智慧型字元辨識系統設計與實現：使用基於函數鏈結類神經網路

Design and Implementation of Intelligent Optical Character Recognition System by Functional-Link Neural Network

李慶鴻(Ching-Hung Lee) ； 劉欣宜(Hsin-Yi Liu) ； 李仲凱(Chung-Kai Lee)

《興大工程學刊》 23卷3期 (2012/11) Pp. 145-154

https://doi.org/10.6287/JENCHU.2012.2303.03

摘要

本文以函數鏈結類神經網路（functional-link-based neural network, FLNN）設計智慧型字元辨識器，辨識系統包括字元特徵值擷取與FLNN分類器。函數鏈結類神經網路以函數鏈結擴充為基底，且擁有抗雜訊和容失真的特性，在隱藏層節點個數較少的情況下，也能擁有很好的效能。本文提出重心距離特徵值和同心圓面積特徵值，其皆具有旋轉不變性以及抗雜訊的特性，適用於理想及旋轉、具雜訊字元之分類。在分類理想字元，辨識正確率為100％；分類非理想字元（具字元位置及大小、具雜訊和旋轉角度之誤差），辨識正確率94％～100％；而對於不同字型之分類（Arial與Times New Roman），辨識正確率亦維持80％以上，這些實驗均可說明所設計字元辨識器之可行性與性能。

關鍵字

字元辨識 ； 特徵抽取 ； 類神經網路 ； 影像處理

並列摘要

In this study, a functional-link-based neural network (FLNN) is adopted to develop an optical character recognition (OCR) system. The proposed OCR system contains the feature extraction and the FLNN classifier. The OCR system is valid for recognizing the capital letters and Arabic numerals. The proposed feature extraction technique includes the distance between each black pixel and concentric circles feature. They have the properties of rotation and noise invariant. In addition, the proposed FLNN with the expanded functional-basis has the properties of anti-noise and distortion characteristics. Finally, some experimental results are introduced to illustrate the performance and effectiveness of the proposed feature extraction and FLNN classifier. The proposed system performs 100% accuracy for actual characters and at least 94% even the characters has variations of location, size, rotation, and noisy. For different fonts (Arial and Time New Roman), the proposed system almost perform well and has 80% accuracy.

並列關鍵字

optical character recognition ； feature extraction ； neural network ； image processing