Research Of Group Lasso-Based Semi-Supervised Hashing For Image Retrieval Optimization And Algorithm

With the advent of large data age,the rapid development of Internet technology and the growing popularity of imaging equipment,data collection of image and other media resources is more and more convenient,images are playing crucial role in medicine,astronomy,criminal investigation,transportation,military,environmental,social networking and every walk of life.Traditional technology of image data analysis and processing are facing problem of huge number of media data,high dimensional features,the need for large storage space and retrieval speed is slow.Facing the increasing number of network images,the requirement of large-scale image retrieval algorithm is increasingly urgent.In real life,especially in the field of image retrieval and recognition,most amounts of data are not labeled and more convenient to collect compared with labeled,so semi-supervised image retrieval has great realistic meaning and actual requirement.Image data has the features of color,property and texture itself hence different dimensions of the image data may have some structural or semantic association,so effective modeling of image data structure is critical for image retrieval.In addition,the design of effective solution algorithm is important for image retrieval,especially on large-scale datasets.In this paper,based on the above realistic background,and on the existing semi-supervised hashing algorithm for image retrieval,we do the following works:1.Group Lasso-based semi-supervised hashing algorithm for image retrieval is proposed.The semi-supervised learning method takes full advantage of all the labeled and unlabeled training data.And the hashing image retrieval algorithm only stores the binary hash code of the image,which saves the storage space,takes constant query time,and can be extended to large scale image dataset retrieval.2.Group Lasso is used to consider the group structure into the image retrieval model so that the features of the same group can be synchronously selected into or removed out of the model.Group Lasso also considers the sparseness between groups,which play a role in feature selection,reduce over fitting and improve precision of the model3.Solve and optimize the image retrieval model quickly by introducing the proximal gradient method.4.Test the model on the MNIST and CIFAR10 standard image dataset and compare it with other existing image retrieval algorithms.