Image Retrieval Based On Sparse Coding And Its Application

The rapid development of Internet technology makes the network information dissemination grow very swiftly.The images from varieties of blogs,forums and social networking sites are flowing into the Internet,which causes the exponential growth of the scale of image databases.It has become more and more difficult to retrieve quickly and accurately the effective information from amounts of image databases.Therefore,efficient image retrieval technologies have attracted wide attention,so the content-based image retrieval technology has become a research focus.With the development of the sparse coding technology and considering its efficiency in respect of image processing,it has a wide range of applications in the many image fields.Therefore,the image retrieval based on sparse coding becomes an important research topic of the content-based image retrieval technology.In this thesis,we mainly research the image retrieval based on sparse coding.The bag-of-visual-word(BOVW)model of image retrieval ignores the image feature spatial structure,thus causing quantization error.In order to overcome this shortcoming,we combine the advantages that the sparse coding technology can efficiently process images and retain the local information of the feature space and take into account the importance of feature space geometric structure,then propose several image retrieval algorithms based on sparse coding.The major contributions are as follows:(1)An image retrieval algorithm of sparse coding is proposed based on modified spatial pyramid matching.Since spatial pyramid structure can effectively preserve image spatial location information,we first segment an image according to the spatial pyramid.Then we establish the sparse regularization item by integrating the local features information with the sparsity of sparse coding,acquired the sparse coding formula will be convex optimization problem.In order to get more accurate retrieval results,a similarity association calculation method is proposed according to image itself structure and two similarity computation ways.This algorithm ameliorates the fault of the BOVW model and improves the retrieval accuracy.(2)An image retrieval algorithm of sparse coding is presented based on sub-region bag-of-visual-phrases model.In the sparse coding process,the encoding coefficients may not be unique because of the over-complete dictionary,which leads to the poor robustness.Aiming at this problem,we build a bag-of-visual-phrases model.An image will be divided into a series of sub-regions according to corners and features,we extract image features of sub-regions and then encode these features.We calculate the feature arrangement histograms of sub-regions,then construct the visual phrases combining the histograms with the sparse coding of sub-regions.The similarities will be calculated through visual phrase histograms of images.This algorithm combines the efficiency of sparse coding with the robustness of BOVW model,not only retaining the local features correlation but also improving retrieval stability.(3)We propose an image retrieval algorithm based on Laplacian sparse coding.Owing to the independent encoding process,the sparse coding loses the local similarity information and ignores the features spatial geometric structure,which leads to big reconstruction error.Laplacian Eigenmap is able to reserve the local adjacency relations of features.Considering the above reasons,we establish the regularization item combining the local information of features with the relevance of encoding coefficients.The similarity matrix is computed according to the distance between features,it will be as the weight matrix used for defining the Laplacian matrix.Then we construct the Laplacian sparse coding formula on the basis of the Laplacian matrix.The optimal encoding coefficients will be solved adopting feature-sign searching algorithm.Proposed retrieval algorithm guarantees that similar features have similar encoding coefficients and improves retrieval efficiency as well.(4)We bring forward an image retrieval algorithm based on Hessian sparse coding.The convergence speed of Laplacian sparse coding algorithm is fast,but it can not deal with complex images very well.Considering the Hessian Eigenmap can efficaciously preserve the local manifold geometrical structure of image features,we map the points on manifold into the local tangent space and define the second order Hessian energy function by the local Hessian quadratic.Then,we construct the sparse coding formula based on Hessian Eigenmap.In order to make better use of the relationship of neighboring features,we build the bag-of-visual-phrases model——n-words model inspired by the ideas of the binary model in text document.n-words sequence are extracted from n-words model,which are high-level visual descriptions.All the n-words sequences will be used as the feature descriptors of Hessian sparse coding.This algorithm strengthens the recognition power for complex images and promotes the overall level of retrieval efficiency.(5)We propose an image retrieval algorithm based on symmetric positive definite kernel sparse coding.Hessian Eigenmap needs to estimate the second order partial derivative,therefore the Hessian sparse coding is extremely sensitive to noise.The calculation of local tangent space on manifold is quite complicated,which makes retrieval speed of the Hessian sparse coding very slowly.The kernel method do not need to complicatedly calculate and approximately estimate and can effectively deal with the nonlinear structure of Riemannian manifold.Therefore,we construct a sparse coding formula based on kernel function utilizing the efficiency of kernel method.We first carve up an image into a series of sub-regions by the 8×8 templet.Then we extract the scale invariant feature transform(SIFT)descriptors of all sub-regions and calculate their covariance matrices,then construct the symmetric positive definite manifold.We define the symmetric positive definite kernel function by Stein divergence,and the symmetric positive definite manifold will be embedded into reproducing kernel Hilbert space using the kernel technology,we can acquire the sparse coding formula based on the symmetric positive definite kernel function.According to the iterative algorithm,we train the Riemannian dictionary,then compute the optimal encoding coefficients by Riemannian dictionary.This algorithm not only can recognize complex images accurately but also can accelerate image retrieval.In conclusion,this thesis focuses on image retrieval algorithms based on sparse coding.We strengthen the image visual description through effectively using image spatial information and improve the image encoding efficiency though further combining with feature multilayer semantic information,and thus image retrieval precision is also ameliorated.