed to consider the feature information of samples in different support classes by means of a loss function to enhance the classification performance of the model.