Research On Unsupervised Image Embedding Learning Methods

Unsupervised image embedding learning aims to learn a feature embedding that has two properties: 1)the embedding features of samples from the same class should be pulled close to each other;2)the embedding features of samples from different classes should be pushed apart from each other as much as possible.Due to the lack of category annotation of images,the main challenge in unsupervised embedding learning is to mine class relationships or weak positive supervision from unlabeled images.To solve this problem,this paper makes a systematic and in-depth research on unsupervised image embedding learning.The main research results are as follows:(1)A joint real-time deep clustering(RDC)and hard samples mining(HSM)method for unsupervised embedding learning is proposed.RDC performs a global clustering operation on the features of all training samples at the initial stage of training,so as to obtain the initialization label and class center of the samples.In order to facilitate the real-time update of pseudo labels and sample categories,we design two dynamic memories: the sample memory storing features and labels of samples and the centroids memory storing class centroids.RDC updates the sample labels by reusing the sample features in forward propagation,so as to avoid additional feature extraction.In this way,in the training process,the sample labels and network parameters are updated at the same time rather than iteratively,which ensures the stability of feature learning.In addition,in order to select more valuable and meaningful samples during training,according to the structural similarity between sample features,we introduce a hard sample mining strategy to accelerate the convergence of the model.Finally,according to the hard samples mined,the structure-level pair-based loss function is minimized,so that the model can learn the high discriminant features of the image.We have conducted a lot of experimental evaluation on the proposed RDC-HSM on three standard datasets,and the image retrieval results show the effectiveness of RDC-HSM.(2)A correlational instance feature embedding(CIFE)method for unsupervised embedding learning is proposed.CIFE exploits the self-correlation and cross-correlation of instances in each training batch by learning a feature embedding with intra-instance variation and inter-instance interpolation.Intra-instance variation synthesizes positive samples to enrich the positive supervision,making the model robust against image noise.Inter-instance interpolation synthesizes mixed samples from two random instances,which is a process of generating semi-positive samples since the samples before mixing are originally negative pairs of each other.Learning through such semi-positive samples is conducive to learning better discriminative representations,resulting in better generalizability on new test classes.Through the image classification and retrieval experiments on four standard datasets,our method is simple and effective,breaks the limitations of instance discrimination methods,and can be applied to the existing instance specificity learning methods.(3)An effective strongly connected subgraph expansion(SCGE)method for unsupervised embedding learning is proposed.SCGE constructs the neighbor graph through the nearest neighbor of sample features to obtain the preliminary sample neighborhood relationship.In order to find more effective or less noisy semantic information between images,SCGE uses the depth first search strategy to mine the strongly connected subgraphs in the feature graph according to the global data manifold formed by all neighborhoods,and the feature points on the strongly connected subgraphs share semantic information.Then,the synthetic feature points are generated by linear interpolation method to expand the strongly connected subgraph.Improve the quality of image feature learning.In addition,we introduce the hard positive enhancement strategy(HPE)to provide a greater gradient for model training.We have carried out a large number of image retrieval experiments on the proposed method,and the results show that our method can effectively mine the semantic information between images and learn high discriminant image features.