Ridge Detection Using Artificial Neural Networks

Ridge detection in digital imagery, unlike edge detection, is an area that has received relatively little attention. Ridges are long, narrow structures that are perceptibly brighter than the background. Examples of ridges include roads in aerial images and vessels in medical images. The ability to rapidly and accurately label ridge pixels in a digital image is an important first step in many computer vision applications. Artificial neural networks (ANNs) have proved useful in various pattern recognition tasks. In this work, we investigated the effectiveness of ANNs for ridge detection. We trained a three-layer feed-forward backpropagation network to detect ridges in digital images. The input and the hidden layers had the same number of perceptrons (roughly equal to the maximum ridge width), while the output layer had a single perceptron. The input training set consisted of digital images with strong ridge content (scenes of lightning strikes, containing ridges of varying width and intensity). The input vector was composed of normalized pixel intensity levels, sampled from a small neighborhood about each pixel. The output data set in this supervised training approach was the normalized intensity output of a more traditional ridge detector. The ridge detection results we obtained using ANNs were equal and in some ways superior to the traditional ridge detector. The traditional ridge detector worked well on thinner ridges, but was not as effective in detecting wider ridges. The ANN-based ridge detector appeared to "generalize" well enough to detect ridges of varying width and brightness, giving it a significant advantage over more traditional approaches.