Study On Na(?)ve Similarity Discriminator-based Deep Adversarial Metric Learning

Metric learning is a basic technology in computer vision,which plays an important role in many tasks,including image retrieval,object recognition,face recognition,person re-identification and so on.Metric learning aims to learn a specific metric function.The metric function projects original samples into low-dimensional feature representations whose similarities are well preserved.Deep metric learning methods utilize the strong learning ability of deep neural networks and consider these networks as metric functions.As an advanced learning spirit,adversarial learning provides many new ideas for a lot of researches and reaches a good performance.This thesis works on deep metric learning method based on the spirit of adversarial learning,aims to against the problem of preserving similarity of high-level feature representations,and discusses the effects of training difficulty of samples,sampling strategies and the architecture of discriminators on adversarial training.This thesis considers deep neural network as metric function,and designs a discriminator with special architecture named Na(?)ve Similarity Discriminator(NSD)to improve the performance on image retrieval and clustering tasks.The main contributions of this thesis are shown as following:(1)The most important thing in metric learning task is to preserve the similarity of extracted features.Unlike existing methods which utilize a well-designed loss function to illustrate the relationship of samples,this thesis takes advantage of the learning ability of neural networks and designs a special discriminator model called Na(?)ve Similarity Discriminator.NSD is used to predict the probability of being similar of the input pairwise samples.With the combination of standard binary cross-entropy loss,NSD provides large loss when distinguishing hard samples,which encourages generator to generate samples with well-preserved similarity.(2)This thesis designs adversarial sampling strategy,taking the full use of samples with different training difficulty in dataset.Three training stages are conducted in this thesis,including warm-up stage,generator training stage and discriminator training stage.Samples with low training difficulty are utilized to help discriminator learn the ideal distribution,while those hard samples would provide large loss to generator when distinguished by discriminator.Alternative optimizations are employed in the training procedure,generator and discriminator are supposed to be stronger during the confrontation.(3)An Offline Easy Sample Mining(OESM)is proposed in this thesis to build a larger training set for discriminator.In order to avoid the repeated use of samples,a spanning tree is built on the global distance matrix of training set,which provides sufficient samples that benefit the training on discriminator.(4)The experiments are employed on three famous datasets in metric learning task,including CUB-200-2011,Cars196 and SOP dataset.By comparing with classical deep metric learning methods and the state-of-the-art approaches on image retrieval and clustering tasks,the proposed method shows its ability on generating similarity preserved feature representations.Besides,this thesis further explores the influence of the architecture of discriminator with ablation experiments.