Research Of Few-shot Image Classification Based On Metric Learning

With the continuous development of artificial intelligence and machine learning,deep learning algorithms have made a breakthrough progress in the field of computer vision.Few-shot learning based on convolutional neural networks can solve problem of data shortage in practical applications,so become a research hotspot.The purpose of few-shot learning is to quickly learn the information about sample category from a small number of samples or just only single sample,and then classify the new sample according to the category information.In this thesis,we investigate the few-shot classification problem.And we propose classification methods accoding to few-shot images based on metric learning for data sets with different complexity.Few-shot data sets with low complexity: Aiming at its characteristics of simple feature extraction and similar sample features,the few-shot image classification model MPro-Net based on Manhattan distance was proposed on the basis of prototype network.Methods:(1)The MPro-Net model adds an MLP optimization module to the prototypical network to assign weights to the embedding vectors mapped to a higher dimensional space,so as to making the embedding vectors of those samples more robust.(2)Manhattan distance is used to measure the distance between the embedding vector and the prototype representations of each type in the target set to predict the category of the sample.(3)A center loss function is added on the basis of the cross entropy loss function,and the network parameters are updated through back forward to reduce the distance within the sample class,so those embedding vectors can be better distributed around the prototype representations of the class.(4)Select some commonly used handwritten characters of ethnic minorities,and independently build a data set of handwritten characters of ethnic minorities based on bilinear interpolation algorithm.Experiment: MPro-Net was tested in two data sets of Omniglot and Minority Handwritten Characters,and compared with other models.Conclusion: The experimental results show that MPro-Net has higher accuracy and better convergence performance in the tasks of low complexity few-shot image classification.Few-shot data sets with high complexity: Aiming at the characteristics of difficult target feature extraction and large feature differences,MPro-Net was improved to improve the accuracy when dealing with complex image classification.Methods:(1)Deepen the depth of the embedded network,add residual learning units and identity mapping,so that the embedded network can obtain higher-level and richer image semantic features.(2)An image pre-processing module based on the FCM algorithm is added in front of the embedded module to reduce features irrelevant to the sample classification and compress the intra-class features of the sample,so that the improved model can obtain more relevant classification on the pre-processed image samples.(3)The mixed loss function composed of the cross-entropy loss function and the adjustment term is used to increase the inter-class distance of the samples,reduce the intra-class distance of the samples,and update the network parameters in reverse through gradient descent.Experiment: The verification is carried out on MiniImageNet.Conclusion: The experiment results show that the improved model has better generalization ability and higher accuracy in image classification of high-complex data sets.Finally,we take a comprehensive analysis about the experimental data of different classification tasks based on the proposed MPro-net model.And the different impacts of different classification tasks on the results of few-shot classification are obtained.Thus provide a foundation for the study of how to reduce the impact of different difficulty classification tasks on few-shot classification.Based on the improved MPro-Net model,experiments are performed on different complexity data sets and different metric learning.The impact of different complexity data sets and measurement methods on the results of few-shot classification is analyzed.It provides an idea for studying the metric learning methods about different complexity data sets.