Research On Hard Sample Mining And Generation Algorithms For Deep Metric Learning

Deep metric learning,which combines deep learning and metric learning,has made great progress in tasks involving similarity matching,such as image retrieval,face recognition,and person re-identification,as a result of the development of deep learning.However,how to make full use of training data and construct hard sample pairs for effective training is crucial to the performance of training results in deep metric learning.In order to solve the above problems,this paper conducts research on the generation and mining of hard samples:First,existing sample generation methods for deep metric learning only focus on the generation of hard negative pairs,and the hardness of generated samples cannot be accurately controlled.Therefore,I propose a two-stage hardness-adaptive sample generation(THSG)framework.The first stage dynamically adjusts the hardness of the generated anchorpositive pairs through the piecewise linear manipulation module,and generates realistic synthetic samples through the conditional generative adversarial network;the second stage uses the reverse triplet loss and the generative network to further synthesize complete hard triplets.The network is finally trained with these hard triplets,and then learns stronger feature discrimination ability.Then,in order to fully exploit the new samples generated by THSG,THSG is further extended and combined with the existing hard mining strategy for hard sample mining and generation.The scalability of the proposed sample generation algorithm is also demonstrated.Finally,I explore methods of deep metric learning in unsupervised scenarios and find that some unsupervised deep metric learning algorithms lack mining hard sample features in dynamic dictionaries and use all instance features indiscriminately when calculating losses,a problem that has a negative impact on model performance.In this paper,I innovatively introduce hard sample mining into a memory dictionary-based contrastive learning framework,propose a difficult instance contrastive loss,and further combine the cluster contrastive loss to design a hybrid contrastive framework guided by difficult samples,which effectively improves the feature discrimination capability of the model.Experiments on several publicly available datasets of person re-identification,the results also demonstrate the effectiveness of the proposed method.