Research On Texture Image Retrieval Based On Multiscale Geometric Analysis

Along with the speedy development of the multimedia and Internet techniques, content-based image retrieval (CBIR) has become one of the hot topics on image processing and computer vision, content-based image retrieval has broad application prospects and important academic value. In recent decades, although many image retrieval methods have been proposed and all of them have achieved good performance in a certain extent so far, image retrieval precision needs to be improved further. How to search for desired images from large-scale image databases quickly and accurately is still a challenging task so far. In order to improve average retrieval accuracy, this paper focuses on similarity measure, rotation-invariant texture feature extraction, multi features extraction, texture image classification and retrieval based on support vector machines, probabilistic texture retrieval.The main works and research results of this dissertation are summarized as follows:A complete texture image retrieval system includes texture features extraction and similarity measure these two parts. Similarity measure is a key technology for texture image retrieval, Kokare. et al compared nine distance similarity measures such as Euclidean distance, Canberra distance, Bray-Curtis distance, Weighted-Mean-Variance distance(WMVD), Manhattan distance, Mahalanobis distance, Chebychev distance, Squared Chi-Squared distance, and Squared Chord distance for texture image retrieval. Their experimental results indicated that the retrieval performances are the best ones by using Canberra distance and Bray-Curtis distance measures comparing to the rest seven distance measures based approaches, but the retrieval rates by using Canberra distance and Bray-Curtis distance measures are not still ideal enough. Finding good similarity measures is a challenging task. This paper presents an effective similarity measure based on distance for texture image retrieval. The proposed similarity measure firstly uses denominator to normalize the difference between the two image features, then the differences in each dimension are squared, and then they are summed before extracting square root. The proposed similarity measure can comprehensively use all features. Experimental results demonstrate that the proposed similarity measure based on distance improves retrieval accuracy, compared with Euclidean distance, Canberra distance, Bray-Curtis distance, Weighted-Mean-Variance distance.In order to extract the more effective rotation-invariant texture feature, many kinds of analytical methods about rotation-invariant texture have been proposed. Among those, model based methods have relatively good effects for rotation-invariant texture feature extraction to a certain degree, but this kind of methods have a common drawback, i.e., the selection of a suitable model for the different textures is considerably difficult. Besides, some frequency-domain based methods have also been proposed and effectively applied in rotation-invariant texture analysis. Han and Ma used Garbor filters to obtain invariant texture features, but the number of Gabor channels is not easy to determine. Do and Vetterli proposed an invariant texture retrieval method by using steerable wavelet-domain hidden Markov models and developed a wavelet-based texture retrieval method by using generalized Guassian density and Kullback-Leibler distance. Manthalkar et al used discrete wavelet packet transform to extract rotation-invariant texture. Discrete wavelet transform has shift sensitivity and poor directionality, so Kokare et al used dual-tree rotated complex wavelet filter and dual-tree-complex wavelet transform to improve rotation-invariant texture image retrieval accuracy. This paper presents an effective approach for the rotation-invariant texture image retrieval with NSCT which has anisotropy and translation invariability. To achieve rotation invariant features, we use the mean of energies and standard deviations of all subbands at each NSCT scale. The proposed similarity measure between two images is not affected by any other images, and the measure need not calculate the statistics of the entire image database in advance, so it has much wider application situation, e.g., internet. Experimental results demonstrate that the proposed method improves retrieval accuracy, compared with some existed existedapproach.Multiple rotation invariant features have better ability to describe the texture of image. This paper presents a rotation-invariant texture image retrieval algorithm using multiscale geometric analysis and multi-features. The subbands coefficients of NSCT can reflect the texture and details of image. NSCT has anisotropy and translation invariability. GLCM of image reflects information about direction, adjacency spacing relationship, and the range of variance change. By computing the average energy and average standard deviation of all subbands at each NSCT scale as well as the mean and covariance of the second moment angle, inertia entropy, inertia moment, contrast points moment, and inertia correlation of GLCM, rotation-invariant features can be achieved. Experimental results demonstrate that the proposed algorithm improves retrieval accuracy from76.09to80.71%on the rotated database (640images), compared with DT-RCWF and DT-CWT-based approach.This paper presents an approach of texture image classification based on nonsubsampled contourlet transform, Local binary patterns and Support vector machines. Nonsubsampled contourlet transform and Local binary patterns are used to extract texture features of images, Support vector machines are used to classify texture images. Nonsubsampled contourlet transform has translation invariability. Local Binary Patterns has rotational and gray invariance. Support vector machines have good performance in a variety of pattern recognition problems. Experimental results demonstrate that the proposed method performs much better than some existing methods. It achieves higher classification accuracy.In order to improve texture image retrieval accuracy, this paper presents an approach for probabilistic texture retrieval using nonsubsampled contourlet transform. Nonsubsampled contourlet transform is used to extract texture features from images, Gamma distribution is used to model the marginal distributions of nonsubsampled contourlet transform coefficients. Image similarity measurement between images is gotten by using the Kullback-Leibler divergences between the statiscal models. Experimental results demonstrate that the proposed probabilistic texture retrieval method performs much better than some existing methods.