Neural Networks on Shape Recognition

• Main Authors: Shen Yu-Sen, 沈玉升
• Other Authors: Chou Wen-Kuang
• Format: Others
• Language: zh-TW
• Published: 1994
• Online Access: http://ndltd.ncl.edu.tw/handle/43141496575791892099

碩士 === 靜宜大學 === 管理科學研究所 === 82 === Pattern recognition has been a well-known complicated problem. Although, numerous efforts have been made based on traditional computer, they still suffered by the time- consumed procedure. By the invent of neural networks, which is an architecture mimicking the spirt of human brain, the research of pattern recognition is promoted based on the new technology. In this research, a hybrid neural system is proposed to attack shape recognition with invariant for rotation, scaling and distortion. In the system, some efficient preprocess are proposed to extract shape features. Based up on those features, the most popular neural networks, back- propagation (BP), is used to learn and recall. The hybrid neural system has been implemented on C language. Also, The benchmark of 2-D plane shape is selected to test the hybrid neural system. The simulation results show that the proposed system achieve 97% recognition rate even thought the test patterns are scaled, rotated and distorted. Since the proposed system is powerful and efficient for recognition of object contour, it has very high potential for real-time system. In other word, it can be applied to objective searching, Chinese recognition, character recognition, and so on. On the information management point of view, the proposed system has achieved a significant contribution on office automation.