Few-shot Image Classification Based On Deep Metric Learning

Deep learning methods have achieved great success in image classification tasks.However,the excellent performance of deep neural network often highly depends on the effective training of the model with a large number of labeled data.When the number of samples is small,the training of the model is limited,which leads to overfitting and affects the results of the classification.In actual application,it is very difficult to obtain labeled samples.Not only that,it's very expensive to label samples.Therefore,it is of great significance that how to efficiently use deep learning model to solve the problem of few shot image classification.The deep metric learning aims to learn an efficient measurement space,completes the nonlinear mapping of the input features effectively and reflects the similarity degree between samples accurately.For extremely special few shot image classification task,deep metric learning still performs well.In this paper,few shot image classification research based on deep metric learning is proposed by us.In order to explore how to make the most of unlabeled data in few shot learning(FSL),we construct a deep metric self-optimizing model(DMSO)based on the metric learning and incremental learning.At the same time,we propose a multi-scale feature and class prototype expression model(MSAP)based on the prototype expression idea,aiming at how to obtain high quality image features and class prototype.The following content is the focus of our research.(1)We construct the deep metric self-optimizing model.The DMSO mainly uses large number of unlabeled samples to solve classification problems of few-shot images.It consists of siamese network classification module(SNC)and confidence self-optimizing module(CSM).Based on the idea of deep metric,we construct a siamese network classifier and use a small number of labeled samples to train the classifier.We use the primary classifier to predict the pseudo-labels of unlabeled samples.In order to solve the noise problem of pseudo-labeled samples,we build confidence self-optimizing module by the incidental parameters and to select pseudo-labeled samples with high reliability.Meanwhile,adding the most trusted pseudo-label to the training set.In this way,the model gets effective selfoptimizing learning and enhance the generalization ability and fitting ability for the model continuously.(2)We propose a multi-scale feature and prototype of class expression model to focus on how to obtain better image features and better representative of the class.Due to multiscale features play a great role in image expression,we build a multi-scale feature fusion module(MSFF)and obtain multi-scale images by gaussian pyramid strategy.At the same time,we used the spatial pyramid pooling method to solve the problem of feature's output for multi-scale image.Finally,we obtain the multi-scale features of the images by attention network.In addition,we build the class prototype expression module(CPE)to learn the higher quality prototypes of all classes.(3)In order to explore the differences of samples within the same class and the different influences between the different samples in the process of obtaining the prototype.Based on the idea of weight,we come up with the concept of weight prototype.Due to the influence of occlusion,background changes and other factors,the contribution of each sample to the class prototype is different.Assigning corresponding weights to different samples to learn excellent the prototype.In this way,we completed the classification task of few shot image accurately.At last,we conducted a lot of experiments for two solutions and achieved excellent results on many FSL datasets such as mini-imagenet and so on.Our method has been fully proved to be feasible and reasonable in few shot image classification task.