비선현 비레 공간에서 명시적 고속 확산을 이용한 불변색상 특징 추출 및 기술

Abstract

In recent years, several works have aimed to find the most effective and accurate feature extraction method. Considering the computational time and simplicity, most of state-of-the-art methods operate on grayscale images by applying color conversion mechanism. In the other hand, conversion to grayscale images causes some important information lost which potentially reduce the performance. In this paper, the authors propose a multi-scale 2D invariant color detection and description algorithm in nonlinear scale spaces. The algorithm exploits and utilizes color information of image in Hue, Lightness, and Saturation (HLS) space. Therefore, nonlinear scale spaces are built separately for each color channel (HLS) and adaptive integrated determinant Hessian responses are calculated in finding keypoints of the image.

We implement Fast Explicit Diffusion (FED) scheme in the nonlinear scale spaces which causes locally adaptive blurring to the image data. It can reduce image noise but retain object boundaries in obtaining superior localization accuracy and distinctiveness, where linear scale spaces system such as Gaussian blurring smooth image to the same degree both details and noise. In addition, the authors propose a Color Modified-Local Difference Binary (CM-LDB) descriptor which is exploits HLS color information concurrently with its gradient information from nonlinear scale spaces. Besides the improvements of precision and accuracy in general image matching, our proposed system (HLS-AKAZE) shows significant performances improvements compared with other grayscale features extraction methods in the certain image which has low / unclear information in its grayscale image. Moreover, this proposed system is mostly invariant in the rotated, scaled, and illumination-changed images.