Edge detection through cooperation and competition

The system presented is an edge detector based on a simplified vertebrate retina model and on a constraint-satisfaction neural network implementation. The objective of this system is to produce long image-correlated edge segments. From a functional point of view a simplified vertebrate retina can be seen as a three-layer network. Light is received and encoded by the first layer, the image features are extracted by the second layer, and motion analysis is performed by the last layer. In the first layer, the image intensity gradient magnitude and direction evaluation are performed by the Sobel operator over the smoothed image. In the second layer, a competition and cooperation process is defined and executed in order to suppress noise effects and determine the relevant features. The feature extraction task model is composed of two synchronous subtasks, and is implemented by a constraint-satisfaction neural network. The model was simulated on a SUN-SPARC with four kinds of images. The edge segments generated by the competition and cooperation process are longer and thinner than the results of traditional edge detectors