Research On Image Retrieval Algorithms Based On Deep Metric Learning

With the rapid development of Internet technology in recent years,the available image data is also growing at a high speed,which has led to the development of computer vision analysis technology and many researchers have started to devote themselves to it.Today,image retrieval has become a hot and difficult area of research in the image field,and developing fast and accurate methods to obtain the desired images from large-scale image databases is the goal of image retrieval.The way to achieve this goal is to accurately extract the highly distinguishable features of various images and measure the similarity between the features,and then get the retrieval results based on the similarity.The emergence of deep convolutional neural networks has made the features of images highly distinguishable,but the intervention of deep metric learning is often required to obtain accurate retrieval results.Deep metric learning research focuses on designing effective loss functions to train an excellent high-dimensional embedding space in which similar images being closer together and dissimilar images being further apart.In this paper,we will study and design effective loss functions for deep metric learning,and thus apply them to image retrieval.The main tasks are as follows:1.Loss functions for deep metric learning are broadly classified into two categories,namely,pair-based loss and proxy-based loss.Pair-based loss tends to provide too much input to the deep convolutional neural network,resulting in high training complexity and slow convergence,but fully exploits the information brought by the samples;the proxy-based loss tends to ignore the rich semantic information between samples,but by replacing all samples of a class with proxy for training,it can effectively reduce the training complexity and speed up the convergence.The Proxy-Anchor loss function has the advantages of both pair-based and proxy-based loss functions,which can exploit the rich semantic information between samples through gradients and has the advantage of low training complexity,but it uses a fixed margin to distinguish between positive and negative samples and does not fully consider the intra-class variance of each class.Therefore,this paper proposes a loss function with a learnable dynamic margin,which is also based on the fusion of pair-based and proxy-based loss,with low training complexity and the ability to make full use of inter-sample information,and solves the disadvantages of the Proxy-Anchor loss function by using a learnable dynamic margin,which can fully consider the intra-class variance of each class.In addition,the semantic measure between proxies is added to the loss function,which not only maintains the inter-class separation,but also participates in learnable dynamic margin training to prevent the overfitting of margin.The experimental results demonstrate that the deep embedding space obtained by using the loss function trained with the learnable dynamic margins is robust and distinguishable,and significantly improves the retrieval accuracy in image retrieval tasks.2.Existing deep metric learning methods learn to obtain embedding spaces with high differentiation by maximizing the differences between classes as much as possible,but these methods ignore the inherent intra-class variances in the learning process and treat all positive samples equally and try to distinguish between positive and negative samples due to the lack of labels,while the ranking of different positive samples is discarded completely.When intraclass variances are ignored,the local structure is unconsciously corrupted and tends to overfit the training set,with low generalization on the test set(classes that do not exist in the training set).In past methods,for positive samples,if it satisfies certain constraints(e.g.,margins)with the anchor point(which can be either a proxy or a sample),it contributes little to the training of the model,but this will waste a lot of information carried by the positive samples.Therefore,to address the above considerations,this paper performs sample generation for positive samples that satisfy the constraints in order to obtain quantifiable intraclass variances from the real positive samples,and designs an intra-class ranking loss function based on the relationship of sample generation using the idea of self-supervision,so that the generated embedding space can be better differentiated for similar samples and their ranking relationships can be maintained in the embedding space,and this loss function can be easily integrated with existing deep metric learning methods.Through experimental demonstrations,it can be found that the sample generation-based intra-class ranking loss function proposed in this paper can achieve superior performance on image retrieval tasks.