一個新的特徵分析演算法改善中文簽名辨識

A Novel Feature Analysis Approach Improve to Chinese Signature Recognition

吳至偉(Chin-Wei Wu)

指導教授 ： 方志鵬 ； 張陽郎

國立臺北科技大學/電資學院/電機工程系研究所/碩士(2007年)

https://doi.org/10.6841/NTUT.2007.00132

摘要

在簽名辨識流程中，需經過特徵抽取的步驟，從原始資料擷取各種的特徵資訊。在簽名辨識研究中，有許多特徵被提出來改善辨識率與安全性，也造成有許多雷同或是多餘特徵產生，普遍而言，特徵數目的多寡，會實際影響到簽名辨識的速度與運算量。因此在分類過程之前，運用特徵分析的處理，可強化特徵的品質與穩定度，並刪去相似度極高之特徵。若是未經由特徵分析的過程，將造成「簽名辨識系統」效能無法有效的發揮效能。
本論文中運用遙測影像降低資料維度的方法：「貪婪特徵空間演算法」 (Greedy Modular Eigenspace, GME)與「特徵尺度齊一化轉換」 (Feature Scale Uniformity Transformation, FSUT)演算法，來減少分類時的特徵數目，即可以降低分類時的運算量，提升「簽名辨識系統」的執行速度。利用統計理論中的相關係數公式，可擷取出特徵之間的相似程度，經由本論文的特徵分析處理之後，可刪除性質雷同的特徵，使得在簽名辨識效能上，可以明顯增加特徵的利用性與有效性。本論文初步建立了一個簽名資料庫，具有動態與外型特徵資訊，經由「貪婪模組特徵空間」�「特徵尺度齊一化轉換」將特徵篩選，驗證我們所提出的方法，可改善特徵抽取的不足。文中並也實作此一個中文「簽名辨識系統」於個人數位助理(Personal Digital Assistant, PDA)之上，可在PDA上實際操作「簽名辨識系統」，證明本文的所提方法除了可提升簽名辨識速度與正確率之外，並可以直接實現於商用中文「簽名辨識系統」中。

關鍵字

特徵抽取 ； 特徵分析 ； 貪婪特徵空間 ； 特徵尺度齊一化轉換 ； 中文簽名辨識

並列摘要

In the online character recognition, the trajectories of pen tip movements are recoded and analyzed to identify the writer. Online character recognition is able to yield higher accuracy than offline recognitions because of temporal stroke and spatial shapes information. It can also correct the errors and improve the accuracies with a large volume of writer information.
The applications of online recognitions include text entries used for form filling and message compositions, personal digital assistants (PDA), computer-aided educations, handwritten document retrievals, and etc. Another useful application is the signature verification that checks whether a specific writer generates a personal handwriting signature. In the thesis, we implement a novel Chinese signature recognition technique, which is originally developed for the classification of the remotely sensed hyperspectral images recently.
Since a huge volume of features is collected to improve the classification accuracy in our proposed GME method, the repeated and redundant features are expected to be larger. In order to improvement performance, we proposed a novel greedy modular eigenspace (GME) method to reduce feature dimensions. In this thesis, the performance of GME is evaluated in comparison with the conventional feature extraction methods. The approach is designed as a feature extractor for Chinese signature recognitions to simplify these redundant features. It can extract the simplest and the most efficient signature feature modules collected from each signature that includes real-stroke and virtual-stroke features. A new developed feature scale uniformity transformation (FUST) is also performed to fuse the most similar signature features from different writers.
The performance of the proposed GME/FUST method is evaluated by a signature database. The experimental results demonstrate the proposed method is an effective scheme not only for the feature extractions of the signature recognitions but also for computational reductions complexity of the signature databases.

並列關鍵字

Feature Extraction ； Greedy Modular Eigenspace ； Chinese Signature Recognition ； Writer Verification

參考文獻

[1].黃文增、餘信緯、沈毅偉、黃仲麟、周慶棟，低功率、低價格與小面積之指紋辨識系統之研究與實作，臺北科技大學學報第三十八之二期，2005，第23-38頁。

[3].Y. L. Chang, H. C. Han, K. C. Fan, K. S. Chen, C. T. Chen, and J. H. Chang, “Greedy Modular Eigenspaces and Positive Boolean Function for Supervied Hyperspectral image Classification,” Optical Engineering, vol. 42, no. 9, 2003, pp. 2576-2587.

[4].Y. L. Chang, H. C. Han, H. Ren, C. T. Chen, K. C. Fan , and K. S. Chen, “Data Fusion of Hyperspectral and SAR images,” Optical Engineering, vol. 43, no. 8, 2004, pp. 1787-1797.

[5].F. Z. Marcos, “Online Signature Recognition based on VQ-DTW”, Pattern Recognition, vol. 40, no. 3, 2007, pp. 981-992.

[6].Liu C. L., Jaeger S., Nakagawa M., “Online recognition of Chinese characters: the State-of-the-art,” Pattern Analysis and Machine Intelligence IEEE Transactions on, vol. 26, no. 2, Feb. 2004, pp. 198-213.