3D Reconstruction Based Image Compositing

As portable equipment, especially smart phones become more and more popular, taking photos and beautifying them afterwards have become a common way of recording beauty in our lives. And along with the wide application of 3D technology, for example, 3D movies and naked eye 3D screen, the meaning of research in 3D image processing area has been more and more important.Based on this, this paper proposes an image compositing technology based on 3D reconstruction. This method can restore a 3D scene easily from two-dimensional photographs, which were obtained from portable equipment. And this method also uses information and result producted by 3D reconstruction to guide the work of image compositing afterwards. The ultimate goal is to strengthen the fidelity of the composite image.In the stage of recovering three-dimensional scene, we propose a three-dimensional reconstruction method based on image segmentation and binocular vision. We propose a method to filter discrete error points in the initial disparity map in order to improve the accuracy of result. We propose a method to generate the corresponding topology of block surfaces. For the merging process for block surfaces, we use union-find data structure to reduce the complexity.In dealing with the image compositing, taking into account that human is more sensitive to light and shading in pictures, and in order to maximize the fidelity of composite image, we studied the issue of light consistency and studied how to estimate light direction in pictures.Since photo taken from a real environment is often more complex, the applicability of a single algorithm for light direction estimation is limited. In order to enhance the applicability of the light source direction estimation algorithm, this paper proposes a method for seeking a simple local image region, which uses information obtained from 3D reconstruction and is in favor of light direction estimation. Then we use this local image region as input for light direction estimation algorithm. In this process, it needs to segment the image, but traditional image segmentation algorithms often divide a region of the same color into pieces, just due to the changes in illumination, which will affect the RGB values in image. To address this problem, we propose an image segmentation algorithm, which can effectively reduce the interference caused by illumination changes. Then, this paper estimates the approximate location of the light source using the least squares method to estimate the light direction for each point. Finally, the algorithm is tested, and the results are given.