Edge representation and recognition using neural networks

In this paper, we propose a new approach to represent and recognize edges of objects using multilayer feedforward neural networks. First, we will show how an edge of an object can be represented by neural networks. This is accomplished by generating two classes consisting of samples that lie on each side of the edge and then by training a neural network to classify the two classes. If the training is successfully accomplished, the resulting neural network will have a decision boundary that matches the edge we want to represent. Second, we will propose a matching algorithm that identifies an arbitrarily rotated and shifted edge. The matching algorithm uses a gradient descent algorithm. The proposed algorithm can be used in the area of object representation and recognition. In addition, we will investigate the relationship between the number of hidden neurons and complexity of edges.