Illumination Direction Estimation And Application Of Surface Texture Without Using Test Image For Training

Texture is one of the basis attributes for the surfaces of the natural objects. In the real life, visual perception helps human beings to realize the world. Recently, work related with 3D surface texture such as data collection, analysis has become an active research topic. Now technology of 3D surface texture has been widely used in the field of computer vision, image processing and computer graphic.Due to the rough surface geometry, 3D surface textures have different appearances under different illumination and view conditions. The study of texture structure and representation seems to be important.This paper describes an approach to estimate the illumination tilt angle of 3D anisotropic surface texture based on Active Basis. The Active Basis model is learned from training images with efficient Haar features by the shared pursuit algorithm. The base histograms of the test image and the training sets are then obtained and employed for estimating the illumination directions by minimizing their Euclidean distances and the entropy differences. Moreover, Mojette transform is introduced to measure texture directions, with which we are able to estimate the illumination direction of a texture excluded in the training set but having similar texture orientations to training samples.The effectiveness of the proposed approach is demonstrated using two data sets. Data setⅠis used to assess how well the method can identify the texture directions and estimate the illumination for the samples of the same class. Data setⅡis used to assess the ability of the proposed method on the different texture classes. Though the data set is limited, the illumination of the test image can be effectively estimated once the texture with similar orientations is found in the database. Differing from the previous work, the illumination direction of a test texture that is not included in the training set still can be estimated. At the end of the paper, the proposed method is tested with an improved 3D reconstruction algorithm. The reconstruction result shows the effectiveness of the proposed method.