Few-Shot Image Recognition Based On Metric Learning

Few-shot image recognition based on metric learning is dedicated to training a few-shot learning model that can learn more discriminative embedding space and metric space with lots of similar meta-tasks,and it can quickly generalize to new tasks that only contain a small number of labeled samples.Compared with the methods based on memory storage and parameter initialization which have complex network structure and high computation cost,the metric learning method not only has the advantages of simple structure,low computation cost,and fast training but also has high recognition performance.Inspired by it,we make further research on few-shot image recognition based on metric learning,and propose a variety of improved methods.The main research content of this paper can be summarized as follows:Firstly,to improve the performance of few-shot fine-grained image recognition,we propose a coupled patch similarity network.The proposed network can not only capture the discriminative features through the feature enhancement module,patch similarity module,and patch weight generator,but also comprehensively measure the similarity of sample pairs through the coupled metric network.It can greatly reduce the metric bias that caused by single metric branching and effectively solves the problems existing in the few-shot fine-grained image recognition task,so as to greatly improves the recognition performance of the model.Secondly,to narrow the deviation between the category prototype and the real cluster center conforming to the sample category,we propose a prototype optimization network,The proposed network utilizes the confidence score to soft-weight the query samples to the support samples and optimize the initial category prototype in continuous iterations.In addition,the network also introduces an adaptive recognizer module and two strategies of rotation and label smoothing,which make the model can not only mine the internal connections between samples,but also learn more perspective features.What’s more,it also reduces the calibration error of the model and improves the confidence score of the model prediction.