Research On Image Retrieval Based Deep Second-order Hashing

Nowadays,with the rapid development of the era,the image and video data on the web grow explosively.Hash technology have been investigated widely for its high storage efficiency and fast query speed of large-scale visual search activities in information retrieval.Due to the excellent performance of convolutional neural networks in computer vision,deep hashing has come into people's field of vision.However,the fundamental image characteristics captured by most deep hashing methods depend heavily on the first-order convolutional feature statistics,failing to take their global structure into consideration.To address this problem,this paper gradually carry out the research of hashing algorithms based on deep second-order feature by covariance estimation,from the perspective of model construction and constraint optimization.It gets rid of the one-sidedness problem of image features,opening up a new direction for hash research.The specific work can be expressed as:(1)A supervised deep second-order covariance hashing is proposed.Based on deep hashing method,the robust covariance is estimated by power normalization,forming an independent structural layer.Then the structural layer is embedded into deep hashing framework in a pointwise way,to explore the second-order statistical information.Under the cross-entropy constraint based on Softmax and Sigmoid,the hash function can be supervised in an end-to-end manner.After that,the global and detailed hash codes can be gained.The model can be evaluated by the index of top-k precision(top-k),precision-recall(PR)and mean average precision(mAP)across four kinds of hash coding bits.Extensive experiments on MNIST,CIFAR-10 and NUS-WIDE show that the performance of the model significantly surpassing other methods only using the first-order conv.feature statistics,which manifests the effectiveness of the proposed approach.It effectively solves the weakly global problem in deep hashing capturing image feature.(2)An architecture of deep covariance estimation hashing is constructed.On the basis of the supervised deep second-order covariance hashing model,pairwise labels are further utilized.Pairs of images are used as input of the network to carry out covariance estimation under dual input streams,performing global pairwise feature interactions.The whole network can be learned by the pairwise similarity constraint in an end-to-end fashion,by which the correlation between features can be explored to generate more robust and powerful deep second-order hash representation.The results on three datasets show that the proposed method is better than the basic network,and is better than the supervised deep second-order covariance hashing.Meanwhile,the comparison results with other representative deep hashing methods again prove that the method is advanced.It is another successful exploration of the deep second-order hashing model.(3)A deep second-order hashing based on category supervision and similarity preserving is proposed.On the basis of architecture construction,a new explorations from the perspective of constrain and optimization is carried out.Given a feature representation generated by global statistics in deep hashing paradigm,the semantic information is added to the dual channel deep second-order hashing framework,establishing a mechanism for common constraints on category supervision and similarity maintenance.In addition,to solve the discrete optimization of hash code,hash-like function is adopted to achieve approximate binarization.In this way,quantization errors can be avoided greatly,while ensuring joint optimization.Compared with the deep covariance estimation hashing model based on pairwise labels,the results on the three datasets exhibit an absolute improvement.The performance of hashing method can reach the level of advanced deep hashing method,which proves that the proposed constraint target and joint optimization method are effective.