The Research On Texture-Based Image Retrieval

The key technology for the image database is the image retrieval, which has become a bottleneck of the mass information processing. In fact, the contents of images are the inbeing of image retrieval. Content-based image retrieval (CBIR) has been an active research area in the field of image processing. Feature extraction and similarity measure are two most important issues of CBIR. In real world, almost every object has its own textural properties. Everyone can recognize texture, but it is more difficult to be defined. Texture as a primitive visual cue has been studied for over twenty years, but the applications of texture analysis to image data and image retrieval have been very limited to-date. In recent years, texture-based image retrieval has already become a hot topic in CBIR.We firstly review and compare different methods of texture feature extraction and texture models, and categorize some common similarity measures, showing strengths and limitations of each method. Our main research work is to study what is the effective texture feature and how to accurately measure the similarities between texture images.In this thesis, we propose the use of Gabor wavelet features for image retrieval. This kind of texture features integrates the advantages of Gabor transform, wavelet and multiresolution. A texture, then, is represented by a vector of values, each corresponding to the energy in specified scale and orientation subband. We also propose a simple method to improve the Gabor texture features. Depended on the characteristic of total energy for each orientation we make a circular shift on the feature map to solve the rotation variant problem associated with Gabor texture features.Similarity measure of texture features is the other important content in this thesis. We propose a novel method to measure similarity of texture features, which we called as Cluster Space Model (CSM). In CSM, the signatures, instead of histograms, are get by clustering algorithm to find dominant clusters and their weights in the feature space, and used to represent the full distributions of image texture in a compact way. The Earth Mover's Distance (EMD) is applied to measure similarities between texture signatures. We do a comprehensive study on the EMD, including its definition, properties, computing theoretic basis and algorithm. We prove the EMD is a metric...