Research On Content-Based Image Retrieval Methods

Along with the rapid development of multi-media technology and storage technology, there will be producing a large number of digital images in medical, engineering, science, photography, advertising field, how to manage and use these images becomes an urgent problem we are facing. Traditional method using text annotation of images clearly can not meet the requirement of dealing with massive images, so scholars proposed content-based image retrieval (CBIR), and it got widely studied.This paper firstly elaborates the development and research of content based image retrieval home and abroad, and gives a depth analysis of each key technologies of content-based image retrieval, these technologies including image feature extraction, similarity measure, relevance feedback and retrieval performance evaluation.Secondly, we studied the image’s color feature extraction methods, for the problem of excessive dimensions of color histogram, we propose a dimensionality reduction method based on sample. Images of the same class should have some color dimension with same distribution (main color) against other image classes. So we can analysis and calculate the distribution of the main color, and extract these color dimensions to compose the new color feature, by this process we can achieve color feature dimension reduction, reduce the computational complexity and improve precision.Thirdly, we studied the image’s shape feature extraction methods, when calculating the shape feature by Zernike moment, we also calculate the background messages beside the image object, and these background messages are not necessary for use to recognize an image. So we use Canny operator to get the edge image of the original image when calculate image’s shape feature, then we use Zernike moment to calculate the edge image as the shape feature. Thus using the pure white background to replace image’s original background and keeping the shape message of the image.Fourthly, we first studied the texture feature extraction methods of image, and then proposed an analogy-relevance feedback image retrieval method using multi-features. The method find the key images by the specified image’s interesting object, and using the key images to re-calculate the similarity calculated directly, and then retrieving images by the modified similarity. Experiments show that the proposed method has a higher precision and recall ratio than the traditional content-based image retrieval method.Finally, we use MATLAB and MFC design and implement a content based image retrieval test system based on local image database to verify the method, the system is simple and can retrieval very fast.