HASHI: Japanese character recognizer

• Main Authors: David, Michael Angelo E., Malanum, Mark David C., Ongkoa, Lester S., Yalung, Walter T.
• Format: text
• Language: English
• Published: Animo Repository 1995
• Subjects: Japanese character sets (Data processing); Japanese language > Writing; Imaging systems; Computer network architectures; Neural circuitry; Computer Sciences
• Online Access: https://animorepository.dlsu.edu.ph/etd_bachelors/14013
• Institution: De La Salle University

Hashi is a handwritten Japanese character recognizer. It uses handwritten OCR techniques and the neural network structure of Backpropagation to attain at least 90 percent recognition accuracy. During the course of its development, two approaches were arrived at and developed to successfully accomplish its objectives. The first used an image-based approach in recognition. This approach used a 16 x 16 single hidden layer BP network. A second approach uses stroke capture and recognition techniques as preprocessing for the neural network. The first approach required the creation of 40 image sets. Training was limited to at most 20 training sets which were taken from the pool of 40 image sets. Five different scenarios were run, each scenario using a different configuration based on the number of units in the hidden layer, number of training sets, and output representation. Each scenario was tested using 4 test sets which also came from the pool of 40 image sets. The stroke-based implementation involved the creation of 20 training sets and 2 test sets. Recognition accuracy was based on the system's performance on the test sets. The first approach attained a recognition accuracy of 88.24% while the second achieved 95.59% recognition. Thus, the second approach was chosen for the system."