Image Retrieval Based On Hierarchical Diffusion Process And Manifold Learning

The image retrieval technology based on the diffusion process mainly exploits the intrinsic manifold structure of images to construct the affinity graph,and then diffuses the pairwise affinity values to other surrounding images through the reachable path of graph.These algorithms enhance the relationship between all images of the database,which have been widely used in content-based image retrieval.The image retrieval method using locally constrained diffusion process algorithm suffers from high computational complexity and large amount of iterations,especially when applying to retrieval large-scale image dataset.The regularized diffusion process algorithm also has a disadvantage of computational expensive.In addition,it spreads the affinity values in the low-level visual feature space of images without considering the high-level semantic information.To solve the above problems,this thesis proposes a hierarchical locally-constrained diffusion process algorithm and a regularized hierarchical diffusion process with relevance feedback algorithm for image retrieval by combining manifold learning with algebraic multigrid and introducing relevance feedback.The specific researches are as follows:1.As to the problem that locally constrained diffusion process for image retrieval method has high matrix iterative computation for large-scale images,which will result in high cost of image retrieval time,a image retrieval based on hierarchical locally constrained diffusion process method is proposed.This approach introduces algebraic multigrid to establish the hierarchical structure of the images in image database.Only affinity values among the query image and the top-level images are diffused,and the diffused values is interpolated from the top layer to the bottom layer to obtain the final retrieval results.Compared with the locally constrained diffusion process algorithm,the theory analysis proves that the proposed algorithm can reduce the computation complexity.What's more,the experimental results demonstrate that the proposed algorithm has a higher retrieval accuracy when returns a certain number of top retrieved images than locally constrained diffusion process algorithm,self-smooth operator algorithm and self-diffusion algorithm.2.For the problem that existing diffusion process methods only consider to diffuse the affinity values in the low-level visual feature space of images and ignore the high-level semantic information,which will cause the problem of “semantic gap”,this thesis proposes a regularized hierarchical diffusion process with relevance feedback for image retrieval method.The algorithm establishes a hierarchical structure of the images in image database and spreads the affinity values among the query image and the top-level images by regularization diffusion.Then relevance feedback is introduced to adjust the similarity between the query image and top retrieved images in the top layer,and the affinity values are diffused again according to labeled information from user's feedback.Finally,the similarity between the query image and all images in the database are obtained by interpolating the diffused results on the top layer from top to bottom.The experimental results show that the regularized hierarchical diffusion process with relevance feedback for image retrieval has better retrieval performance than the manifold ranking and regularized diffusion process when returning a certain number of top retrieved images.3.For retrieval of medical CT images,the influence of feature selection on retrieval performance of hierarchical locally constrained diffusion process method and regularized hierarchical diffusion process with relevance feedback for image retrieval method is verified.The thesis uses the GIST,LBP and HOG feature and their serial fused features to conduct comparative experiments in the IRMA dataset.