Research On Deep Metric Learning Method Based On Multi-Similarity Loss

With the popularity of digital cameras and smartphones,thousands of images are generated from these devices every day,making accurate image retrieval from large datasets a huge challenge.Deep metric learning aims to obtain a more discriminative embedding space by using deep nonlinear mappings to improve the accuracy of image retrieval tasks.Most deep metric learning methods focus on designing loss functions to improve the utilization of sample similarity.Hard sample mining is an important means to improve the performance of deep metric learning,which essentially evaluates the difficulty of samples accurately by fully utilizing sample similarity.Therefore,introducing multiple similarities into the loss function to mine hard samples has high research value.However,existing deep metric learning methods often use a single similarity for hard sample mining in the loss function,which is not conducive to accurate mining and full utilization of hard samples.To solve this problem,this paper proposes two deep metric learning methods based on multisimilarity loss.The main content is as follows:(1)In order to accurately mine hard samples,the paper introduces proxy similarity in hard sample mining and proposes a new deep metric learning method called Deep Metric Learning Based on Proxy Similarity.Based on this,a new loss function,Proxy Contrastive Loss,containing multiple similarities is designed.This method focuses on the relationship between special samples and samples,which is very similar to the relationship between samples and proxies by introducing proxy similarity,thereby improving the aggregation degree of samples in the embedding space.Compared with other related algorithms,this method obtains relatively good experimental results on three benchmark data sets of image retrieval.(2)In order to make loss function fully utilize the generated hard samples,the paper introduces the feature storage structure into the sample-generated hard sample mining method and proposes a new deep metric learning method called Deep Adversarial Metric Learning Based on Feature Memory Bank.This method stores the generated hard samples in the feature storage structure,enabling the current loss function to utilize the positive sample similarity of generated hard samples,thus making them cluster around their positive samples.In addition,the paper uses Label-Preserving Loss to control the difficulty of generating samples,effectively suppressing the evolution of generated hard samples towards noise.This paper verifies the proposed algorithm on several benchmark data sets for image retrieval.Compared with other similar methods,the proposed algorithm has significant performance improvement.