ART-based Neural Networks for Character Recognition

• Main Authors: Jung-Liang Hou, 侯俊良
• Other Authors: Pau-Choo Chung，Yu-Kuen Ho
• Format: Others
• Language: en_US
• Published: 1994
• Online Access: http://ndltd.ncl.edu.tw/handle/52986108434092270999

碩士 === 國立成功大學 === 電機工程研究所 === 82 === As nearest neighbor classifiers and multilayer networks with backpropagation algorithm have been widely used for practical pattern recognition，both types of networks have their superior and unique characteristics．Therefore，it has been proposed to combine the superior coming from these two types of networks， so to enhance the performance of the nearest neighbor classifier through iterative backpropagation training．The network is called nearest neighbor isomophic network (NNIN)．However， as using the NNIN for pattern recognition，one problem still exists, that is the determination of the number of clusters within each class． In practical applications， it is common to find that the patterns of each class distribute differently．Some class has a unique high density distribution． Some class distribute sparely，or separated into a few groups. Hence some way to allow the network automatically construc the required network architecture provides the applications robustness． In this paper，the ART networks are slighty modified to incorporate the Mahalanobis distance in order consider the different pattern deviations in various directions．The modified ART are used in the front layers of the nearest neighbor classifier to robustly construct the optimal cluster neurons．The proposed network was extensively tested by hand-written digits． Simulation results indicate that this network combining the adaptive ART networks， the concept of nearest neighbor classifier andf the modified backpropagation learning rule provides a high accurate classification capability with a fast learning rate．