The Study of COmmercialization for Multi-country False Currency Recognition System

[[abstract]]多國真偽鈔辨識系統商品化之研究 摘 要 類神經網路 (Artificial Neural Network ) 這個領域和其它的科學相比算是相當年輕的，儘管如此，在最近幾年裡，它卻一直是研究工作者鑽研的對象之一。因為，我們渴望能夠發展出比目前的機器更聰明、靈巧的機器來服務人類。因此、在許多領域的研究者，像生物學、計算機科學、電機工程、心理學、數學、哲學、機械．．．等等，都同心栛力地從各個不同的角度來探討類神經網路。目前，類神經網路大都被應用在作圖形識別、分類、控制．．．等研究上，至於商業化的產品，仍未多見。 本篇論文主要根據的理論是，一般各國所用以製造鈔票的紙張，乃是一種特殊製造的紙質，在各個國家中都是嚴格管制，且各個國家所用以印刷鈔票的油墨顏料都是具有一定相對比例；綜合以上二種製鈔的關鍵技術，都是偽鈔集團所無法突破的，故本論文基於此一特點結合類神經網路，發展出一套適用於多國真偽鈔辨識的系統。 本篇論文將提出運用特殊光學影像擷取技術、數位影像處理 (Digital Image Processing) 及結合可塑性認知網路 (Plastic Perception Neural Network) 來實現多國真偽鈔辨識系統的商用化研究。文中所運用的可塑性認知網路的架構是針對類神經網路中回傳式神經網路 (Back-propagation Neural Network) 的學習法則以及其網路架構做改進，來克服傳統的回傳式神經網路在訓練時學習速度太慢、不易收斂、及當刪除或加入新的訓練樣本 (patterns) 時所必須全盤重新訓練等問題。這些缺點成為實現回傳類神經網路系統的嚴重障礙。為了克服這樣的障礙，我們結合了平行分散處理的觀念，對回傳式神經網路的網路架構做一改進，以增快其學習速度，並克服全盤重新訓練的問題。 本文尚包括特殊的光學影像擷取方法、影像物件擷取、雜訊去除以及影像物件特徵值的抽取；特殊的光學式影像擷取方法有助於能夠順利的擷取下隱藏於鈔票中的特殊特徵，例如：防偽線，浮水印，雙面套印．．．等；另一般的鈔票影像辨識乃是採用整體辨識的方法，本系統則採用將鈔票中主要特徵影像，切割成數個特徵物件，再對其做適切的特徵值抽取來加以辨識，本文所採用的特徵向量是結合白色跑長碼（White Run-Length ）與圖素密度（Pixel Density ），可以分別表現出物件的結構性與整體性，以達到高準確的辨識率。[[abstract]]The Artificial Neural Network (ANN) compares younger with the other sciences. In spite of this, it is worth researching for research workers since our thirst for more clever machines. Therefore, the researchers in different fields such as biology, computer science, electrical engineering, psychology, mathematics, philosophy and mechanical engineering etc., are exploring the ANN in different points of view. For the moment, ANN is generally applying to graph recognition, classification, and control. It is seldom applying for commercial manufactures. The foundation of theory in this thesis is that banknotes are made up of particular papers which are close confinement, and the printing inks and pigments used to print banknotes is a fixed ratio in every countries. Above-mentioned two key points aren't surmounted for false currency groups. Therefore, this thesis combines the two key points and ANN to develop a system suited for multi-country false currency recognition. This thesis will provide techniques as Particular Optics Image Capturing, Digital Image Processing and Plastic Perception Neural Network for implementation of multi-country false currency recognition system in commercial manufactures. The framework of Plastic Perception Neural Network in this thesis is the improvement on the learning rule in Back-propagation Neural Network and network framework of ANN, it solves problems as slow learning, difficult convergence, and overall retraining when you delete patterns or add new ones. To overcome these barriers of the learning rule in Back-propagation Neural Network, We combine the concepts of parallel and distributed process to improve the network framework of Back-propagation Neural Network. For this reason, I increase learning speed and solve problem of retraining. This thesis involves Particular Optics Image Capturing, Object Image Capturing, Noise Suppression and Eigenvalue of Object Image Capturing. Particular Optics Image Capturing is contributive to capture special characteristics hid behind the bank notes such as Metallic Thread，Watermark，Overptints etc.. On the side, The general recognition of the bank note is introduced into the whole recognition, This system takes some characteristic segments separated from principal characteristic images of the bank note, and captures eigenvalue of the segments to recognize. The characteristic vectors combine White Run-Length and Pixel Density to display apart from the construction and macrocosm of Object in order to have high accuracy of recognition