基于FPGA的手写蒙文字体转换系统设计及实现

Design and Implementation of FPGA-Based Handwritten Mongolian Script Conversion System

摘要:

蒙文字体转换在促进蒙文应用和推广、丰富中国文化多样性以及促进蒙古族地区经济繁荣方面具有关键作用。针对蒙文字体转换的效和准确率低的问题，提出一种基于轻量化卷积神经网络(CNN)和FPGA硬件加速器的方法。即使用CNN进行手写蒙文识别，并结合字体转换库，通过识别结果和字体的映射关系实现了简单高效的蒙文字体转换。相比于其他方法，该方法结合了高效的CNN和FPGA硬件加速器的优势，既提高了转换效率又满足了设备成本、功耗和便携性的需求。使用Xilinx公司的XC7Z020CLG400-2完成网络模型电路的设计和优化工作，在此基础上实现了手写蒙文字体转换系统，测试结果表明，手写体蒙文转换为目标字体的准确率为95.62%，转换时间为1.43ms，功耗为0.341W，加速器峰值吞吐量为6.64Gops。研究成果对于促进蒙古族文化传承和经济发展具有重要意义。

Abstract:

The conversion of Mongolian fonts plays a pivotal role in promoting the application and dissemination of the Mongolian script, enriching the diversity of Chinese culture, and fostering economic prosperity in Mongolian regions. The method based on lightweight convolutional neural network and FPGA hardware accelerator is adopted to improve the efficiency and accuracy of Mongolian font conversion. By utilizing CNN for handwritten Mongolian script recognition and integrating it with a font conversion library, a simple and efficient Mongolian font conversion is achieved through the mapping relationship between recognition results and fonts. In comparison to other methods, this approach combines the efficiency of CNN and FPGA hardware accelerators, improving conversion efficiency while meeting requirements for device cost, power consumption, and portability. The design and optimization of the network model circuit were accomplished using Xilinx's XC7Z020CLG400-2. On this basis, a handwritten Mongolian font conversion system was implemented. Test results indicate an accuracy of 95.62% for converting handwritten Mongolian script to the target font, a conversion time of 1.43ms, a power consumption of 0.341W, and a peak throughput of 6.64Gops for the accelerator. The research findings hold significant importance for promoting the preservation of Mongolian culture and facilitating economic development among the Mongolian population.