Research On Image Retrieval Based Deep Hashing

With the rapid development of mobile Internet technology,image,text,video and other data show exponential growth.In large-scale image data,how to retrieve user demand information quickly and effectively has become a major issue in the field of image retrieval.The Approximate Nearest Neighbor?ANN?search technique is closely related to the rapid search technology of massive images,and the hashing method has received extensive attention.In recent years,the deep hashing method combined with deep learning,which trains the deep network to learn the hash function and converts the image pixel information into binary coding,has achieved significant performance improvement compared with the traditional hashing method,and has become a research hotspot in image retrieval.On the basis of existing hashing methods,the thesis further studies image retrieval based on deep hashing,and proposes three deep hashing learning methods.?1?A supervised hashing method for convolution column sampling based on deep feature is proposed.In the research of traditional supervised hashing method,column sampling discrete supervised hashing method uses hand-designed feature to learn binary coding of image.Since the deep feature of the image extracted by the convolutional neural network is superior to the manual feature,the deep feature of the image is employed to column sampling discrete supervised hash coding learning.In retrieval,the proposed method has the best performance compared with the traditional hash coding method,and it has achieved better recognition accuracy than the classical method in palmprint recognition application.?2?A second-order pooling deep supervised hashing method is proposed.In existing deep hash networks,convolution layer features are used to represent images,while second-order statistics in deep networks show good performance in image classification.Based on the DeepO₂P model,the second-order statistical representation of O₂P is embedded in the deep hashing network,and the second-order statistical deep hashing network architecture is constructed to learn the second-order statistical information and hash coding simultaneously.The experimental results on commonly used hash learning datasets show that the proposed second-order deep hashing method has better retrieval performance than the convolutional feature deep hashing method.?3?A high-order statistical deep hashing method based on multi-layer convolution feature aggregation is proposed.Second-order information is obtained by cross-correlation of cross-layer convolution features and autocorrelation aggregation of same-layer convolution.In hash coding and end-to-end optimization of second-order statistical information,pointwise learning and pairwise learning are employed.The proposed high-order statistical deep hashing method is evaluated on CIFAR-10 and MNIST datasets.The experimental results show that the method of high-order statistical deep hashing is more effective than that of using only first-order statistical information,and also better than several supervised deep hashing methods.