摘要
近年来,卷积神经网络在图像处理方面得到了广泛应用,然而其存在计算复杂,移动端资源有限,无法存储过多数据、进行大规模计算等缺点。提出一种基于CNN的汉字识别系统的硬件实现方法。在TensorFlow框架下用casia数据集训练出20个常用汉字的CNN网络架构,测试集识别率达98.36%,并采用卷积核复用、定点化等方法降低资源消耗,在FPGA上搭建优化后的CNN。最后,将摄像头实时采集的图片输入到上述CNN,实现硬件端的汉字识别。实验结果表明,在结构简化、速度相较CPU提高6.76倍的同时,在FPGA上所构建的CNN达到几乎无损的97.58%的准确率。
In recent years,convolutional neural networks have been widely used in image processing.However,their computation is complex,mobile resources are limited,so it is impossible to store too much data and perform large-scale calculations.In this paper,we propose a hardware implementation method of CNN Chinese character recognition system based on FPGA.In the TensorFlow framework,the CNN network architecture of 20 commonly used Chinese characters was trained with the casia dataset,and the test set recognition rate was 98.36%.Convolutional core multiplexing,fixed-point method are used to reduce resource consumption,and the optimized CNN is built on the FPGA.Finally,the image captured by the camera in real time is input to the CNN to realize the Chinese character recognition on the hardware side.The experimental results show that while the structure is simplified and the speed is increased by 6.76 times compared with the CPU,the CNN constructed on the FPGA achieves an almost non-destructive accuracy of 97.58%.
作者
潘思园
王永
黄鲁
Pan Siyuan;Wang Yong;Huang Lu(School of Microelectronics,University of Science and Technology of China,Hefei 230026,China)
出处
《信息技术与网络安全》
2019年第9期44-49,54,共7页
Information Technology and Network Security
关键词
FPGA
CNN
汉字识别
硬件加速
FPGA
CNN
Chinese character recognition
hardware acceleration
