Model-Based Object Recognition by Maximization of Combined Mutual Information of Intensity and Orientation
Yong Sun Kim, Jae hak Lee, Jong Beom Ra · 2006
We propose a model-based object recognition algorithm to establish accurate correspondence between a 3-D model and a 2-D image. We try to solve this problem by matching the 2-D image with various projections of the 3-D model. On the contrary to feature based approaches, we use normalized mutual information (NMI) as a similarity measure, which is known to be accurate for multimodal image registration. NMI is usually regarded as a measure representing statistical dependence of intensities between two input images. However, when the number of samples is small, the statistical dependence of intensities may not be reliable and a complementary term is needed. In this paper, we introduce image orientation information as an additional NMI measure. The experimental results show that the proposed NMI method provides a better registration performance than the conventional intensity-based method.