Study On Relationship-aware Hard Sample Generation Based Deep Metric Learning

Metric learning mainly studies the accurate measurement method of distance relationship between data samples.Through the powerful function fitting capabilities of neural networks,the network model is trained and used as a metric function to accurately measure the similarity relationship between samples.At present,as an important research content in the field of machine vision,depth metric learning plays a key role in image retrieval,fine-grained image classification and face recognition.In recent years,a series of hard sample synthesis methods have been proposed,which can generate hard samples that are difficult to distinguish from the original data,further improve the training efficiency and performance of the metric model,and have achieved good performance in the research of deep metric learning.This thesis mainly works on the deep metric learning based on the synthesis of hard samples,analyzes the three research difficulties in the existing methods,and proposes corresponding solutions.The innovative and main contents of this thesis are as follows:(1)Aiming at the problem of insufficient consideration of the global distribution relationship of samples in the process of sample synthesis,we use the graph structure to learn the global relationship between sample data.Different from many existing methods using random sampling strategy,this thesis constructs a minimum spanning tree structure sampler to constrain the sample synthesis process,so that the synthetic samples conform to the original data distribution,which is more conducive to the training of the model.(2)Aiming at the problem that the optimization information provided by the synthetic samples is not clear,this thesis takes the data class center as the reference point of sample optimization direction,and constructs the angle loss constraint in the process of hard sample generation.In this way,a clearer synthetic sample generation interval is constructed in the metric space,thereby encourage the original sample to move toward its cluster center in the optimization process,improving the overall performance of the model.(3)Aiming at the problem of insufficient utilization of synthetic samples during model training,this thesis adds adaptive dynamic weight to the synthetic sample loss according to convergence of the model in each iteration of training,so that the synthetic sample can automatically adjust its contribution in the total loss according to the training results,and further enhance the promotion effect of synthetic samples on model training.Combining the above three aspects,this thesis proposes a deep metric learning method RHSGML(Relationship-Aware Hard Sample Generation in Deep Metric Learning).Through the experimental comparison with 7 excellent deep metric learning methods such as Triplet Loss,Angular Loss,N-pair Loss and HDML on Cars196,CUB-200-2011 and SOP three public datasets,and multi-angle self-comparative experimental evaluation,the effectiveness and superiority of RHSGML are verified.