Fusion Of Color And Shape Features Of The Image Retrieval Method

Image is an information carrier which can describe the objective world directly and vividly. It is bottleneck that the query image is found from the massive image database fast and efficiently. In the real life, research to the image retrieval has practical significance. For example, multimedia digital libraries, satellite remote sensing images, medical image management and computer-aided design and manufacturing, criminal identification system, geographic information systems, trademark and copyright management, all need the strong support from the research to the image retrieval.Essentially the early image retrieval is a text search which is based on the text annotation, it is subjective, linguistic background differences and time consuming. Content-based image retrieval(CBIR) overcoming the shortage of text annotation relies on the image feature information extraction, expression and similarity matching which are based on the characteristics of image-color, texture, shape and so on. As an important subject, the image retrieval algorithm based on one or more of the underlying characteristics have been widely concerned. Color-based image retrieval algorithm usually uses color histogram to describe image features, because it has many advantages, such as simple computation, significant effect, computation efficiency, insensitive to the position of lens, image distortion and rotation invariant. However it can not reflect the color distribution spatial information, results in a higher error rate. Retrieval based on texture features are commonly used to search this kind of image that has rich content and is difficult to distinguish between objects and background. Although Chain code, Fourier descriptors, Hu invariant moments, Zernike moments and pseudo Zernike moments can be used for image retrieval, but the rotation invariance, scaling invariance and translation invariance are not well resolved. For example Hu invariant moments has the rotation invariance, scaling invariance and translation invariance, but need to do a lot of image preprocessing and large amount of calculation. This paper summarizes results of previous studies and proposes an integrated color and shape of the image retrieval algorithm. Using the normalized moment of inertia (NMI) to describe the shape of the target image is innovation of this paper. The theoretical basis of this paper is that the normalized moment of inertia can reflect the spatial information and using multi-feature image retrieval can get a more comprehensive description of the image.In this paper, color space is proposed non-uniform quantization, the hue, saturation and brightness are quantified for the six, four, three-level, color space images are quantified for the 72 level. Non-uniform quantization can reduce the color feature dimension and system overhead. A target image normalized moment of inertia of the formula (1) is established. Using external normalization theory gets image feature vectors, and by equation (2) this paper gets similarity measure. The experimental search algorithm is carried out in MATLAB environment, and the experimental results are analyzed. Known by experiment, the traditional color image retrieval based on color histograms can't include the spatial distribution information of color, and leads high noise ratio. In this paper, the integration of color and shape of the image retrieval algorithm uses the color histogram for color features, and the Normalized moment of inertia to express the target image shape information, at the same time the adjustment for weight-based relevance feedback technique is used. The retrieval result is more accurate. And how to achieve high-level semantic features and special layer of precise communication, how to interact through the retrieval of user feedback in order to get a more satisfactory search results, needs to be addressed in further research..f (i, j) represents the image, m (f (i, j)) is the target image quality, CG(i,j) is recorded as the focus of the target area. NMI(f(i, j)) is the normalized moment of inertia of the target image.Hp andHQ are the image color feature vector, NMIP and NMIQ are the shape feature vector of the target images.