Research And Improvement Of The Inverse Compositional Image Alignment

Image alignment or image registration is the process of overlaying two or more images of the same scene taken at different times, from different viewpoints, and/or by different sensors according to some criterion. It has been widely involved in a lot of applications, such as medical image analysis, computer vision and remote sensing. In general, image alignment algorithms can be divided into two groups, pixel-based image alignment and feature-based image alignment. We mainly cover one of the most typical pixel-based algorithms—inverse compositional image alignment (ICIA).In section 2 we introduce the derivation and main content of the ICIA algorithm and its three important extensions: the inverse compositional algorithm with a weighted L2 norm, the inverse compositional iteratively reweighted least squares algorithm and the inverse compositional algorithm with a prior. In section 3 we present how to apply the ICIA algorithm in Active Appearance Models (AAM) and then develop the program of the original AAM algorithm and the inverse compositional AAM algorithm. We testify that the inverse compositional AAM is more efficient than the original one by an experiment. Traditional ICIA algorithm compares intensity values between a template image and an input image. Although it is efficient because most of its computationally expensive calculations are done at pre-computation phase, it is easily affected by the variations of the lighting conditions, which would lead to poor convergence or even divergence. However, the edge structure of an image tends to be less sensitive to lighting conditions than intensity values. Hence, in section 4 we propose the inverse compositional gradient (ICG) algorithm, which is a novel ICIA algorithm based on the local orientation. We demonstrate that our algorithm could efficiently overcome the influence of the lighting variations on the matching results. In addition, the evaluation of the local orientation and gradient is essentially equivalent. Thus we used 4 different gradient operators (first order difference operator, Roberts operator, Sobel operator and Prewitt operator) to compute the local orientation, and then compared the performance of ICG algorithms using above 4 operators under different lighting conditions and noise. Experimental results show that the ICG algorithm using the first order difference operator performs best if there is no noise in images. And the ICG algorithms based on Sobel operator and Prewitt operator would give better matching results if there is noise.