Research On The Key Technology Of Content-based Image Retrieval

With rapid development of multimedia and Internet technology, images are used extensively in all walks of life. Image data management and image retrieval have become increasingly important. However, the images contain abundant semantic information and complicated vision features, and the information redundancy is a serious problem. The traditional hand-marking and images indexing as text-based image retrieval methods cannot meet people's needs. Content-based Image Retrieval (CBIR) is to solve the above problems. In CBIR systems, a set of feature attributes, such as shape, color and texture are used to represent a digital image. Such a set of feature attributes are called feature vector or feature descriptor. The method based on the feature attributes (such as color, shape or texture) is to build feature database and index for images automatically. The need for efficient and accurate image retrieval systems is to facilitate the use of massive information more acute. The CBIR technology involves computer vision, image processing, pattern recognition, cognitive psychology, and many other technologies, which has important theoretical significance and practical application value.In this paper, lots of exploratory research work has been done around the key technologies of CBIR, which can be summered as follows,1. Some of the CBIR key technologies, including color space, low-level features of the image descriptors in MPEG-7, the low-level feature extraction, similarity measuring, image retrieval system evaluation standard, and relevance feedback were elaborated.2. In the study of color quantization and color histogram, a new image retrieval algorithm based on HSV non-uniform quantization dominant color blocking histogram was proposed. This method made use of HSV non-uniform quantization and image blocking. It improved the traditional histogram which is lack of space information. The basic idea of this algorithm was to divide image space into 3 by 3 blocks equally, and then to calculate the dominant color of each block according to the HSV non-uniform quantization. This method can greatly reduce the computation and storage space, and extract more efficient color feature. User can assign different weights to each block, according to its distribution in whole image. Finally, to make the weight of each block more accurate, a relevant feedback mechanism was utilized. Compared with classical color histogram, the results indicated that the proposed algorithm in this paper had better retrieval recall and precision.3. Though study the extraction of shape feature, a new method, which combined edge contour and Hu moment invariants, was proposed. This method can extract the information from features of image edges and features of the whole region of image. This method enhanced the precision of the shape feature. In order to obtain the accurate contour of the target, the eight neighborhood criteria was used to track the outline of edge image, and then the target area of image was calculated by seven moment invariants eigenvalues, which is called Hu moment invariants. The experiment results indicated that the proposed method can accurately describe the shape of target region, which had better retrieval result.4. In order to much more fully describe the image content, a comprehensive multi-features retrieval method was presented in this paper. This method included feature of both color and shape. User can assign different weights to the above both features respectively. The experimental results showed that this method was more effective than other methods using single feature, so retrieval recall and precision were enhanced. Finally, the ImageRetrieval systems as an experimental platform to verify the correctness of the proposed algorithm was designed according to the modular method, and then the structure, functionality and implementation of the ImageRetrieval system were described.