EDGE DETECTION USING NEURAL NETWORKS

Abstract

Artificial neural networks have been shown to perform well in many image processing applications such as coding, pattern recognition and texture segmentation. In a typical multi-layer model of this class, neurons in each layer are linked by synaptic weights to the layer below. The input image itself is linked to the lowest layer. A two stage encoder-detector network for edge detection is proposed here. The single layer encoder stage compresses data from an input receptive field. It drives a two layer feedforward detector network whose two outputs represent components of an edge vector corresponding to the central pixel of the receptive field. The encoder network is trained in an unsupervised learning mode using the Hebb rule. Training is performed over an input image rich in edges. Competition, which is induced by lateral inhibitions, and negative feedback pull the weight vectors of the neurons apart. The Hebb rule maximizes the output variance of each neuron. The result is that an efficient encoding is performed by the network. The detector network is trained using the supervised back propagation learning algorithm. After training, the detector network is able to successfully map the input code vector from the encoder to an output edge vector. Post-processing of the edge image delivered by this edge detection network converts it into a binary edge image. For intensity step edges, the network performs much better than conventional methods of edge detection such as the Sobel and Prewitt operators, the LoG operator and surface fitting techniques. With Gaussian noise added to the input images, the performance of the network is equivalent to that of the Canny operator, which is considered to be an optimal ed3e detecting filter under these conditions. The ability of the network to learn adaptively from its training images, however, makes it suitable to applications involving other types of edges, including textures.