| This work explores the potential of a new Attributed Vector Quantization (AVQ) method for pattern acquisition and recognition. One of the main contributions of this work is the introduction of a novel way to determine suitable weights for the input vectors such that the resulting codewords form a good representation of the structure of the pattern to be acquired and recognized.; Another main contribution of this work is the design of a codeword selection algorithm that is conducive to pattern acquisition.; To explore and test this approach, two problem domains are chosen: medical imaging and robot vision in manufacturing. The rationale behind the focus on medical imaging is that at the low level, medical images are often noisy and suffer from contrast inhomogeneities. At a higher level, anatomic regions of interest are often non-trivial to recognize even for the medical professional. For these reasons, it is believed that this data poses a significant challenge to this approach and is a useful benchmark. As we shall see, application of this approach to robot vision in parts identification during manufacturing allows the exploration of the potential of this approach to semi-automate the vision system's training process. (Abstract shortened by UMI.)... |
