Tadaaki Kitamura

Original Contribution: Extracting algorithms from pattern classification neural networks

It has been shown that a pattern classification neural network can be synthesized in three layers (one hidden layer) if each class is separated into a single convex region by hyperplanes and in four layers if not. Based upon this, extraction of a pattern classification algorithm from a network is discussed, namely, first separation hyperplanes are extracted from a network trained by the backpropagation algorithm, and then the weights corresponding to the separation hyperplanes and those connected to the output neurons are successively tuned to improve the generalization ability of the network. Classification is made according to which sides of the hyperplanes a test datum (or the second layer output of a test datum for four layers) is on. Finally, weights of the neural networks, generated by the backpropagation algorithm for a number recognition system, are tuned and classification algorithms extracted.

Neural networks as a tool to generate pattern classification algorithms

Extraction of algorithms from pattern classification neural networks is discussed. It is shown that input-to-hidden weights for three layers, or otherwise, hidden-to-hidden weights for four layers are determined so that corresponding hyperplanes separate all the training data in a class from the remaining classes. The extraction of a classification algorithm from a trained network is discussed, and its generalization ability is compared with that of the original network. Weights of the neural networks, generated by the backpropagation algorithm for number recognition, are tuned, and classification algorithms are demonstrated to be extracted.<<ETX>>

VEHICLE EXTRACTION USING SPATIAL DIFFERENTIATION AND SUBTRACTION

In this paper we propose a method, which is executable on todays commercially available image processors, for extracting vehicles. The method is summarized as follows: (1) First, the method generates an initial background image by extracting a background region using logical OR operations among sequentially inputted and subtracted images. (2) Second, on the basis of the change ratio of the brightness of visible road surface, the method dynamically updates the background image by estimating brightness of road surface regions hidden by vehicles. (3) Finally, the method extracts the vehicles using images subtracted by the background image and spatially differentiated images. In preliminary experiments we obtained a good background image by updating it with moving shadows, even under bad conditions at dusk. Furthermore, we had good accuracy in traffic flow measurements, with more than a 95% recognition rate of vehicles and only 3% overdetection. We confirmed the effectiveness of the method by the above evaluation experiments.