Super-Resolved Image Mosaicing from Image Sequence Taken by Moving Camera

在这项研究中，我们的目标是开发一种方法来合成超分辨率的图像，提高空间分辨率的输入图像合并多个输入图像由一个移动的相机。当移动的摄像机拍摄图像序列时，图像的某些部分在不同的帧中重叠。在这种重叠部分中，检测特征点以获得帧之间的逐点对应。该对应信息通过应用因式分解方法提供对象的形状。虽然因式分解方法的执行仅提供了对象形状的一部分，但是形状的多个部分被合并以生成整个形状。将输入图像渲染到生成的物体形状上，从而可以重建具有纹理的3D模型。由于可以将多个图像渲染到物体的同一表面上，因此可以提高空间分辨率。由重建的三维模型投影生成的具有纹理的图像可以看作是具有超分辨率的拼接图像。在实验中，我们可以生成4-16倍空间分辨率的超分辨率图像，并提出了一种利用射影几何方法合成运动相机的超分辨率图像的新方法。在该方法中，输入帧之间的逐点对应可以由从射影几何导出的基本矩阵生成。为了使相应的误差最小化，我们还实现了一种优化技术，在该技术中，图像的高频分量被估计用于优化评估。实验结果证明了该方法的有效性。

In this research, we aim to develop a method for synthesizing super-resolved images, that improve spatial resolution of input images by merging multiple input images taken by a moving camera. When a moving camera5takes image sequence, some part of image is overlapped in different frames. In such overlapped part, feature points are detected for obtaining point-by-point correspondence between the frames. This correspondence information provides shape of the object by applying Factorization method. Although the execution of Factorization method only provides a part of the object shape, a number of the part of the shape are merged for generating whole shape. The input images are rendered onto the generated object shape, so that the 3D model with texture can be reconstructed. Since a number of images can be rendered onto the same surface of the object, spatial resolution can be improved. The generated images projected from the reconstructed 3D model with texture can be regarded as mosaic image with super resolution. In the experiment performed in this research project, we can generate super resolved images that have 4-16 times spatial resolution.We also propose a novel method for synthesizing super-resolved images from uncalibrated moving camera by using projective geometry. In this method, point-by-point correspondence between input frames can be generated by fundamental matrices that are derived from the projective geometry. For minimizing the corresponding error, we also implement an optimization technique, in which higher frequency component of the image is estimated for the optimization evaluation. The experimental results demonstrate the efficacy of the method.