A Rapid Image Approximation for Keypatches Building

A new type of local features with visual semantics, keypatches, was proposed to improve performances of image retrieval methods. These features are obtained by approximating the content of scanning windows using the predefined geometrical patterns. To build keypatches quickly, this paper proposes a new approximation based on a novel compact gradient mapping table (CGMT). The method includes two stages. First, the approximation candidates can be obtained quickly by constructing and looking up the compact gradient mapping table based on the distribution sharing of the contours characteristics of the patterns. Then, Hough transform needs to be applied only in a very small parameter space defined by the candidates for the final approximation. To demonstrate the effectiveness of the proposed method, we have approximated 1628 image patches from 100 images. Results show the proposed method significantly outperforms the previous Hough transform-based approximation in speed with a very small accuracy loss.