Image Feature Extraction Algorithms Based On Color Index And Clustering

Feature extraction is one of the important aspects in computer vision, which consists of CBIR(Content Based Image Retrieval,CBIR), image classification and so on. It is significant to extract image information exactly. Image feature extraction mainly utilizes auto-extraction of low-level visual feature (color, texture, profile, shape etc.) from image data, to deeply describe high-level semantic information. This paper first states the basic knowledge of CBIR: color space model, Image denoising algorithm, similarity measurement and comparison method if image retrieval performance etc., then emphases in the feature extraction algorithm based on image color and texture.In respect of color feature,a new color feature extraction method-color index local correlations and statistics(CILCS)is proposed:Pixels'color is represented sparsely in a manner similar to the color indexing of color histogram. The local and overall correlations of color index distribution in a certain image area are calculated, the correlations can reflect more color texture details.and have shown some rotation-invariance, scale-invariance and shift-invariance. This method improves the image retrieval performance with N-S dataset to CILAC by 9.02%-11.29%and with Corel image dataset, the recall vs precision curve is more steady and Classification accuracy improved by 2.33-3.2%.In respect of texture feature, In this paper, multi-resolution multi-direction transform domain method is intensively analysed, the Dominant Texture Descriptors (DTD) based on Gabor Translation is proposed, which fully utilizes magnitude and phase information in image translation domain, combines clustering algorithm to determine image domain texture composition and spatial distribution. In addition, Dominant Color Descriptors (DCD) is proposed. Then, combining both descriptors to form Dominant Feature Descriptors (DFD), can not only reflect texture distribution, but also include color information of features.A series of image retrieval and classification experiments are designed to test the related parameters and performances on the above-mentioned algorithm. The experimental results of image retrieval and image classification demonstrate their effectiveness.