Research Of Few-Shot Image Classification Based On Deep Representation Learning

In recent years,with the popularization and rapid development of Internet technology,humans can obtain a large amount of labeled data through the Internet easily.As a result,big data and artificial intelligence technologies have made great progress in many fields,such as image recognition and speech recognition.However,most of the current image recognition and classification algorithms are based on the deep neural networks and rely heavily on a large amount of labeled data.When there are few labeled samples,these algorithms often fail to perform well as expected.Therefore,research based on small data sets,i.e.few-shot learning,is also very important.In order to solve this problem,this paper studies the semi-supervised few-shot representation learning based on the image classification task,including the adversarial variational autoencoder and the mutual information maximization neural networks.This paper first proposes an algorithm for learning feature extraction based on an adversarial variational autoencoder.This algorithm extracts the feature information of the image samples through the encoder,and represents them with high-dimensional vectors,then restores the samples through the decoder,so as to learn the hidden feature representation of each sample.At the same time,by introducing the adversarial training mechanism,it can not only work as data augmentation operation,but also improve the model's learning ability of the discriminative features of each sample.Finally,through the distance-based metric learning method,the classification task is done by calculating the most similar labeled sample for each test sample and mark them as the same class.Considering in the case of very limited data in the previous algorithm,the model would suffer from overfitting,so this paper further proposes an algorithm based on transductive inference and mutual information maximization to learn the feature representation of each sample.This algorithm first extracts the local feature maps and global feature vectors of all images through a convolutional neural network,and the model is trained to maximize the mutual information between the global feature vectors and local feature maps of these samples under the unsupervised manner.At the same time,a clustering loss function is introduced to cluster the sample features from the same category.Finally,the similarities between samples are calculated by a distance-based metric,and each unlabeled sample is classified as the most similar labeled sample.The main contributions of these algorithms described in this paper are as follows:(1)The first algorithm bases on the adversarial autoencoder network,it can achieve semi-supervised learning strategy with a small number of labeled samples,and strengthen the encoder's capabilities of extracting discriminative feature and representation for each sample;(2)The second algorithm introduces the tansductive inference method to maximize the mutual information of the local feature maps and global feature vectors of all samples,which maps the samples to the feature vector space with more robust discriminative boundaries;(3)By designing and utilizing of clustering loss function with hard margin,the model is able to clustering the feature vectors of the samples with the same label,which can effectively improve the classification result of the distance-based metric learning method.This paper has conducted experiments to verify the two algorithms proposed on the commonly used few-shot image classification datasets,and the experimental results confirmed the feasibility and effectiveness of the algorithms proposed in this paper.