運用影像處理技術之仿冒中文字跡辨識

字跡是一種富含資訊的生物識別特徵而字跡辨識也在法務鑑定上扮演非常重要的角色。然而，由於行為特性導致個體字跡的變異程度大，使得字跡辨識仍然是一個具有挑戰性的研究題目。在本篇論文中，我們將會提出兩種能增加辨識率的中文字跡辨識系統。在第一個方法中，我們利用區域特徵和支持向量機來進行字跡辨識。首先藉由組合各種偵測子和描述子，包括高斯差、哈里斯角點、尺度不辨特徵轉換描述子以及方向性亮度區塊描述子，來產生區域特徵。由於每張圖片所找出的特徵點數目不同，因此需用K均值分群法先建立編碼簿。然後根據詞袋模型，每份手寫字跡便能以編碼頻率直方圖表示，進而成為支持向量機的輸入特徵向量。之後我們也開始研究全域特徵在字跡辨識的應用，所以另外一個提出的方法則是以全域特徵作為基礎。這個系統利用了對數賈伯特徵、進階不變矩和灰階共生矩陣的擷取特徵，並得到比前者更佳的辨識率。藉由這些特徵的結合，這個系統在傳統辨識問題上顯現出更卓越的穩健性。除此之外，我們也提出了一個更有彈性的分類架構。雖然支持向量機提供了準確的結果，但卻為訓練資料的數量所限。為了避免在訓練資料不足或是不平衡的情況下發生過度擬合，另外一個分類方法僅以加權歐氏距離平方作為辨識基準。模擬結果顯示，我們提出的字跡辨識系統在有/無搭配支持向量機的架構下分別能達到92.7% 和83.5%的準確度，而這也超越了其他現有的辨識方法，包括區域二元模式、區域方向模式、賈伯特徵、分段筆劃直方圖以及分段曲度編碼。

關鍵字

字跡辨識；高斯差；哈里斯角點；尺度不變特徵轉換；方向性亮度區塊； K均值分群法；支持向量機；對數賈伯特徵；不變矩；灰階共生矩陣；加權歐氏距離平方

並列摘要

Handwriting is an informative kind of biometrics and writer verification plays a very important role in forensics. However, writer verification remains a challenging topic due to the large variations caused by the behavioral trait of individuals. In this thesis, two systems of writer verification are proposed to improve verification accuracy. The first method is to perform writer verification with local features and the support vector machine. These local features are generated by different combinations of detectors and descriptors, including the difference of Gaussian, the Harris corner, the SIFT descriptor and the oriented intensity patch. Since the amount of keypoints are various, a construction of codebook is required. K-means clustering is applied to build the codebook. Then, by the bag-of-word model, each handwriting image can be represented by a histogram with the codewords being indices. The histograms are the input feature vectors used for the support vector machine, which is a famous technique in machine learning. Later, global features are also explored in this research. So another method based on global features is proposed. This verification performs even better by using log-Gabor features, some advanced moments and features extracted from gray level co-occurrence matrices. By combining these features, the system shows superior robustness for traditional verification problems. Besides, a more flexible classification framework is proposed. Though the support vector machine leads to accurate results, it is confined to the amount of training data. To prevent overfitting, another classification method based on the weighted squared Euclidean distance is devised for the case of insufficient or unbalanced training data. From the results of simulation, the accuracies can reach about 92.7% and 83.5% for the proposed framework with/without the support vector machine, respectively, which outperform other popular identification or verification methods, including the local binary pattern, the local directional pattern, Gabor features, the stroke fragment histogram and the curve fragment code.

並列關鍵字

writer verification ； difference of Gaussian ； Harris corner ； scale-invariant feature transform ； oriented intensity patch ； K-means clustering ； support vector machine ； log-Gabor feature ； moment invariant ； gray level co-occurrence matrix ； weighted squared Euclidean distance

參考文獻

A. Feature Extraction Basics

[2] Manjunath, Bangalore S., and Wei-Ying Ma. "Texture features for browsing and retrieval of image data." Pattern Analysis and Machine Intelligence, IEEE Transactions on 18.8 (1996): 837-842.

[3] Winston H. Hsu. "Texture and shape for image retrieval and multimedia analysis and indexing." National Taiwan University, 2013.

[4] Ojala, Timo, Matti Pietikäinen, and David Harwood. "A comparative study of texture measures with classification based on featured distributions." Pattern recognition 29.1 (1996): 51-59.

[6] Jabid, Taskeed, Md Hasanul Kabir, and Oksam Chae. "Facial expression recognition using local directional pattern (LDP)." Image Processing (ICIP), 2010 17th IEEE International Conference on. IEEE, pp. 1605-1608, 2010.

國際替代計量

運用影像處理技術之仿冒中文字跡辨識

主題瀏覽