Augmented Distinctive Features With Color And Scale Invariance

Yan Liu,Xiaoqing Lu,Yeyang Qin,Zhi Tang,Jianbo Xu

DOI: https://doi.org/10.1117/12.2005879

2013-01-01

Abstract:For objects with the same texture but different colors, it is difficult to discriminate them with the traditional scale invariant feature transform descriptor (SIFT), because it is designed for grayscale images only. Thus it is important to keep a high probability to make sure that the used key points are couples of correct pairs. In addition, mean distributed key points are much more expected than over dense and clustered key points for image match and other applications. In this paper, we analyze these two problems. First, we propose a color and scale invariant method to extract a more mean distributed key points relying on illumination intensity invariance but object reflectance sensitivity variance variable. Second, we modify the key point's canonical direction accumulated error by dispersing each pixel's gradient direction on a relative direction around the current key point. At last, we build the descriptors on a Gaussian pyramid and match the key points with our enhanced two-way matching regulations. Experiments are performed on the Amsterdam Library of Object Images dataset and some synthetic images manually. The results show that the extracted key points have better distribution character and larger number than SIFT. The feature descriptors can well discriminate images with different color but with the same content and texture.