[论文解读] Recognition of Non-Compound Handwritten Devnagari Characters using a Combination of MLP and Minimum Edit Distance

Sandhya Arora, Debotosh Bhattacharjee+3 more|arXiv (Cornell University)|Jun 30, 2010

Handwritten Text Recognition Techniques参考文献 15被引用 44

一句话总结

本文提出了一种用于离线手写非复合天城文字符的两阶段识别系统，结合多层感知机（MLP）分类器与最小编辑距离（MED），采用阴影特征和链码直方图特征。该方法在7,154个样本的数据集上实现了90.74%的识别准确率，通过加权多数投票进行初始分类，并利用哈里斯角点检测解决形状相似字符的识别问题。

ABSTRACT

This paper deals with a new method for recognition of offline Handwritten non-compound Devnagari Characters in two stages. It uses two well known and established pattern recognition techniques: one using neural networks and the other one using minimum edit distance. Each of these techniques is applied on different sets of characters for recognition. In the first stage, two sets of features are computed and two classifiers are applied to get higher recognition accuracy. Two MLP's are used separately to recognize the characters. For one of the MLP's the characters are represented with their shadow features and for the other chain code histogram feature is used. The decision of both MLP's is combined using weighted majority scheme. Top three results produced by combined MLP's in the first stage are used to calculate the relative difference values. In the second stage, based on these relative differences character set is divided into two. First set consists of the characters with distinct shapes and second set consists of confused characters, which appear very similar in shapes. Characters of distinct shapes of first set are classified using MLP. Confused characters in second set are classified using minimum edit distance method. Method of minimum edit distance makes use of corner detected in a character image using modified Harris corner detection technique. Experiment on this method is carried out on a database of 7154 samples. The overall recognition is found to be 90.74%.

研究动机与目标

• 解决由于形状变化和视觉相似性导致的手写非复合天城文字符高精度识别挑战。
• 通过将视觉相似字符分离到不同处理路径，减少其误分类。
• 通过特征多样性与基于加权多数投票的分类器融合策略，提升识别性能。
• 开发一种稳健方法，利用几何与结构特征区分易混淆的字符对。
• 通过结合神经网络与基于字符串编辑的分类策略，实现高总体识别准确率。

提出的方法

• 从输入字符图像中提取阴影特征与链码直方图特征，用于双MLP分类。
• 训练两个独立的MLP：一个使用阴影特征，另一个使用链码直方图特征。
• 通过加权多数投票方案融合两个MLP的决策，生成前三位候选字符。
• 计算前三位候选字符之间的相对差异值，将字符划分为两类：形状不同字符与易混淆字符。
• 对易混淆字符集应用最小编辑距离（MED），并利用改进的哈里斯角点检测算法提取角点。
• 基于角点序列的MED用于度量结构相似性，并为最相似的参考字符分配最终标签。

实验结果

研究问题

• RQ1 结合MLP与最小编辑距离的混合方法是否能提升手写天城文字符的识别准确率？
• RQ2 结合加权多数投票的阴影特征与链码直方图特征在初始分类中的有效性如何？
• RQ3 基于角点表示的最小编辑距离在解决视觉相似天城文字符混淆问题上的效果如何？
• RQ4 将字符划分为形状不同与易混淆类别是否能提升整体识别性能？
• RQ5 特征多样性与分类器融合对非复合天城文字符识别准确率的影响如何？

主要发现

• 所提出的两阶段方法在7,154个手写天城文字符样本的数据集上实现了90.74%的整体识别准确率。
• 使用双MLP结合阴影特征与链码直方图特征，通过特征层面的多样性提升了初始分类的可靠性。
• 加权多数投票方案通过聚合两个不同特征集的决策，有效减少了误分类。
• 结合角点检测的最小编辑距离方法显著提升了对视觉相似字符对的识别效果。
• 将字符划分为形状不同与易混淆集合，使系统能够采用定制化分类策略，增强了整体鲁棒性。
• 改进的哈里斯角点检测技术成功提取了对MED分类中模糊字符至关重要的结构特征。