一种基于多维表示的汉字识别方案

A Chinese Character Recognition Solution Based on Multidimensional Representation

针对复杂场景下汉字特征提取难的问题,提出了一种基于多维特征表示的汉字识别方案。首先,提出一种融合空间信息的关键笔形特征提取方法,能够利用少量关键特征实现汉字的唯一识别;其次,通过多任务网络提取多维特征,增强特征提取能力,从而提高汉字识别的准确性;最后,应用字符相似度算法消除噪声,优化识别结果。实验结果表明,相较于可插拔的部首感知分支(PRAB)模型,本方案在场景数据集、网页数据集、文本数据集和手写数据集中的性能分别提升了1.62百分点、1.09百分点、0.15百分点和1.27百分点,证明了该方案的有效性。

To address the challenge of extracting features for Chinese characters in complex scenarios,this paper proposes a Chinese character recognition solution based on multidimensional feature representation.Firstly,a key stroke feature extraction method integrating spatial information is introduced,capable of uniquely identifying Chinese characters using a small number of key features.Secondly,multidimensional features are extracted through a multitask network to enhance feature extraction capability and improve the accuracy of Chinese character recognition.Lastly,a character similarity algorithm is applied to eliminate noise and optimize recognition results.Experimental results show that compared to the PRAB(Pluggable Radical-Aware Branch)model,the proposed solution achieves performance improvements of 1.62 percentage points,1.09 percentage points,0.15 percentage points,and 1.27 percentage points in scene,web page,document,and handwritten datasets,respectively,demonstrating the effectiveness of the proposed solution.