| Summary: | This thesis presents a hybrid system for embedded machine vision combining programmable hardware for the IP tasks and a digital implementation of an Artificial Neural Network (ANN) for the pattern recognition and classification tasks. This thesis also researches techniques for the efficient implementation of IP algorithms into programmable hardware, and introduces a library of hardware implementations of common IP algorithms that has been compiled to facilitate the development of embedded vision applications on the hybrid system. A common problem encountered during the hardware implementation of algorithms is how can the evaluation of similarity between two edge images be performed. This thesis includes the development of a novel metric that has proved to be more robust and accurate, in the similarity measurement, than existing metrics. Finally, a protype of the hybrid system is presented, and the applicability of the hybrid design is demonstrated with two vision applications introducing new ways to combine IP and ANNs: microbiological rapid imaging and Road Sign Recognition (RSR). In particular, the RSR application demonstrates the ability of the prototype system to achieve a real-time performance, as well as its ability to dynamically adapt the values of the neural array during execution. These results indicate the suitability of the hybrid system in a wide number of vision applications ranging from robot navigation to security, as well as its potential use in fields other than machine vision.
|
|---|
