叠层模型驱动的书法文字识别方法研究

Calligraphy Character Recognition Method Driven by Stacked Model

基于二维图像的书法文字识别是指利用计算机视觉技术对书法文字单字图像进行识别,在古籍研究和文化传播中具有重要应用.目前书法文字识别技术已经取得了相当不错的进展,但依旧面临很多挑战,比如复杂多变的字形可能导致的识别误差,汉字本身又存在较多形近字,且汉字字符类别数与其他语言文字相比更多,书法文字图像普遍存在类内差距大、类间差距小的问题.为解决这些问题,提出叠层模型驱动的书法文字识别方法(Stacked-model driven character recognition,SDCR),通过使用数据预处理、节点分离策略和叠层模型对现有单一分类模型进行改进,按照字体类别对同一类别不同字体风格的文字进行二次划分;针对类间差距小的问题,根据书法文字训练集图像识别置信度对形近字进行子集划分,针对子集进行嵌套模型增强训练,在测试阶段利用叠层模型对形近字进行二次识别,提升形近字的识别准确率.为了验证该方法的鲁棒性,在自主生成的SCUT_Calligraphy数据集和CASIA-HWDB 1.1,CASIA-AHCDB公开数据集上进行训练和测试,实验结果表明该方法在上述数据集的识别准确率均有较大幅度提升,在CASIA-HWDB 1.1、CASIA-AHCDB和自建数据集SCUT_Calligraphy上测试准确率分别达到96.33%、99.51%和99.90%,证明了该方法的有效性.

Calligraphy character recognition based on two-dimensional images means to recognize single calligraphy character based on computer vision,which has important applications in ancient book research and cultural dissemination.At present,calligraphy character recognition has made considerable progress,but still faces many challenges,such as recognition errors caused by complex and variable font shapes,the existence of many similar characters in Chinese,and the number of Chinese character categories is extremely large.Calligraphy character images generally have large intra class differences and small inter class differences.In order to tackle these issues,we proposed a calligraphy character recognition method based on stacked model(SDCR).By using data preprocessing,node separation strategy and stacked model,and the characters with different font styles in the same category is subdivided according to the font style.To address the issue of small inter class differences,the calligraphy character training set image recognition confidence level is used to divide the characters with similar style into subsets.Nested model enhancement training is conducted on the subsets,and in the testing stage,a stacked model is used for secondary recognition of characters with similar style to improve the recognition accuracy of shape near characters.In order to verify the robustness of our proposed method,we train and test on self-generated dataset SCUT_Calligraphy and publicly available datasets CASIA-HWDB 1.1,CASIA-AHCDB.The experimental results showed that the proposed method significantly improved the recognition accuracy of the datasets mentioned above.The testing accuracy on CASIA-HWDB 1.1,CASIA-AHCDB and SCUT_Calligraphy reached 96.33%,99.51%,and 99.90%,respectively,which proves the effectiveness of the method described in this article.