Research On Key Techniques Of Content-Based Image Retrieval

With the rapid development of the Internet and multimedia techniques,the amount of multimedia information is increased rapidly.Therefore, the rapid and effective management for large-scale image databases becomes an important research topic. In this research area, image database system plays an important role in the multimedia information system because of its abroad employment in many important applications such as: multimedia information system, digital library, medicine, geographical information database and military defense. Currently,rapid and effective searching for desired images from large-scale image databases becomes an important and challenging research topic. Content-based image retrieval (CBIR) is the set of techniques to address the problem of retrieving relevant images from an image database based on automatically derived image features. In this dissertation,lots of exploratory research work has been done around some key techniques of image feature extraction of CBIR, and the low level visual features extraction of images have been studied systematically, which include color, shape, texture, color spatial distribution features and so on. The presented study is the current research focus of image processing and information retrieval. Thus, its research has both the theory and the application value.The main contributions of this dissertation are summarized as follows:1. Several key techniques and algorithms of CBIR are deeply analyzed and discussed,such as,color model, the low-level feature descriptions including color, shape, and texture,the similarity measure between the features,image database indexing methods and the evaluation methods of image retrieval algorithms. Moreover, by some experiments tested under the same conditions, we report the comparison results of many classical methods.2. A new kind of color image retrieval algorithm based on color and spatial features is presented.The model of generalized image is introduced into the retrieval systems .In the first place, the generalized image can be attained by the model of generalized image.The spatial-color moment of generalized image can be attained based on the annular color histogram. The spatial-color moments of every kind of color can be used as the character vector of color images. Experiments indicate that this method gives better performance than SCH and geostat.The SCH algorithm and geostat algorithm reflect the color information of isolated point. This algorithm expresses the statistical character of color block and can denote the information of color and spatial features. Thus, this algorithm has a higher retrieval-rate than those of SCH and geostat.3. A new kind of image retrieval algorithm based on multi-scale Radon transform is presented. The shape character vector with shift, scaling and rotational invariant is constructed based on the Radon transform of edge images, which is gained by the wavelet modulus maximum. Experiments indicate that this method has the property of translation, scaling and rotational invariant and has a higher retrieval-rate than that of wavelet HU moments.Since the Radon transform is robustness to additive white noise, this method can denotes the steady features of image. Thus, this method can give better performance than the wavelet HU.4. A new retrieval algorithm based on the multiwavelet energy and entropy is presented. The preprocessing image can be decomposed by the multiwavelet transform.The repeated row preprocessing is used in this paper because the repeated row preprocessing has more texture information than other preprocessing methods. The multiwavelet energy and entropy can be used as the texture feature of image. Experiments indicate that this method gives better performance than FWHT and multiwavelet histogram technology. Bacause the scalar wavelet has no the property of symmetry, orthogonality, shorter support for given approximation order and high vanishing order at the same time, the retrieval-rate of classic scalar wavelet algorithm is restricted to a great extent. Thus, this algorithm has a higher retrieval-rate than those of FWHT and multiwavelet histogram technology.5. Two new retrieval algorithms based on the Radon and wavlet transform is presented based on the current texture retrieval algorithm. According to statistical property of the Radon transform, the translation- and rotation- invariant texture retrieval algorithm is proposed and the translation-, rotation- and scale-invariant texture retrieval algorithm is proposed. Experiments indicate that the two methods is robustness to additive gauss noise and the geometrical transform, and have a higher retrieval-rate than that of other methods.