Photometric Stereo Methods For Object Surfaces With Various Reflectances

Photometric stereo has been widely employed for surface reconstruction since the pioneer work by Woodham. In photometric stereo, an object surface is illuminated by multiple directional light sources and its images are captured using a camera with fixed viewpoint. The works dealing with object surfaces with various Bidirectional Reflectance Distribution Functions (BRDFs) are mainly categorized as the Lambertian-Enforced ones and BRDF-Based ones. Though recent works have achieved good performance, there are still problems to be resolved, namely, (1) improving the normal estimation accuracy, (2) reducing number of light directions, and (3) improving high efficiency. This thesis investigates these problems. The main contributions are summarized as follows:1. A photometric stereo method is proposed to improve the collinear light source based photometric stereo. The highlight detection is converted to a pattern classification problem using multiple classifiers and multiple training data. An î–¢ norm regularization term is further employed in final normal estimation to improve its robustness to noise and outliers.2. A photometric stereo method based on the general characteristics of object reflection is introduced. The reflection is modeled by three characteristics, i.e., the smooth variation of diffuse reflection, spatial concentration of specular reflection, and low-intensity nature of shadow. A graph of light directions is constructed to model these characteristics as the summation of local variation, group sparse and weighted l1 norm term, respectively. The optimization problem is cast as a second order cone programming problem. Experimental results validate that the method is effective for both isotropic and antisotropic materials.3. A photometric stereo method using kernel regression is presented. The reflection is implicitly modeled using kernel regression with variable parameter. The problem is transformed to an eigen decomposition problem, whose computational cost is only the inverse of a Gram matrix and eigenvector computation on a 3x3 matrix. The best parameter of the variable kernel is determined via leave-one-out cross-validation, which is further accelerated by fast matrix computation and proper normal initialization. Experimental results validate the accuracy and efficiency of the method in normal estimation.It is worth noting that, different to the current BRDF-Based methods, the object reflectance is simplified as a function of single variable, i.e., light direction, in the 2nd and 3rd works. This treatment reduces the dimensionality of reflection function, and consequently makes the proposed works perform quite well on object surface with various reflectances.