Content-based Image Retrieval Based On Multi-feature Fusion Of Interest Points

Content-based image retrieval based on multi-features fusion of interest pointsWith the continuous development of multimedia technology and the wide application of the Internet and the widespread use of digital imaging equipment, more and more large-scale digital image databases form. How to find images that user needed in these databases attracts people’s attention, and image retrieval technology developes rapidly. Text retrieval which needs artificial marked text description such as keywords was used at the beginning, but it no longer meet the need of image retrieval with increasing of the size of the image database. So content-based image retrieval technology is put forward and widely used, it can be automatically carried out by the computer. Content-based image retrieval (CBIR) refers to the retrieval that uses features of image such as color, texture, shape, semanteme to indicate image content, and uses these features to complete image retrieval. CBIR uses low-level features to express image content, attempts to find similar images on the basis of studying the image content, which has characteristics like small amount of calculation, rapid retrieval, less artificial intervention, and so on. The traditional CBIR uses single feature to achieve image retrieval, it has the obvious problem that lack of semantic coherence, so the results of retrieval are not very ideal. In order to overcome these problems, a lot of improvement methods of CBIR are put forward and used. Local area feature extraction and multi-features fusion technique are the mainly two aspects.This article puts forward a method of image retrieval that based on multi-features fusion of interest points, which bases on the study and research of the interest points of image combining with the characteristics of image retrieval and reading a lot on the basis of the related literature at home and abroad. The method uses interest points which are less contained in the image as the key of retrieval, extracts image features related to interest points (color, texture and discrete degree) for image retrieval, and obtains good retrieval results. Specific content as follows:On the basis of learning and studying existing interest points extraction methods, we put forward a new method to extract interest points. Interest points in image are the pixel which contains a large amount of information, are also regarded as the visual concerns. Interest points have little proportion in images, and are convenient for the various applications and operations. There are various of extraction methods of interest points, of which the most widely used is Harris operator method, it has good computability, stability and robustness. The parameters k in Harris method is not easy to determine, this paper puts forward a new method to improve. The new method takes the characteristic value of the autocorrelation matrix as the basis to judge interest points, when two characteristic value of the matrix are both large value s, the pixel of the matrix center is the interest point. Considering that it is complex to get the characteristic value, we use the ratio of determinant and trace of autocorrelation matrix as response operator. The operator has the same calculation complexity to Harris operator, which can avoid the influence of selecting parameter k to the result. Select the certain proportion value of local maximum as a threshold in the process of filtering the points that are not interest points. So the interest points extracted by the method accord with the whole situation of the image, is more sensitive to the change of the gray value than Harris, can find more reasonable interest points.Extracting the image characteristics include color, texture and average discrete degree which related to interest points as image features, on the basis of extracting interest points and combining the characteristics of image retrieval. Color is the most common used characteristic of image features, can reflect the information of image well such as the content, and is not sensitive to the operation of rotation and translation. This paper bases on interest points, takes color histogram of interest points to characterize image color characteristics, join spatial distribution characteristics and color of image together well. We find centroid according to the distribution of interest points, and regard centroid and the maximum distance from interest point to centroid as the center of a circle and maximum radius, and construct three equidistant concentric rings. Then we find interest points in each annular region, extract the color value of the pixels in certain neighborhood to make a color histogram. This method puts spatial distribution characteristics into the color when it makes the histogram. The texture is an important feature of the image, plays an important role in guiding the retrieval. This paper brings the idea of block truncation coding in the process of extracting texture feature. First we take interest points as the center to establish four small pieces according to four directions of upper left, upper right, lower left and lower right, encoding the pixels in the block to binary code in accordance with the difference to mean value. Then we take the decimal of small piece after encoding as its texture value, construct histograms according to the values in four directions of interest points of the texture, and the histograms can characterize texture feature of image. In order to depict the distribution of interest points in images,we present the concept of the average discrete degree of interest points. It can reflect the polymerization of interest points, and reflects structure of the image content to a certain extent. We synthesize characteristics including color, texture and discrete degree based on interest points and realize the image retrieval based on multi-features fusion.First calculate the similar degree of three low-level features of their characteristics respectively, and then use the Gaussian normalization method to dispose different similarity, and fuse the similarity disposed of the features as the image similarity. This paper proposes the experiment results and analysis of the image retrieval method, the results show that the image retrieval method based on multi-feature fusion of interest points is feasible and effective.