Affine transformed image matching using Hough transformed planes

在图像中搜索对象是计算机视觉中的一项重要技术。在基于模型的匹配系统中，由于模型视图和真实图像之间的旋转和平移，目标对象在图像中的外观会随着模型的纵横比和比例的不同而有很大的不同。此外，如果输入图像的背景未知或有杂波，则匹配过程将进一步复杂。提出了一种利用Hough变换和Gabor变换在未知场景图像中搜索目标对象的有效方法。在后者中，通过对输入图像进行Gabor变换来获得输入图像的光谱信息。谱信息具有以下特点：(1)每个Gabor函数在频域展开。因此，很少有Gabor函数能够覆盖原始图像的全部频谱。这意味着它提供的特征数目很少，使得计算效率很高。(2)响应于模型的旋转，谱平面旋转。(3)响应模型的尺度变化，谱平面以恒定的速度改变其尺度，同时改变其幅度。利用第二和第三个特征，可以正确地检测出目标-目标，并通过光谱匹配计算出姿态。该方法包括四个步骤：(1)在模型图像的中心像素点和场景图像的所有像素点计算Gabor展开系数的幅值谱；(2)采用对数极轴采样对幅值谱进行变形；(3)考虑旋转和尺度变化的影响，作为仿射不变的同心特征，计算所有对数极坐标采样谱的傅立叶展开系数幅值谱。(4)在场景图像上的像素中找出与模型图像具有相同傅里叶幅值谱的特定像素。目标对象被认为位于该像素处。

Searching objects within an image is an important technology in computer vision. In model-based matching system, the appearance of the target-object in an image can be greatly different with the aspect and scale of the model, because of the rotation and translation between the views of the model and the real image. Furthermore if the background of the input image is unknown or with clutters, the matching process will be further complicated. This work presents an efficient method to search a target-object in an image with unknown scene by using the Hough Transform and the Gabor transform. In the latter, spectra information of the input image can be obtained by applying the Gabor transform to the input image. The spectra information has the following characteristics.(1)Each Gabor function is expanded in the frequency domain. Therefore, few Gabor functions could cover whole the frequency spectrum of the original image. This mean it provides very few numbers of features, resulting in the highly efficient calculation.(2) Responding to the rotations of model, the spectra plane rotates.(3) Responding to the scale changes of model, the spectra plane changes its scale while changing its magnitude with constant rate. Using the second and the third characteristics, target-object can be correctly detected and the pose is also calculated by spectra matching. The approach consists of four steps.(1) Amplitude spectra of Gabor-expansion coefficients are calculated at a central pixel on the model image and at all pixels on the scene image.(2) Amplitude spectra are deformed by a log-polar sampling.(3) Considering the effect of rotation and scale change, as an affine-invariant concentric feature, amplitude spectra of Fourier-expansion coefficients are calculated for all the log-polar sampled spectra.(4) A specific pixel having the same Fourier amplitude spectrum as the model image is found out among the pixels on the scene image. The target object is considered to be at the pixel.