<title>Neural network image synthesis method for use in computer vision</title>

Abstract

ABSTRACT This paper describes an image synthesis method with neural network and demonstrates its potential applicationsin computer vision. Various 3-D object models with freeform surfaces and different texture can be generated using this method by training the network based on the main geometric feature of objects. It is also very flexible andcapable of providing different photometric geometry which are desired in many computer vision applications. Experimental results indicated that the method can be used as a modeling tool in computer vision. 1. INTRODUCTIONIn recent years, synthetic images are not only used in computer graphics, visualization and computer aideddesign, but also have been widely used in computer vision research, especially for the three dimensional object recognition and localization, such as model-building, algorithm evaluation, shading computation, model selection, andshape estimation [1,2,3,4,5].The recognition and localization of 3-D objects in a scene is a key problem for the research in computer vision,especially for robot vision. There are several approaches have been developed for recognition of 3-D objects. Wallaceet a!. [6] described an appmach in which a model of each object in the model data-base is matched against the scene