Image Retrieval Based On Texture

Over the past few years, lots of digital images in the fields of civilian satellites, medical, health work and fingerprinting have produced every day due to the wide application of the technology of computer and web. In order to retrieve these images efficiently, the retrieval methods based on Gabor wavelet, dual-tree complex wavelet transform, generalized Gaussian density, KLD, Local texture statistic model and Log-likelihood statistic are well studied in the thesis, therefore the paper presents research on texture based image retrieval.Firstly, a novel texture image retrieval approach based on generalized Gaussian distribution statistic model in the Complex Wavelet Domain is presented. The paper presents a statistical view of the texture retrieval problem by combining the two related tasks, namely feature extraction and similarity measurement, into a joint modeling and classification scheme. Feature extraction is obtained by the dual-tree complex wavelet transform (DTCWT) which is a valuable enhancement of the traditional real wavelet transform that is nearly shift invariant and, in higher dimensions, directionally selective (filter coefficients give texture information strongly oriented in six different directions). Besides, the DTCWT subband coefficients can be modeled by generalized Gaussian density (GGD). Therefore feature extraction and similarity measurement are combined into a joint modeling and classification scheme. The statistical scheme leads to a new wavelet-based texture retrieval method which is based on the accurate modeling of the marginal distribution of complex wavelet coefficients using generalized Gaussian density (GGD) and on the existence of a closed form for the KLD between GGDs. To check the retrieval performance, the texture database is created. Extensive experimental results on a database of 600 texture images indicate clearly that the proposed method significantly improves retrieval rates, e.g., from 82.50 to 89.46, from 70.66 to 89.46 in comparing with complexwavelet transform and Gabor based approach. The proposed method also retains comparable levels of computational complexity.Secondly, a retrieval approach based on local texture statistic model combining with improved Log-likelihood statistic is presented. Local binary patterns (LBP), a very simple method, is very sensitive to rotation and has 256 possible texture units. For these problems, the"uniform"patterns LBP which can achieve gray-scale invariance and rotation invariance is introduced. Different radiuses and directions can be designed according to various cases in this method. Nine"uniform"patterns LBP8 r,1iu2, seventeen"uniform"patterns LBP1 r6i,u22, twenty-five"uniform"patterns LBP2 r4iu,32 are designed in the thesis. In most cases a single texture measure cannot provide enough information about the amount and spatial structure of local texture. Better discrimination of textures should be obtained by considering joint occurrences of two or more features. The 2,riuLBPP R operator is a gray-scale invariant, i.e., its output is not affected by any monotonic transformation of the gray scale. It is an excellent measure of the spatial pattern. However, it discards contrast by definition. Therefore, we can measure it with a rotation invariant measure of local rotation invariant measure of local variance VAR P, R.The VAR P, Ris by definition invariant against shifts in gray scale. Since LBPP r,iRu2and VAR P, Rare complementary, and their joint distribution 2,riuLBPP R/ VAR P, Ris expected to be a very powerful rotation invariant measure of local image texture. Owing to lower precision of log-likelihood statistic distance measure in texture image retrieval, this paper applies improved Log-likelihood statistic. Extensive experimental results on a database of 600 texture images indicate clearly that the retrieval rate of nine"uniform"patterns LBP8 r,1iu2is 89.7083%, which significantly improves retrieval rates, e.g., from 82.50% to 89.7083%, from 70.66% to 89.4083% in comparing with complex wavelet transform and Gabor based approach.Finally, the proposed methods above are applied to the medical image retrieval. Immense improvement has been obtained in the field of content-based image retrieval (CBIR) in the past few years; nevertheless, existing systems stillfail when applied to medical image databases. In medical image processing, texture is one of the most important features because it is difficult to classify human body organ tissues using shape or gray level information. Medical images are similar, multimodal, heterogeneous and higher dimensional with temporal properties, which distinguish them from images in other domain. Because the character of medical image is different from that of natural image, and some traditional methods of texture feature extraction can not get satisfying results. For these problems, the paper applies the proposed methods above to the medical image retrieval.In order to prove the feasibility of these methods, a database of 175 medical images including magnetic resonance imaging (MRI) and Computed Tomography (CT) is built. Experimental results indicate clearly that these methods which are fit for texture image retrieval can not fit for medical image retrieval. The method of texture image retrieval based on generalized Gaussian distribution model in the complex wavelet domain can not attain very good experiment result whose retrieval rate is only 72.7143%, however, the retrieval rate of the texture image database based on this method is nearly 90%. At the same time, the local information is considered in the method of texture image retrieval based on local texture statistic model, thus the retrieval rate is high. In the medical image database, Log-likelihood statistic is nearly feeble, but the improved Log-likelihood statistic can attain the superiority.