Research On Key Techniques Of Fast Image Search

With the onset of the big-data era,the need to search multimedia information has grown rapidly.Nowadays,image data carries important multimedia information and has become a key component of data sources.Rapid and efficient search and retrieval from the massive,unstructured high-dimensional image data with a minimal amount of time and storage space has become an important focus of research in the computer vision and information retrieval fields.This is primarily because of the geometrical increase of the tremendous amount of multimedia data on the Internet and urgent need for large-scale image search capability.This problem not only needs to be solved urgently,but also generates more demanding requirements and challenges for rapid image retrieval technologies.Existing image retrieval studies have focused primarily on topics such as the learning and description of image features,high-level semantic information modeling of images,image similarity search,sorting image search results,and research of distributed image search systems.This paper explores and discusses the methods for image feature extraction and description,high-level semantic information modeling,approximation algorithm similarity search,and the realization of distributed systems with respect to constructing a large-scale imaging system.The main content and innovations of this paper are summarized as follows:1.This paper describes a proposed multi-instance package construction approach based on a semi-supervised learning method.Using a Gaussian mixture model and an agglomerative information bottleneck(AIB)algorithm for joint modeling,the proposed method ingeniously converts the traditional complex multi-instance questions into a simple,single-instance learning problem.In addition,a classifier ensemble algorithm,based on a gradient-boosted decision tree,is adopted to effectively improve the learning outcomes and generalization capability.Compared to other existing multi-instance algorithms,the validation results confirmed that this algorithm can improve the search accuracy and effectiveness.2.This paper also discusses a proposed approximation search algorithm that uses additive homogeneous kernel mapping to achieve an image approximation search based on kernelized locality-sensitive hashing.To address problems with the search accuracy performance instability of an unsupervised image hashing function and declination of search-time performance with an increase of the hashing bits,this paper describes a proposal to combine the additive explicit homogeneous kernel mapping and image feature histograms to construct a search algorithm based on the locality-sensitive hashing function.Moreover,to address the problem of semantic gaps caused by the use of image data lacking type information in semantic modeling,this paper describes an approximation searching algorithm based on the homogeneous kernel mapping of similarities between pairs of images and dissimilar constraint relationships.The image search experiments in this study confirmed that the algorithm can construct a locality-sensitive hash function more accurately,thus effectively improving the similarity search performance.3.A proposed hashing retrieval method based on the deep learning framework to address the problems with poor performance and low efficiency of using existing artificial handmade design features for searches,is discussed.The proposed method takes into account higherorder-dependent information in the neighboring blocks within the same area and combines the hashing index framework without dependence on any artificial design features,or other previous sparse coding algorithm.This method thus becomes more suitable for large-scale image retrieval.Further,a hashing algorithm of supervised images is proposed based on a deep Convolutional Neural Network(CNN).This algorithm can achieve simultaneous learning of specific regional features and a series of binary hashing function set by adding a hidden layer under the framework of in-depth CNN learning.In addition,the algorithm can improve search accuracy owing to a double-layer sorting and filtering learning process through layer search.The validation experiment proved that both algorithm frameworks can achieve higher search accuracy and retrieval efficiency compared to other artificial feature-based learning algorithms.Moreover,the two algorithm frameworks have good versatility and expandability for medium to large data sets.4.This paper describes the design and implementation of a distributed image search system that possesses features such as retrieving large-scale image data sets and sorting the search results.Compared to previous image retrieval systems,the proposed system has great advantages on search efficiency and the throughput of large-scale data processing.