Research And Implementation Of SFS Problem Under Probability Model

Shape from shading is a basic and widely used algorithm in surface reconstruction work and it attracts researchers' attention for the algorithm can reconstruct the object's shape only based on one image.For a single synthetic image under experimental conditions, restoring its coordinates of surface by the brightness of the image pixel is an ill-posed problem.For real images, the complexity of its lighting environment and surface reflectancewill bring a variety of noise to the reconstruction work.This paper proposes an improved shape from shading algorithm for both synthetic and real images. Good results have been achieved.The paper uses an improved propagation shape from shading algorithm to reconstruct the synthetic image. Because of the large error which may be caused by propagation method, this paper uses a framework based onprobabilistic method to optimize the reconstruction result. The framework combines different prior-knowledge including a silhouetteprior-knowledge together to constrain the reconstruction process.For real image, due to the limitation of its surface reflection properties and unknown sources, this paper chooses a human-face image to impose reconstruction work. On one hand, a human face's shapeis regular. On the other hand the shape of the organs on human's face is very complex. We use a CANDIDE-3 model to fix the reconstructed result. The algorithm uses a linear SFS method to apply the first step of the reconstruction work. Then we use wavelet transform method to get the height information in the frequency domain of the result of first step. At last, a probabilistic framework is proposed to impose the prior knowledge to our coarse reconstruction result. The process is done in the frequency domain and the final result should be obtained by inverse wavelet transform.The experiment chooses four synthetic images and one human face image which is captured in certain conditions. For synthetic images we use ground-truth height to calculate the MSE of reconstruction result. The experiment result implies that our method is efficient and accurate.