|Abstract: ||此計畫將為一系列研究發展的開端,其目的在落實在中文環境下之電腦化服務的提供。在此計畫裡,我們選擇了銀行系統裡的提存款單的手寫款額數字之辨識為研究主題。換句話說,我們想研究發展出一電腦系統來摹擬人的能夠幾乎百分之百地辨識手寫的以下十七個中文數字:零、壹、貳、參、肆、伍、陸、柒、捌、玖、拾、佰、仟、萬、億、元和角。我們選擇的執行工具是簡單的類神經網路(Artificial neural networks)系統---例如線性結合系統(Linear associator)、Kohonen網路系統和Hopfield網路系統。在此計畫裡,我們想研究了解如此簡單的類神經網路系統是否能達到以下之要求:(1)在系統學習時僅需要少量的範例,和(2)只要考慮網路的規模(Size)大小是否合適。摹擬實驗的結果顯示,此十七個中文數字的筆劃是相當地複雜。當使用的類神經網路的規模大小適當時,只要手寫數字與其範例之間的差距(雜訊量,Noise level)不超過45%,其辨識率可達百分之百。可是有很多實際的手寫數字的雜訊量是超過50%。由此計畫,我們了解了簡單的類神經網路系統是不能滿足我們的期望。|
This project is the initialization of a series of the implementations of providing the computerized service. In particular, the goal of this series is to realize the providing the computerized service in the Chinese environment. In the Chinese banking system, there are 17 numeral characters that are written necessarily to display formally the amount of money stated in the deposit (or withdraw) slip: 零, 壹, 貳, 參, 肆, 伍, 陸, 柒, 捌, 玖, 拾, 佰, 仟, 萬, 億, 元 and 角. We have studied the possibility of developing an intelligent system that can recognize nearly 100% perfectly the statements stated in the deposit (or withdraw) slips. For accomplishing this, we may develop a neural networks system that can recognize nearly 100% perfectly the handwritings of these 17 characters. In achieving this, providing the computerized service in the Chinese environment goes further one big step. In essential, this task belongs to the category of setting up an associative memory system. Thus, several kinds of the associative memory systems of the artificial neural networks, included the linear associator, the Kohonen network, and the Hopfield network, were studied in depth here. Based upon the analysis, we have made the following propositions to render the system able to mimic the intelligence of human beings. We propose that there should be mere few training exemplars corresponding to the standard form and the variability of the 17 Chinese characters. Furthermore, we also propose that whether the size of the network is proper is the only consideration here. The simulation with mere 17 exemplars corresponding to the standard forms of these 17 Chinese characters is used to test the conjecture. The simulation's results show that, as we expected, the graphs of these Chinese characters are complex and distinct and there is a proper size of the network. Then, the real, reasonable handwritings are tested. Unfortunately, the testing results are negative. The reason of the negative testing results in the real, reasonable handwriting is that the difference between the real, reasonable handwritings and their corresponding exemplars are sometimes larger than 50% noise level. Just as we expected from the beginning of this project, it is difficult to achieve the goal of developing a system which can mimic the human beings' intelligence to recognize nearly 100% perfectly the handwritings of these 17 Chinese characters. However, the advantages and the limits of these neural networks systems, theoretical and practical, had been studied here. The knowledge we obtained from this project can, and will, be adopted in the further projects and researches.