Reaserch On Discriminative And Compact Image Feature Representation Methods

Image feature representation is a vital research issue in computer vision and pattern recognition communities,and plays an important role in image recognition and retrieval performance.Most existing shallow image recognition and retrieval models use the extracted features to recognize and search images.However,there still exist some intrinsic limitations among them.Sparse representation based classification(SRC)and nuclear-norm matrix regression(NMR)use each class training samples to reconstruct a testing sample.The error vector,formed from all the reconstruction errors,also contains discriminative information for image feature representation,but currently under-studied.Therefore,the error vector,a high-level representation feature,can also be used to represent an image.Similarly,most learning to hash image retrieval models seldom make full use of the previously learned hashing features.Deep learning,as a multi-step feature learning model,can learn highly discriminative image feature,and achieve impressive classification and retrieval performance.However,deep learning relies on large-scale data and computation resources for numerous hyper-parameters training with complicated back-propagation optimization.Obviously,straightforward training of deep neural networks from scratch on small-scale data can not achieve expected performance.Therefore,in order to develop efficient architecture that are specifically adapted for small-scale data,different from shallow models,a cascaded feature representation based image recognition and retrieval model is proposed.Meanwhile,due to the data sparsity problem,there are not always sufficient instances used to learn effective hashing feature on the domain of interest.To solve such problem,an Optimal Projection Guided Transfer Hashing is proposed,which is inspired by transfer learning.The proposed model can be used for compact image feature representation and retrieval for heterogeneous images.Specifically,the proposed three models are described as follows.(1)Deep cascade model based face recognition(DCM)model.The model is based on SRC and NMR,and uses three-level pyramid structure to compute error vector.The error vectors are cascaded to obtain better image representation.The proposed model inherits the merits of deep learning such as hierarchical learning,nonlinear feature transformation and multi-layer connection.The contributions include four aspects.First,an end-to-end deep cascade model for small-scale data learning without back-propagation is proposed.Second,a multi-level pyramid structure is integrated for local feature representation.Third,for introducing nonlinear feature transformation in layer-wise learning,softmax vector coding of the errors with class discrimination is proposed.Fourth,the existing representation methods can be easily integrated into the proposed DCM framework.(2)Multi-level cascaded hashing(MCH).The model uses supervised methods as basic learning to hash(BLH)model to learn new hashing codes.BLH first learns high-dimensional hashing codes in the former layers.Then,all the learned hashing codes from the preceding levels and the original appearance features are concatenated as the input of the subsequent levels.The compact binary codes in the final level are obtained for retrieval task.The contributions are threefold.First,a hashing-in-hash architecture is designed in MCH for visual retrieval.The excellent traits in deep learning are inherited,such that discriminative binary features benefiting to image retrieval can be effectively captured.Second,in each level the binary features of all preceding levels and appearance feature are fed as inputs of all subsequent levels for higher level hashing.The simple yet effective feature connection way improves the training efficiency of our deep-layered model.Third,a basic learning to hash(BLH)model with label constraint is proposed for hierarchical learning.In fact,existing hashing methods can be easily integrated into our hashing in hash framework.(3)Optimal Projection Guided Transfer Hashing.Recently,learning to hash has been widely studied for image retrieval thanks to the computation and storage efficiency of binary codes.For most existing learning to hash methods,sufficient training images are required and used to learn precise hashing codes.However,in some real-world applications,there are not always sufficient training images in the domain of interest.In addition,some existing supervised approaches need an amount of labeled data,which is an expensive process in terms of time,labor and human expertise.To handle such problems,inspired by transfer learning,this paper proposes a simple yet effective unsupervised hashing method named Optimal Projection Guided Transfer Hashing(GTH),where the images of other distribution different but semantic related domain,i.e.,source domain are borrowed to help learn precise hashing codes for the domain of interest,i.e.,target domain.GTH aims to learn domain invariant hashing projections of source and target domains.Besides,this paper proposes to seek for the maximum likelihood estimation(MLE)solution of the hashing projections of target and source domains due to the domain gap.Furthermore,an alternating optimization method is adopted to obtain the two projections of target and source domains,such that the projection aligment is achieved progressively.Extensive experiments on various benchmark databases verify that the proposed GTH method outperforms many state-of-the-art learning to hash methods.