Abstract
To recover 3-D structure from a shaded and textural surface image involving textures, neither the Shape-from-shading nor the Shape-from-texture analysis is enough, because both radiance and texture information coexist within the scene surface. A new 3-D texture model is developed by considering the scene image as the superposition of a smooth shaded image and a random texture image. To describe the random part, the orthographical projection is adapted to take care of the non-isotropic distribution function of the intensity due to the slant and tilt of a 3-D textures surface, and the Fractional Differencing Periodic (FDP) model is chosen to describe the random texture, because this model is able to simultaneously represent the coarseness and the pattern of the 3-D texture surface, and enough flexible to synthesize both long-term and short-term correlation structures of random texture. Since the object is described by the model involving several free parameters and the values of these parameters are determined directly from its projected image, it is possible to extract 3-D information and texture pattern directly from the image without any preprocessing. Thus, the cumulative error obtained from each pre-processing can be minimized. For estimating the parameters, a hybrid method which uses both the least square and the maximum likelihood estimates is applied and the estimation of parameters and the synthesis are done in frequency domain. Among the texture pattern features which can be obtained from a single surface image, Fractal scaling parameter plays a major role for classifying and/or segmenting the different texture patterns tilted and slanted due to the 3-dimensional rotation, because of its rotational and scaling invariant properties. Also, since the Fractal scaling factor represents the coarseness of the surface, each texture pattern has its own Fractal scale value, and particularly at the boundary between the different textures, it has relatively higher value to the one within a same texture. Based on these facts, a new classification method and a segmentation scheme for the 3-D rotated texture patterns are developed
Date of this Version
1-1-1991
Comments
U.S. Army Research Office Contract DAAL03-89K-0032