| Verification of computer vision theories is facilitated by the development and implementation of computer simulation systems. Computer simulation avoids the necessity of building actual systems; they are fast, flexible, and can be easily duplicated for use by others. Development and implementation of computational models in computer vision are both interesting and challenging. It involves research in diverse areas and requires integration of both science and technology. This dissertation addresses the computer modeling and simulation techniques for two computer vision problems: object recognition and image sensing process. Image sensing process investigates how an image is sensed by specifying the input characteristics of the object and the imaging devices, while object recognition is a high level processing of the sensed image. We present a neural network model to solve the problem of 3-D object identification and pose estimation. The network is divided into two stages, namely Feature Extraction Stage and Feature Detection Stage to extract the feature vectors and to identify the objects, respectively. 3-D moments are used as input feature vectors to the network. Therefore, unoccluded objects are required. We also present a useful computational model to explore the image sensing process. This model decouples the photometric information and the geometric information of objects in the scene. Therefore, it is computationally tractable. Finally, we extend the proposed image sensing model to simulate the formation of moving objects and stereo imaging applications. All the models presented here have been implemented and the implementations are efficient, modular, extensible, and user-friendly so that others can easily reproduce and/or verify their experiments on a broader set of computer vision theories. |
