An integrated multiple neural network architecture for reading alphanumeric characters in complex scenes
Kong-Wah Wan, Qi Tian, Kah-Chan Low, Soo-Leng Lau, Ho‐Chung Lui · 2002
An integrated multiple neural network (NN) architecture is proposed to handle the automatic extraction and recognition of machine-printed alphanumeric characters in a complex scene. The computational framework is illustrated by our implementation of a multi-expert, multi-level reading system comprising of five NNs for both segmentation and recognition. A window-based alphanumeric NN is trained as a first-pass character-level primary recognizer. A window-based alphabet NN and a window-based digit NN are then used as group-level secondary recognizers. To further boost the recognition accuracy, a Fourier-descriptor-based probabilistic neural network is trained on characters which were misrecognized by the above three networks. In the event of mis-segmentation of merged or damaged characters, a special character horizontal position neural network is used to detect the horizontal positional occurrence of characters. Our system has been applied to the recognition and verification of the 11 machine-printed alphanumeric ID on containers.>