Research On Small Sample Image Classification Problem Based On Distance Metri

With the rapid development of deep learning and the mass popularization of highperformance GPUs,deep learning models have comprehensively recognized enemies and even surpassed humans in image classification.Due to lack of a lot of annotation data,the deep model of the network is easy not to work in many constrained scenarios,and its performance is not satisfactory.How to produce a classifier with better generalization performance under a small amount of data has become a hot issue in the field of artificial intelligence.Research on few-shot learning provides a possible solution to this type of problem.Among existing methods,metric learning is an important branch of few-shot learning.Specifically,the model based on metric learning mainly includes a feature extraction module and a similarity measurement module.In this paper,some researches are carried out on the measurement module,the feature extraction module and the loss function,etc.,to try to improve the generalization ability of the model.Aiming at the problem that the current metric-based small sample classification calculation only focuses on the calculation of the first-order statistic relationship in the feature extraction module,we try to use the second-order statistics to characterize the data features.Aiming at the problem that the measurement function in the similarity measurement module is too single,this paper proposes a covariance matrix measurement method.This method uses the second-order statistics information at different scales to update the model parameters,and combines the similarity measurement method of the covariance matrix to make full use of the relationship between the second-order statistics between samples.The experimental results show that the performance of this method is better than The traditional algorithm has a certain improvement.In view of the relatively simple assumptions of the category prototype in the prototype network,we try to use the characteristics of the different scales between the multi-layer networks to construct the category prototype,trying to build a richer information,closer to the true distribution Prototype hypothesis of the data.At the same time,the loss function of many current metric learning methods simply uses the prediction results of the query set to perform a loss accumulation.We try to add a new loss function of the similarity between classes,so that the metric module of the model can learn better District category information.Make the entire model more closely map the characteristics of similar data and measure the similarity,and make the feature mapping and similarity measure of heterogeneous data more distant,so as to achieve the task of better generalization of few-shot learning.The experimental results prove that our multi-layer prototype representation and the addition of the loss function between classes have positive significance.Aiming at the problem that the previous few-shot learning algorithms always focus on using the global features of samples or learning global regions,but do not pay attention to the potential information differences of local images or regions,we propose a module based on local attention mechanism to try to Focus on local sensitive areas in the data.At the same time,for the problem that the feature extraction module is relatively simple,which leads to the problem that the extracted features are not robust enough,we try to add a multi-head attention mechanism to assist the feature extraction module for training.Finally,we utilize the joint loss of the two modules to assist the overall model training.Experimental results show that our method alleviates these two problems to a certain extent.