Image synthesis methodology for algorithm testing and vision system design

Although it is well-recognized and widely accepted that vision adds considerable flexibility, and it has also been shown that numerical simulation can aid in image understanding and vision system design (significantly reducing the engineering time to design and implement such systems), the utilization of image synthesis as an aid in algorithm and system design still remains a largely underexplored area. In machine vision applications, accuracy of the image generally outweighs image appearance. Unfortunately, the focus of most commercially available simulation methods is on photorealistic image synthesis; this is insufficient to design vision systems or evaluate and compare image-processing algorithms for part-presentation tasks: physically accurate, rather than photo-realistic, synthesis methods are necessary to sufficiently simulate captured image grey-scale values. This paper presents a methodology to generate physically accurate synthetic images efficiently in order to provide an accurate, flexible and practical means of evaluating the performance of image-processing algorithms for numerous hardware/software configuration combinations and a wide range of parts. While the synthesis methodology cannot fully compensate for the real environment, it can be used efficiently to study the effects of vision system design parameters on image accuracy. This provides an insight into the efficacy of the design and the ability of suggested image-processing algorithms to perform adequately for specific applications; furthermore, it may provide a means for correcting apparent errors in image-processing results.