The developers of previous computer graphic image rendering algorithms have not accounted for the fact that synthesized pictures are ultimately intended to be viewed by a human observer. This has led to the creation of algorithms that can waste large amounts of time refining areas of an image that are already visually acceptable, while neglecting regions of an image containing perceptible artifacts. The human visual system has a varying acuity for error that depends on the context in which the error is viewed. Exploiting the subjective nature of perceptual response holds the key to improving the performance of image synthesis algorithms.

This dissertation presents a new perceptually based adaptive sampling algo­rithm. This algorithm makes subjective quality assessments during the progression of a rendering algorithm. These assessments are used to focus the effort of the rendering algorithm on the regions of the image containing the most perceptible artifacts. In this manner, images of a given visual quality are produced faster than is possible with existing image synthesis techniques. The new algorithm also allows the user to select a perceptual quality for the output image. This feature eliminates the guesswork involved in halting a rendering and allows the production of visually consistent results.

This work includes a number of important and novel contributions. The first is the development of a new and comprehensive error metric for Monte Carlo ray tracing. This metric is used to characterize and control the objective accuracy of a rendered image. This metric is additionally employed to determine the optimum number of rays to spawn from each surface intersection. The next major contri­bution is the design of a new high speed, color visual difference predictor. This predictor is capable of rapidly assessing the perceptual impact of objective differ­ences between two color images. The third contribution is the design of a frequency based adaptive sampling algorithm. This algorithm synthesizes images directly into the frequency domain. This permits the use of a simple, frequency dependent image quality metric to control the placement of samples. The final and most important contribution is the development of a second generation, perceptually based adaptive sampling algorithm. This algorithm employs the high speed, color visual difference predictor to control image sampling in accordance with the perceptibility of error in the reconstructed image. This technique is shown to improve the performance of a realistic image synthesis algorithm.