A Hopfield neural network for adaptive image segmentation: An active surface paradigm

This paper presents an adaptive thresholding technique for separating objects from noisy and non-uniformly illuminated images. The construction of the threshold surface is formulated as an active surface optimization problem, which is then solved by a Hopfield neural network. We proposed four constraints which ensure the active threshold surface to conform with the underlying image topography. Compared with Yanowitz and Bruckstein's method, this method produces superior segmentations particularly when the edge segments are sparsely distributed in the image and under non-uniform illuminations. Using three types of artificial and real images, we show that this method converges faster and produces better segmentations compared with previous interpolation-based adaptive thresholding techniques.