Based On Online Learning Hashing For Image Retrieval Method Research

The continuous development of Internet technology and the portability of mobile communication devices have provided indispensable convenience for building an information-based society.Today,with networked and digital platforms as an important social method,the carrier of information is not limited to text,but more image,audio,video and other information-rich forms are presented.The explosive growth of these multimedia data poses huge challenges for efficient data storage and rapid retrieval of information.How to achieve the effective management of multimedia data,the extraction of core information,and the mining of its internal relations has become one of the key issues of concern in the world today.The nearest neighbor search based on hash technology,with its efficient storage capacity and fast calculation speed,shines in the field of large-scale image retrieval,and is one of the frontier directions of great theoretical significance and practical value.At present,based on hash technology for the image retrieval task is to implement information retrieval in a batch-processing manner on a predetermined stationary dataset.The retrieval method has high computational repetition,time-consuming,and greatly affected by data distribution,and it is often difficult to balance calculation efficiency and time cost.To address these problems,this paper proposes a study of online image retrieval based on hashing technology,which breaks through the constraint of predicting stationary datasets,reduces sample repetition training,and improves computational efficiency and accuracy.The specific research contents and main contributions of this paper are as follows1.Aiming at the phenomenon of "data imbalance" in the learning process of multiple batches of new data streams on nonstationary datasets,a type of label projection online learning model(Label Projection Online Hashing for Balanced Similarity,referred to as LPOH)is proposed.Compared with the existing online hash model,the model realizes the expression of the similarity between the new data stream and the existing data;considering the indication function of label information,the method of engraving the similarity measurement of the same label sample data is introduced to alleviate the influence of the information imbalance caused by the graph-based similarity matrix on the system is analyzed,and the ratio of each sample label is also adjusted;in the optimization process of the model,in view of the discrete constraints of the binary code,an iterative optimization scheme is designed to alternatively learn the binary code.A large number of comparative experiments with multiple benchmark methods show that online hash image retrieval based on tag projection can significantly improve retrieval performance.2.In response to the continuous accumulation of existing data in online image retrieval,the similarity matrix calculation complexity and storage complexity continue to increase,and a class of Scalable Supervised Online Hashing for Image Retrieval(Scalable Supervised Online Hashing for Image Retrieval,referred to as SSOH)is proposed.In order to construct the similarity measurement between the new data stream and the existing data,a new loss function is introduced in the SSOH model;in the optimization process,the pre-calculated intermediate variables are used to replace large-scale pairwise similarity calculations.The calculation cost is effectively saved;in addition,the introduction of intermediate variables also effectively alleviates the impact of data imbalance;in order to reduce the quantification error in the Hamming space,we design a discrete optimization algorithm to make online discrete optimization possible.Experiments on multiple benchmark datasets show that scalable supervised online hashing significantly improves retrieval accuracy and efficiency.3.In view of the difficulty of obtaining sample labels in online learning or the inability to label them at any time,a type of unsupervised Angular Quantization Online Hashing for Image Retrieval(AQOH)is proposed,which integrates unsupervised hash code learning and online Learning and semantic error reconstruction are integrated into a unified framework.In order to mine the information of the sample data provided by the users,the angle embedding method is firstly used to establish the error metric between the binary code and the sample data itself with the aid of cosine similarity.Then,the reconstruction error between the sample space of the new data stream and its corresponding binary code space is further minimized,and the compact binary code of the new data stream is studied with emphasis.In addition,in order to achieve the above operations,two simple and efficient algorithms are designed to optimize the model in a continuous and discrete manner,and the effectiveness of the method is verified in the comparison test with the benchmark method.