使用加入空間資訊之形狀內容特徵應用於自然場景中字元辨識

Using Shape context with Spatial Information for Character Recognition in Natural Images

黃士瑋(Shih-Wei Huang)

指導教授 ： 蔡宏營

國立清華大學/工學院/動力機械工程學系/碩士(2013年)

https://doi.org/10.6843/NTHU.2013.00327

摘要

自然場景影像中的字元通常包含多種不同字體，而且因為拍照或是環境因素的影響，可能導致字元的變形與破碎，造成辨識上的困難。根據形狀內容特徵之特性，可以用來針對自然場景影像中不同字型之字元進行辨識，並且容許字元有些微的變形，因此本研究選用形狀內容作為特徵來對自然場景中的字元影像進行辨識。
傳統上利用形狀內容特徵進行辨識時需要進行多次迭代來作對應，每次迭代都是使用匈牙利演算法對特徵點進行最佳對應。由於匈牙利演算法需要耗費大量計算時間，時間複雜度為O(n3)。因此本研究保留特徵點的二維空間資訊，對於形狀內容特徵點給予不同的空間標記，做特徵點對應時僅需要對同一標記之特徵點進行一次性對應，而不需要透過迭代方式，藉此提升辨識速度與效率。
本研究針對ICDAR 2003所提供的自然場景字元影像資料集(數字0~9與大寫英文字母A~Z，共5100張)進行辨識，得到最佳化形狀內容特徵參數，並且討論不同空間資訊參數對辨識結果的影響。相較於傳統形狀內容特徵對應方法，本研究所提出之方法，在辨識率與處理速度都有大幅的提升。

關鍵字

形狀內容 ； 自然場景影像 ； 數字辨識 ； 空間資訊

並列摘要

Natural scene images contain a variety of characters in different type of fonts. The camera and environmental factors could cause the characters to be deformed and be broken. The deformable and broken images make it hard to be recognized. Based on the property of shape context, this method can be used for natural scene images of the characters in different type of fonts, even allowing a few deformed in characters. Therefore, this study selected the shape contexts as feature for character recognition in natural scene images.
Traditionally, the shape contexts method requires multiple iterations to make feature point matching and each iteration used the Hungarian algorithm to optimize for feature point correspondence. Because the Hungarian algorithm requires a lot of computing time, the time complexity is O (n3). Therefore, this study added the two-dimensional spatial information of feature points, each feature points given the label from different spatial information. Only the corresponding feature point with the same label would be matched, without the need for iteration. The proposed method will improve character recognition speed and efficiency.
This study used the data set of ICDAR 2003 (digits 0 through 9 and the uppercase letters A ~ Z, a total of 5100 images) for character recognition. Based on the experimental results, this study got the best shape context parameters and the effect of different parameters of spatial information could be discussed. Compared to the traditional shape contexts of the corresponding method, the proposed method’s recognition rate and the processing speed improved dramatically.

並列關鍵字

Shape contexts ； Natural scene images ； character recognition ； spatial information

參考文獻

[1] O.D. Trier, A.K. Jain, and T. Taxt, “Feature extraction methods for character recognition—a survey,” Pattern Recognition, vol. 29, pp. 641–662, 1996.

[3] T.E. Campos, B.R. Babu, and M. Varma, “Character recognition in natural images,” in Proceedings of the International Conference on Computer Vision Theory and Applications, 2009.

[4] S. Belongie and J. Malik, “Matching with Shape context,” in Proceeding of IEEE Workshop Content-Based Access of Image and Video Libraries, pp. 20–26, 2000.

[6] Y. Wan, X. Xu, and L. Yao,” An efficient license plate character recognition algorithm based on Shape context,” SPIE Image Processing and Photonics for Agricultural Engineering, 2013.

[7] M. K. Hu, “Visual Pattern Recognition by Moment Invariants,” IRE Transactions on Information Theory, vol. IT-8, pp. 179–187, 1962.