Research On Theory And Key Technology Of Image Retrieval Based On Content

With the development of technology and living standards of human being,the use of computer is quite prevalent around the world.In the meantime,numerous appliances with function of taking digital images,such as digital camera and smart phone,have been widely applied in our daily lives.In such circumstances,a large quantity of digital images are produced every single day.Furthermore,people prefer to communicate through images rather than traditional text messages due to the changing habits and the acceleration in pace of life.Additionally,digital images are tightly related to,and even a significant role in the field of vocational work.Therefore,some novel objectives have been set for image data management in order to retrieve vital information of concerns and to discard information redundancy rapidly and accurately while dealing with massive image data.Given this urgent demand,content-based Image Retrieval technology(CBIR)came into being and quickly became a popular topic in the field of multimedia information processing technology.This paper focus on the key elements of retrieval technology and the improvement of retrieval performance in the context of content based image retrieval,the main contents are organized as follows:1.In order to describe an image with abundant texture information more concisely and accurately,we propose an image retrieval algorithm which is based on statistical models.Firstly,the gray scale image is decomposed into several sub-bands of frequency and orientation using the non-subsample Shearlet transform.Then we use the Bessel K distribution model to describe the coefficients of Shearlet high frequency sub-band.Both experiment and theory indicate that the BKF distribution is highly matched with the statistical features of coefficients within high frequency sub-band.Therefore,we use the BKF parameters as the texture feature to represent the characteristics of sub-band.Finally,we use Euclidean distance to measure the similarity of different images and respond to user with retrieval results.Simulation experiment shows that the algorithm is of low time complexity while it achieves an ideal retrieval result due to the reduction in data dimension and the outstanding representation method.2.In order to address the problem of incompleteness and incomprehensiveness of single feature representation,we propose a color image retrieval method based on feature fusion.The algorithm is based on the theory of radial harmonic Fourier moments(RHFMs)and the pyramidal dual-tree directional filter bank(PDTDFB).Firstly,we convert color images from RGB to Lab color space,and applying RHFMs to component a and b to extract the color feature of the images.Secondly,we use the PDTDFB to decompose the image in L component and extract image texture feature according to the variance and entropy of each sub-band based on the amplitude and phase.Finally,we carry out the parameter selection process to determine several parameters applied in the algorithm,e.g.the order of RHFMs,the level of PDTDFB and the optimal proportion for color and texture features.Then the color and texture features are normalized and assigned with the corresponding weights respectively.The simulation experiment shows that our proposed method,which make use of the fusion of multiple features rather than single feature,is able to provide retrieval algorithm with a more precise and comprehensive representation of the image.3.In order to address the open issue of “semantic gap”,namely one can never achieve an ideal outcome using low-level features and lose the sight of high-level semantic,we propose a color image retrieval algorithm which is based on relevance feedback mechanism.Firstly,we apply a novel active learning method to reduce the labeling effort of the user and mitigate problems of unbalanced and biased set of relevant and irrelevant images.Then we apply a classic kernel K-means algorithm which is based on diffusion distance to the process of choosing unlabeled samples and clustering them into the training set.Finally,we use an efficient SVM that can significantly improve the computational efficiency by setting a fixed slack variable for outliers.Simulation experiments show that the proposed algorithm,which is composed of the optimization of relevance feedback and sample labeling can effectively decrease the time of feedback,meanwhile achieve a comparatively ideal retrieval result.