Container Number Recognition System Using Artificial Neural Network

[[abstract]]本論文將提供一種以類神經網路及數位影像處理技術為主之貨櫃編號辨識系統，其係於貨櫃車進入場區時，由數位照相機攝取貨櫃編號之數位影像資料，其中貨櫃編號影像經前處理去除雜訊、字元切割、正規化及特徵抽取單元擷取特徵向量值後，作為類神經網路學習與辨識單元之輸入資料，以於辨識過程中做類神經網路之參數，經運算後得到正確之辨識結果，使之可自動登錄貨櫃進出場區之情形，以達節省人力、加速貨櫃管理及自動化之目的。 運用可塑性認知網路（Plastic Perception Neural Network）來完成貨櫃編號辨識系統。文中所運用的可塑性認知網路的架構是針對類神經網路(Artificial Neural Networks)中倒傳遞類神經網路(Back-propagation Neural Network)的學習法則以及其網路架構做改進，來克服傳統的倒傳遞類神經網路在學習時學習速度太慢、不易收斂、及當刪除或加入新的學習樣本(patterns)時必須全盤重新學習等問題。這些缺點成為實現即時的倒傳遞類神經網路系統的嚴重障礙。為了克服這樣的障礙，我們結合平行分散處理的觀念，對倒傳遞類神經網路的網路架構做一改進，以增快其學習速度，並克服全盤重新學習的問題。本文尚包括文字切割、雜訊去除，以及字型特徵值的抽取，適切的字型特徵值抽取有助於字型的辨識，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），可以分別表現出字型的結構性與整體性，以達到高準確的辨識率。[[abstract]]This thesis brings up the implementation of container number recognition system by the neural network and image processing. As the cargo truck is approaching the digital camera, it will capture the image of the container number. Next, noises will be eliminated from the image and character segmentation will be performed respectively. After the image is normalized and extraction of feature is performed, the feature vectors are collected. All collected parameters are now ready to be fed into the container number recognition system for further calculation. The calculation’s results will provide the computer an accurate image of the container number, and the computer will be able to automatically keep track of the container. With this system available, less labor is required, and people will be able to speed up the overall efficiency of the container tracking management. We will also be able to receive the benefit of modern automation. The application of Plastic Perception Neural Network (PPNN), is utilized to distinguish container number. In order to take full advantages of PPNN, the thesis proposed the improvement of learning algorithm and network structure of Back-Propagation Neural Network (BPNN), which is a subunit in Artificial Neural Networks (ANNs). The problems with traditional BPNN, including longer learning period, not prone to convergence and re-training while deleting or adding new patterns, make the realization of real time BPNN system impossible. My proposition is to integrate the parallel distributed process concept into the modification of the BPNN structure. These could accelerate the learning speed and solve the re-training problem in BPNN. The character segmentation, noise removal and extraction of feature are also discussed. Adequate extracted features help the recognition of characters easier. The adoption of White Run-Length and Pixel Density could clearly display the structural and integral properties of characters respectively. It could facilitate the goal of higher recognition accuracy