| In recent years,deep learning has made great progress in many fields and has attracted widespread attention.However,the success of deep models relies heavily on high-performance computing equipment and a large number of annotated datasets that meet the task requirements.In contrast,humans can quickly learn a new category with only a few pictures.few-shot learning aims to bring machine learning closer to human thinking,with the ability to learn as quickly as humans.Siamese neural network belongs to the category of metric learning,which is an effective method to solve the few-shot problem.This thesis conducts research on the feature representation method and measurement module of the Siamese neural network,and conducts experiments on public data sets.The main research contents are as follows:(1)Aiming at the issue that the feature representation method of image-level for Siamese neural networks ignores the local feature information of samples,leading to sensitivity to target position,intra-class differences,and complex backgrounds,we propose a dual representation Siamese network based on dual representation(DRSiamese Net).DR-Siamese Net introduces local feature descriptors into Siamese neural networks and uses dual features(image-level features and local feature descriptors)to jointly represent images.Compared with image-level feature representation,the use of dual feature representation can capture more abundant local feature information on the basis of considering global image features.In order to fully utilize the local feature and global feature information of the image,DR-Siamese Net measures two different levels of features separately and uses an adaptive learning weight strategy to weight the global feature measurement results and local feature results,thereby obtaining the final result.Experimental results on small sample benchmark datasets,fine-grained image datasets,and crop disease datasets show that DR-Siamese Net can effectively extract the local features of samples and improve the model’s generalization ability.(2)Aiming at the problem that a single metric method for the metric module of Siamese neural network cannot effectively learn more distinguishable features on complex background images and the use of class feature(class prototype)metric may ignore the diversity of samples,resulting in class center deviation,this thesis proposes a multi similarity weighted metric Siamese network based on multi-similarity(MSSiamese Net).MS-Siamese Net designs a weighted metric strategy in the metric module and replaces the class feature metric with an adaptive learning weight that comprehensively considers the metric between the query set samples and the class features,and the metric between the intra-class samples.At the same time,multiple similarity criteria are used to replace the single learnable metric in the Siamese neural network.A weighted metric strategy is used for each of the multiple metric criteria to jointly measure the similarity between the query set samples and the classes in the support set.The experiments on small sample benchmark datasets and ablation experiments show that using multiple metric criteria can comprehensively consider the advantages of different metric methods,learn more distinguishable features,and using weighted metrics can alleviate the problem of class center deviation and improve the classification results. |
