地圖中道路名稱之辨認

Abstract

本論文提出一個利用索引表來確認地圖中道路名稱的辨識方法。一般 來說，地圖中提供諸如︰地理標誌、國家、城市、道路、格線、河流以及 建築物之名稱。地圖中道路名稱分佈不均勻，它們可能會旋轉或具有不同 的大小和字型，或是與線段接觸在一起，甚至有的道路名稱和建築物名稱 混合在一起，這些問題使道路名稱的抽取和辨認非常困難，本論文提出一 些方法來解決這些問題。首先，本論文提出一個以相連單元(connected component)為本的方法來抽出地圖中的索引表(index table)以及此表中 的文字，這些方法包括二值化(binarization)、相連單元分析、規則檢查 以及瞭解(understanding)。其次，提出一個串級式光學文字辨識(OCR)系 統來辨認索引表內的字，這些文字包括明體的中文字、英文字以及數字。 在這個OCR系統中包括四個步驟︰前處理、特徵抽取、分類以及後處理。 在前處理中使用二值化和遮照(mask)運算，在特徵抽取中計算四個特徵︰ Walsh轉換、像素分佈(pixel distribution)、筆劃穿越數(crossing count)以及長的水平和垂直筆劃。在分類中使用大分類和串級式細分類。 在後處理中使用道路名稱在地圖中出現的頻率次數。然後取前兩名當作辨 認結果。最後，提出一個以幾何為本演算法來確認地圖中的道路名稱。這 個演算法包括利用道路的延伸性、符合最短距離、規則檢查以及滿足一些 限制。在我們的實驗中，共使用九張地圖，其中索引表和表內的文字完全 抽出，索引表內的文字辨識率為98.23%，地圖中的道路名稱確認率 為95.54%。這些實驗結果，證明我們所提出的系統是相當有效的。 This thesis presents a system to identify road names from a map. A map consists of a large number of entities, such as geographic landmarks, cities, rivers, roads, grid lines, country borders, institution names, and city borders. Road names often distribute nonuniformly in a map. They may rotate, vary in size or font type, touch lines or touch each other seriously. These problems make character extraction and recognition of road names very difficult. This thesis presents some methods to solve these problems. We first propose a connected-component-based method to extract the index table and its characters. The operations include binarization, connected component analysis, rules checking, and understanding. Second, we propose an cascade OCR system to recognize the characters of the index table. The characters include Ming font, multi-size Chinese characters, English characters and numerals in a map. In this OCR system, statistical and structural features are used to recognize characters. This system consists of four phases: preprocessing, feature extraction, classification, and postprocessing. The preprocessing phase performs thresholding and masking. In the feature extraction phase, we compute four types of features, which are Walsh transformation, pixel distributions, crossing counts, and long horizontal-vertical strokes. In classification phase, preclassification and cascade fine matching are performed. Three kinds of features are used serially in the fine matching. In the final postprocessing phase, candidates outputted from the fine matching module are tuned according to the frequency rate trained from characters collected from a tour book. At last, the top two candidates with the highest scores are chosen as the recognition results. In the last part, a geometry-based algorithm is proposed to identify the road names of the index table. This algorithm includes road line extension, shortest distance coincidence, rules checking, and some constraint satisfaction. The testing maps of our experiments contain nine maps. All index tables and characters in the index tables can be extracted. The recognition rate of characters in index tables is 98.23%, and the identification rate of road names is 95.54%. These experimental results show that the proposed system is rather effective.