Illumination Estimation of 3D Surface Texture Based on Active Basis Junyu Dong Liyuan Su Department of Computer Science Ocean University of China Qingdao, China Email: {dongjunyu,suliyuan}@ouc.edu.cn Yuanxu Duan Alcatel-Lucent R&D Qingdao, China Email: duanyx1999@163.com Abstract—This paper describes an approach to estimate illumination directions of 3D surface texture based on Active Basis. Instead of applying Gabor wavelet transform to extract texture features, we represent our texture features with a simple Haar feature to improve efficiency. The Active Basis model can be learned from training image patches by the shared pursuit algorithm. The base histogram can then be obtained based on each model. We measure the illumination directions by minimizing the Euclidean distance and the entropy difference of base histograms between the test image and the training sets. Experimental results demonstrate the effectiveness and accuracy of the proposed approach. Keywords-3D Surface Texture; Illumination Estimation; Ac- tive Basis; Feature extraction; I. I NTRODUCTION Texture analysis plays an important role in computer vision and computer graphics. Real world textures are seldom flat and may consist of rough surface geometry and various reflectance properties. These textures can be called 3D surface texture. Changes in incident illumination angles upon a 3D surface texture can significantly affect its appearance [1]. Estimating illumination directions of 3D surface texture is therefore important in many tasks in computer vision and realistic graphics rendering. Early studies on illumination direction estimation [2][3] are based on techniques from Shape from Shading. There are many constraints have to be imposed in order to find a reasonable solution. Recently, Koenderink and Pont [4] de- velop a theory based on second order statistics for estimating the illuminant’s azimuth from a single image formed under the Lambertian model. Their algorithm is restricted by the fact that they require the surface is an isotropic, Gaussian random surface with constant albedo. Varma and Zisserman [5] then extend the method by Koenderink and Pont to deal with cases where the albedo is spatially varying. Although it can recover the azimuthal angle much more accurately, this estimation mechanism fails on strongly anisotropic textures. This paper proposes a method to estimate the azimuthal angle of illuminant for 3D surface textures. The method con- sists of two phases. Firstly 3D surface texture classification is performed to determine texture classes by using wavelet- packet features [6] and support vector machines [7]. Then a new method called Active Basis is employed to reflect illumination direction features. The Active Basis model can be learned from training image patches by the shared pursuit algorithm, which was originally developed by Wu et al. in [8]. Finally, the histograms of Active Basis are compared to estimate the illumination direction of surface textures. The basic idea of this paper is based on the fact that illumination from certain direction can cause the surface bumps towards the light source become brighter and exhibit ”edge” property (Figure 2). Apparently, when the texture is anisotropic, due to the high response between strong sur- face texture structures and illumination, ”edge” information corresponding to the illumination direction can be easily modeled by Active Basis. Differing from the work in [5], our method can correctly deal with the anisotropic texture surfaces and also performs well on some isotropic textures. The rest of the paper is organized as follows. Section II describes 3D surface texture classification. In section III we present details on learning of the Active Basis model and estimation of illumination directions. Results produced by our method are presented in section IV. In section V we draw our conclusions. II. THREE- DIMENSIONAL SURFACE TEXTURE CLASSIFICATION We first determine the texture classes using wavelet-based features and support vector machines. A. Feature extraction based on wavelet-packet decomposi- tion Since feature extraction based on wavelet transform can provide more details and can be effectively used in distin- guishing textures classes, we use wavelet-packet decompo- sition to represent characteristics of different textures. We apply a two-scale and full Daubechies wavelet packet decomposition on each texture image. Since each wavelet packet tree has large number of possible decomposition, an entropy-based cost function is then employed to decide the optimal of decomposition. We can generate the feature vector: F = {m 1 ,v 1 ,m 2 ,v 2 ,m 3 ,v 3 ,m 4 ,v 4 ,m 5 ,v 5 ,m 6 ,v 6 ,m 7 ,v 7 }, where m and v are the means and variances of selected wavelet sub band coefficients (See [6] for more details). 2010 International Conference on Pattern Recognition 1051-4651/10 $26.00 © 2010 IEEE DOI 10.1109/ICPR.2010.219 874 2010 International Conference on Pattern Recognition 1051-4651/10 $26.00 © 2010 IEEE DOI 10.1109/ICPR.2010.219 874 2010 International Conference on Pattern Recognition 1051-4651/10 $26.00 © 2010 IEEE DOI 10.1109/ICPR.2010.219 870 2010 International Conference on Pattern Recognition 1051-4651/10 $26.00 © 2010 IEEE DOI 10.1109/ICPR.2010.219 870 2010 International Conference on Pattern Recognition 1051-4651/10 $26.00 © 2010 IEEE DOI 10.1109/ICPR.2010.219 870