Research On Content-based Image Retrieval Technology

Due to the development of multimedia and Internet techniques, people can get more and more image information. Therefore, it is critical for people to retrieve valuable information quickly and efficiently from large-scale image databases. Content-Based Image Retrieval (CBIR) technology emerged as the times required, and becomes an important and challenging research topic. In this thesis, lots of exploratory research work has been done around some key techniques of CBIR, which include low-level feature extraction, image retrieval algorithms and so on. The main works of this dissertation are sununarized as follows.Firstly, the Krawtchouk moments are applied into both texture-based and shape-based image retrieval. The Krawtchouk moment invariants based on Krawtchouk polynomial have the invariance on translation, rotation, scaling changes, so a novel method based on Krawtchouk moment invariants is proposed for shape-based image retrieval. At the same time, using the nice image representation capability and minor noise susceptibility of Krawtchouk moment, a novel method based on Krawtchouk moment texture feature is realized. Compared with Zernike method, experimental results show the proposed methods are effective.Then, an image retrieval method using the global and local shape features is realized. Firstly, an image is segmented, and then the Compactness and Fourier Descriptor as local features are extracted. In order to remedy the effect of image segmentation on feature description and improve retrieval performance, global feature is extracted by Krawtchouk moment invariants. Finally, this approach uses the combined local and global shape features as feature vectors to achieve image retrieval. With plentiful experiments, it is proved that the precision can be enhanced by using this method.Next, this paper has proposed an image retrieval method using texture feature and shape feature. Firstly, the image is decomposed to the low-resolution image with NSCT (Nonsubsampled Contourlet Transform), extracting the contourlet coefficients of each decomposition level of different directions and using the statistical features of the coefficients as texture features of the image. Then, the Krawtchouk moments are employed to extract shape features of the image. Finally, the image retrieval is achieved based on distance measure. Through the image retrieval experiments,better image retrieval performance can be achieved by combining two kinds of features.Finally, a method of binary trademark retrieval based on regional Gaussian descriptor is introduced. Aiming at the importantance of shape feature in binary trademark retrieval, using the regional Gaussian descriptor to achieve image retrieval.This descriptor has some advantages including the invariance on translation, rotation, scaling changes and reflection, high recognition rate and matching rate and edge variations. Experimental results show this method is effective on trademark retrieval and obtains high recall and precision.