Handwritten Postal Zone Number Recognition System

[[abstract]]本論文主要是建立在一個以可塑性認知類神經網路(Plastic Perception Neural Network)為主的手寫郵遞區號辨識系統。 在文中所用的可塑性認知類神經網路，是以具有監督式(supervised)學習的倒傳遞類神經網路(Back-propagation Neural Network)為基礎，並藉由網路架構的修改來克服傳統類神經網路，訓練耗時、不易收歛及增加新圖樣時，需整個網路重新訓練等缺點，來達到快速學習，且具有彈性加入新的圖樣類別單獨訓練即可。 本文尚需應用一些預處理技術，包括雜訊去除、文字切割、正規化。適當的特徵抽取方法是有助於提高辨識效率；本文中所採用的特徵抽取是採白色跑長碼(White Run-Length)和圖素密度(Pixel Density)的結合，抽取出65個特徵向量值。 我們取二組不同數量的樣本分別訓練及測試，最後再以信封上書寫的郵遞區號來測試，所得到的辨識率分別為88%、95%及92%，成效令人滿意。[[abstract]]In this thesis a handwritten postal zone number recognition system is proposed, based on the Plastic Perception of neural networks. The Plastic Perception Neural Network(PPNN) is based on the supervised learning of Back-propagation Neural Network(BPNN), and improved the traditional neural network which has the problems of time-consuming, difficult convergence and retraining of additional new patterns .In this thesis, we need some preprocessing technologies, such as noise elimination, text extraction and normalization. The feature extraction, in this thesis combine White Run-Length Coding and Pixel Density, to extract 65 elements of feature vector. From Two sets of samples with different quantity are trained and tested. At last, the postal zone numbers on the real envelopes are also examined. The satisfying recognition ratios are 88%, 95% and 92% respectively.