Image registration is a challenging problem for computer vision, and accurate and effective image registration is still required in various computer vision applications, e.g., 3-D scanning, autonomous navigation, and augmented reality. However, image registration becomes difficult due to the presence of noise and photometric changes. This paper presents a novel image registration method with unsupervised inverse intensity compensation (ICIR). This methodology uses weighted vectors to compensate for areas affected by radiometric variations. This is a 5-D vector body composed of RGB, brightness, and gradient, that is, each pixel is represented by a 5-D vector in its neighborhood. When performing image registration, the vector angle metric robust to illumination effect is used to calculate cost volumes. Then the selected cost metrics are aggregated based on RGB-Gradient tree structure. Experiments performed on stereo images of the Middlebury datasets and ours demonstrate this methodology in calculation accuracy and time all have good performance.

图像配准对于计算机视觉来说是一个具有挑战性的问题，并且在各种计算机视觉应用中仍需要准确且有效的图像配准，例如3-D扫描，自主导航和增强现实。然而，由于噪声和光度变化的存在，图像配准变得困难。本文提出了一种新的无监督逆强度补偿（ICIR）的图像配准方法。这种方法使用加权矢量来补偿受辐射度变化影响的区域。这是一个由RGB，亮度和渐变组成的5-D矢量主体，也就是说，每个像素都由其附近的5-D矢量表示。在执行图像配准时，对照明效果具有鲁棒性的矢量角度度量将用于计算成本量。然后，基于RGB梯度树结构汇总所选的成本指标。对Middlebury数据集的立体图像进行的实验以及我们的实验证明了这种方法在计算准确性和时间方面均具有良好的性能。