A methodology for the development of machine vision algorithms through the use of human visual models

In many operations the ability of a machine to "see" is what will determine its effectiveness in its particular domain of operation. For example, in a bin picking problem the ability of the sensing system of a robot to determine the position and orientation of the individual parts will ultimately determine the system's success or failure.; Most systems that require this level of sensing, utilize machine vision in which computers are integrated with image acquisition devices to provide the information required for guidance; as would be needed in a feedback loop for example. The development of algorithms that allow these computers to accomplish the image interpretation has turned out to be less than trivial. This is especially true in the area of natural products such as, meat products, fruit or textile; where, because of their natural variability the ability to develop machine vision algorithms to automatically inspect these products reliably has been problematic.; The goal of this thesis is to attempt to determine a methodology for the integration and streamlining of the process of algorithm development so as to be able to more efficiently develop effective and robust algorithms for this class of problems. Humans, are currently still the best available solutions to these problems. This thesis will examine an approach towards the development of machine vision algorithms using the primate visual system as a model.; The approach taken in this work defines three levels of processing for the visual signal these are sensing, ecoding/transfer, and classification. In particular we examine the processes of encoding/transfer derived from the results of research in the area of human/primate biological visual processing and their representations. We focus on the use of the receptive field mechanisms that are commonly observed in the human visual system and their processing of contrast in the scenes. We also show that features derived from the responses of these mechanisms are useful for image classification.; Algorithms for implementing these operations are developed using the technique and demonstrated. The other aspect of the approach provides for user guidance by allowing an expert to teach the system by identifying things that are of interest in a particular scene. We then demonstrate development of solutions to three inspection problems using the approach.