The observed colors of the object are influenced by many factors of the physical world, including the shape and material of the object, the illumination, the position of the viewer, etc. Estimating and processing the shape, color, reflectance and shading the observed image are a fundamental problem in computer vision and computer graphics. In this thesis, we proposed a model for intrinsic decomposition and discussed two related topics. The main contributions of this thesis are as follows: 1. We proposed an intrinsic decomposition method which used the single image and noisy shape observation data as the input data. This method can decompose the input data into the refined shape data, reflectance image and the shading image. As this problem is underconstrained, the shape smoothness constraint, reflectance smoothness constraint and the illumination prior are used to alleviate the ambiguity of the decomposition. Compared to the related work, our method can achieve more accurate shape estimation result by using the shape observation data. Our problem is a nonlinear least squares problem and solve it using sparse Levenberg-Marquardt algorithm and the coarse-to-fine optimization strategy. 2. We proposed an edge-preserving smoothing method for the depth map. This method uses the second order smoothness prior to avoid the staircase effect which exists in recent methods. It uses the binary line process variables to overcome the over-smoothing problem caused by the high order smoothness prior. Therefore, the salient edges can be sharply preserved. As the binary variables make the original problem hard to solve, a practical optimization strategy is used to achieve the approximate solution. We demonstrate the effectiveness of our method in the applications such as depth map smoothing, cartoon image denoising, non-photorealistic rendering and image detail magnification. 3. We proposed a consistent segmentation based local color correction method for coarsely registered images. A consistent segmentation method is proposed to alleviate the negative effect rising from inaccurate registration. The region confidences and the bilateral-filter-like color influence maps are used to improve the color correction result. The experiment shows the proposed method achieves improved color correction results compared with related work. Finally, the research results are summarized, and the future work is discussed.
修改评论