Our model here is that a human expert sets down the initial several pixels of an image boundary, and a neural network continues the task by learning the local landscape and continuing through similar image territory as originally identified. One characteristic of neural nets is an adaptability to noise, and thus if the initial image territory is noisy, the network could learn to navigate through it, addressing the first concern above. In addressing the second concern, we note that hole-scene analysis, a straightforward task for a human expert, has proved exceedingly difficult to automate. The expert/network combination we set forward capitalizes on what each does best: the expert to provide global perspective and context, and the network to quickly analyze and work through similar local neighborhoods.
We have focused on neural networks as the learning mechanism due to their very general abilities. In earlier studies, we demonstrated their facility in learning non-linear region discriminations[1].