Research On Deep Metric Learning Algorithm And Its Application In Object Retrieval

Learning distance metrics is an important and fundamental topic in the computer vision community.Recently,due to the rapid development of deep neural networks,deep metric learning has been widely studied.The method of deep metric learning is to constrain the deep neural network to learn an embedding that satisfies the distance relationship in the embedded space.Although the current deep metric learning method has made great progress in comparison with traditional metric learning,it still faces some challenges.As far as Contrastive Loss and Triplet Loss,the two basic loss function forms of deep metric learning,most of the research is based on Triplet Loss with relative distance constraint,Contrastive Loss with absolute distance constraint has received relatively little research attention.Although the original Contrastive Loss has many training problems such as difficult tuning hyper-parameters and easy over-fitting,it still has significantly low time/space complexity and good performance potential.Therefore,how to overcome its shortcomings is still a valuable topic.On the other hand,the margin hyper-parameter is an important factor affecting the performance of deep metric learning algorithms.How to design a deep metric learning algorithm without margin is also a major challenge for deep metric learning.In addition,how to design a new loss function form that is different from Contrastive Loss and Triplet Loss is also an important topic for deep metric learning.We explored the above topics of deep metric learning.Our works is summarized as follows:(1)We propose an optimal transport based weighted method for Contrastive Loss.The loss matrix of the Contrastive Lossis used as the transport cost matrix in the transport problem and assigned 0/1 weights for every pair by solving the maximum optimal transport problem.In other words,we propose a hard sample pairs mining strategy for Contrastive Loss.This strategy ensures that each sample in a mini-batch is only included by one pair for the optimization and avoids possible gradient direction conflicts.It also ensures that the mined sample pairs are the largest loss value combination.Our method can obtain better retrieval and clustering performance on the CUB-200-2011,Cars196,and In-shop Clothes Retrieval datasets than compared baseline methods.(2)We propose a structured deep metric learning algorithm without margin hyperparameter based on retrieval sequence.Our method uses all the positive and negative samples with incorrect relative position relationships(discordant pairs)in the retrieval sequence to construct triplets with the query sample to calculate the loss.Through the structure of the correct retrieval sequence,each discordant negative sample is pulled to the back of all positive samples in the retrieval sequence.Therefore,the margin hyper-parameter to ensure the discrimination between positive and negative pairs in the general triple loss is eliminated.We also propose a weighting strategy based on the rank of negative sample subsequences.The more negative samples close to the front of the sequence,the more penalized.Our method can obtain the state-of-the-art retrieval and clustering performance on the CUB-200-2011,Cars196,and In-shop Clothes Retrieval datasets.(3)We propose a novel proxy-based Histogram Loss which using the similarity between samples and proxies instead of the similarity between samples.Through the idea of proxy-based method,Histogram Loss without margin parameters is improved,and the instability of model collapse caused by training on fine-grained classification datasets is solved.Proxy-based Histogram Loss can obtain better retrieval and clustering performance on the CUB-200-2011,Cars196,and In-shop Clothes Retrieval datasets than Histogram Loss.We also explore the proxy-based Contrastive Loss,and its disadvantages such as overfitting are discussed.We extend the application scope of the proxy-based method on deep metric learning algorithms.(4)Based on the structured loss function proposed in(2),we design and implement an object retrieval system.The system is developed using the.NET framework and Python,with a separate front-end and back-end design.The system will return the Top-10 most similar sample from the selected dataset as the retrieval result based on the images submitted by the user,and give the confidence score of the Top-1 result at the same time.