Abstract

Image processing and computer vision research is frequently difficult to evaluate because it is based on an experimental approach that is inadequately defined or insufficiently detailed. This paper suggests a model-based research paradigm for digital image processing and computer vision. Specifically, we present techniques for generating artificial scenes and simulating imaging systems that form a versatile and precise research test-bed. This approach is flexible, allowing for a variety of scenes and systems; accurate, providing exact control over all variables; and portable, facilitating replication on various computer systems. Two types of computer generated scenes accurately model the important characteristics of real scenes. Fourier scenes are 2-D Fourier series. Techniques for generating ensembles of Fourier scenes allow precise control over important statistical characteristics. Digital scenes are digital images with superresolution—high resolution relative to the sampling rate. Techniques for generating digital scenes provide precise control over spatial structure. The imaging-system simulation software provides a flexible end-to-end imaging environment that can be varied and controlled much more easily than physical systems. The software simulates the important components of the imaging process from the device point spread function, sampling, and noise in image formation to the image reconstruction of the display device.

© 1988 Optical Society of America