A Statistical Image Retrieval Method Using Color Invariant

Content based image retrieval is an essential task in many image processing applications, among which, color based methods have been receiving constant attentions in past years, because color information is a discriminative descriptor for image retrieval, especially in case of large database. A limitation of previous color based methods is their unsuitability for retrieving similar scenes under varying lighting conditions as color is sensitive to illuminations. Besides image descriptors of some existing methods are with large dimensionality and thus computational expensive. As betterment, an adaptive method is proposed in this paper, which integrates the color invariant with some spatial information of images. Different from previous work, the number of states during the quantization of the color space is not manually determined. Instead, it depends on the context of the image itself, using an adaptive clustering technique: Firstly, feature map consisting of color invariants is established for images. Secondly, the Markov chain model is employed to capture the image both color and spatial information. Thirdly, an image descriptor is computed for each image, not under the frame of the entire fixed color space. To practice our method, similar images are retrieved with a similarity measure based on a two-stage weighted distance. Experiments show that, this method has improved simplicity and compactness without the lost of efficiency and robustness.