摘要
卷积神经网络是深度学习的一种重要模型,广泛应用于图像处理等领域.常用的神经网络模型因结构复杂,参数众多,不适于放在移动端运行.本文基于模块化和硬件复用的思想,给出了一种基于FPGA的手写数字字符识别网络的硬件实现,基于MobileNet的原理改进结构,在实现了算法硬件加速的同时,有效地降低了网络的参数数量和整体运算量.基于MNIST数据集的实验结果表明,对比传统结构的神经网络,改进结构的参数量减少了23.26%,计算量减少了31.32%,在保持速度不变的前提下,用更少的资源和更低的功耗实现了整个网络.
Convolutional neural networks are an important model of deep learning and are widely used in image processing and other fields. The commonly used neural network model is complex and has many parameters, which is not suitable for running on the mobile end. Based on the idea of modularization and hardware reuse, this paper presents a hardware implementation of handwritten digital character recognition network based on FPGA. Based on the principle of MobileNet, the structure is improved, and the algorithm hardware acceleration is realized, and the number of parameters of the network and the overall calculation amount are effectively reduced. The experimental results based on the MNIST dataset show that compared with the traditional neural network, the parameter size of the improved structure is reduced by 23.26%, and the calculation amount is reduced by 31.32%. The entire network is implemented with less resources and lower power consumption while maintaining the same speed.
作者
李炳辰
黄鲁
LI Bing-chen;HUANG Lu(School of Microelectronics, University of Science and Technology of China,Hefei 230026, China)
出处
《微电子学与计算机》
北大核心
2019年第9期7-11,共5页
Microelectronics & Computer
基金
国家自然基金面上项目(61874102)
