Research On Image Retrieval Algorithms Based On Intelligent Computation

Content Based Image Retrieval(CBIR) starts from the content of images and supplies more effective retrieval methods to automatically retrieve images satisfying users'requirements. CBIR involves with many subjects including image processing, pattern recognition, computational intelligence, database,and so on, and it has been paying attention all the time. Along with the development of computer, multimedia, network, and digital communications, effective image retrieval algorithms have been applied in many fields closely interrelated with people's life, such as image retrieval engine, network environment purification, safe city construction, safe production monitoring, key frame extraction, and many other specialized fields. The research of retrieval algorithms includes the key technologies of feature extraction, similarity matching, relavance feedback(RF), and so on. Along with the development of research and the increasingly complex applications, CBIR desires more intelligent and effective algorithms, so exploring new intelligent methods becomes an important research content in CBIR field. In this thesis, some CBIR algorithms are proposed based on intelligent computational methods of Quotient Space Granular Computing Theory, Particle Swarm Optimization(PSO), and Manifold Learning. The main contents of this thesis include:1. Summarize the key problems of CBIR comprehensively, and typical methods of each part are described in detail, mainly including basic framework of CBIR system, feature extraction methods, similarity matching methods, relavance feedback strategies, and performance evaluation methods.2. Traditional image retrieval algorithms aim at a certain profile of a certain feature, which realize retrieval through single feature extraction and matching. But the content of images is complex, single feature can't represent image content all sidedly. Considering the richness and complementarity of features, some ordinary feature fusion algorithms synthetize several single features together simply. The retrieval effectiveness has received certain increasement due to the embodiment of multi features, but how the features are fused doesn't have credible theoretical foundation, and the information provided by the image doesn't be utilized effectively enough. Aiming at above problems, in this thesis we import Quotient Space Granular Computing Theory into image retrieval field, base on the structure categoricalness of Quotient Space Granular Computing Theory, and imitate the granulating and granularity combination characteristics of people getting to know things. Firstly, we describe the image retrieval problem with Quotient Space, and then make use of different descriptive abilities of different granularities to design optimal guide line function according to combining principles. A granularity combination retrieval method on the attribute function level is proposed, and color and texture features under different granularities are used as examples to verify the validity of the proposed method.3. Aiming at the problems of RF algorithms in CBIR that few training samples, high real-time requirement, asymmetry of training samples, existing large quantity samples not marked, and inflexible re-weighting, PSO algorithm is imported into RF. A PSO relevance feedback algorithm based on image coding(PSO-RF) and a re-weighting relevance feedback algorithm based on PSO(PSORW-RF) are proposed. PSO-RF codes the image features, and proposes effective solutions for the key points of coding, space continuity, constraint condition, evaluation criterion, and terminal condition. PSO-RF algorithm is composed of user feedback process and particle evolution process. When feedback is required, the particle swarm is initialed firstly, then user selects relevant images as feedback results, and uses individual average value as current global best position to supervise the direction of particles evolution. In particle evolution process, the speed and position of particles are adjusted according to fitness, and unduplicated discrete image points in each particle's minimum neighborhood are determined as the discrete solution of RF after current round. PSO-RF algorithm displays the advantage of PSO that effective study and quick convergence fully in RF, overcomes ubiquitous problems, and has received good retrieval effect. PSORW-RF algorithm aims at the inflexible re-weighting problem of RF, combines PSO, code is done to the feature weightings, particles'moving directions are guided macroscopically by the variances of positive and negative user feedback samples'feature components, and the random search characteristic of PSO is also exerted fully at the same time, weightings are adjusted to optimal value among wide particle search space, which achieves good retrieval effect.4. In order to maximize the distinguishing features when reducing dimensionality efficiently, we research Locality Preserving Projections(LPP) method in CBIR, which has combined the advantages of linear and nonlinear dimensionality reduction methods, and a Multiple Weight Matrices Synthesis Based Locality Preserving Projections CBIR Algorithm (MWSLPP) is presented. MWSLPP designs a multi-feature weight matrix synthesized with several single feature weight matrices for LPP, which is more suitable for CBIR. The single feature weight matrices are established and locally adjusted through RF first, and then the multifeature weight matrix is globally adjusted according to positive and negative samples, these operations can simulate a better manifold structure among images for retrieval. MWSLPP includes two different phases: LPP forward application and retrieval phase, and relevance feedback backward spreading phase. In the phase of LPP forward application and retrieval, firstly we extract original features for images, and then construct adjacency graphs and weight matrices. We use the original single features weight matrices to synthesize a global weight matrix, and then do LPP for original features using the synthesized weight matrix, and the initial retrieval result can be returned. In the phase of relevance feedback backward spreading, we locally adjust each LPP weight matrix separately through RF, then globally adjust the multi-feature weight matrix, and finally use the optimized global multifeature weight matrix to accomplish retrieval in a new round. Experiments show that the MWSLPP can optimize the feature structures of similarity to accomplish image retrieval efficiently, and satisfying retrieval effects are achieved.Therefore, in this thesis we have summarized the key technologies and main methods in CBIR field. Deep researches on Quotient Space Granular Computing Theory, PSO, and Manifold Learning have been done, some image retrieval algorithms based on these intelligent computational methods are presented, and we have verified the validity of the algorithms through experiments. The research results of this thesis have enriched the research contents of intelligent computational methods in CBIR field, and have certain theoretical significance and application value in the fields of Quotient Space Granular Computing, PSO, and Manifold Learning. These researches also supply significative methods for intelligent computational methods based CBIR problems.