| Summary: | 碩士 === 國立臺灣海洋大學 === 電機工程學系 === 92 === Segmentation tasks often require parallel processing capability in order to handle massive computation load. This thesis presents a novel neural-based approach to a fast and efficient solution for image segmentation. The approach incorporates two neural network models, namely Growing Cell Structures (GCS) and Probabilistic Neural Networks (PNN). Training algorithm consists of initial (pre-process) phase and learning phase. In the initial phase, GCS divides gray level of input image into m (m>1) clusters and accordingly the input image is roughly segmented in the sense that the initial probability of an arbitrary pixel (i, j) belonging to the kth cluster is assigned by GCS. In addition, a new gradient operator called selective average operator (SAO) is developed to extract prominent edges that is useful for the subsequent labeling operations on the rest of pixels in the input image. In the learning phase, the value of the probability of an arbitrary pixel (i, j) belonging to the kth cluster PNN is iteratively adjusted by referring to their respective immediate neighboring pixels until convergence. Experimental results have verified that our approach can efficiently achieve accurate segmentation.
|
|---|
