基于CNN的手写识别芯片关键计算模块硬件加速方案设计
首发时间:2019-04-02
摘要:手写识别作为输入法核心的一部分,其快捷性和准确性一直是近年来科技研究的重点方向。近几十年来,在众多学者和科研人员的的努力下,手写识别技术取得较大进步,但由于移动设备CPU、内存等性能的制约下,手写识别仍存在速度较慢以及识别准确度低等问题,这种延迟会造成功耗大、缩短设备寿命、造成用户体验感差等问题。基于CNN理论指导下,对手写识别应用进行硬件加速是解决此类问题的有效方法。我们通过对Lenet-5结构进行并行性分析,利用多播片上网络的扩展性好和通信带宽高的优势,在片上网络的节点挂载硬件加速单元,协调各硬件加速单元的操作完成整个识别过程的硬件加速。
For information in English, please click here
Design of Hardware Acceleration Scheme for Key Computing Module of Handwritten Recognition Chip Based on CNN
Abstract:Handwritten recognition, as a core part of input method, has been the focus of scientific and technological research in recent years for its rapidity and accuracy. In recent decades, with the efforts of many scholars and researchers, handwriting recognition technology has made great progress. However, due to the constraints of CPU, memory and other performance of mobile devices, handwriting recognition still has some problems, such as slow speed and low recognition accuracy. This delay will lead to large power consumption, shorten the life of devices, and result in poor user experience. Under the guidance of CNN theory, hardware acceleration of handwriting recognition application is an effective way to solve such problems. By analyzing the parallelism of Lenet-5 structure, we use the advantages of good scalability and high communication bandwidth of multicast network to mount hardware acceleration units on the nodes of network-on-chip, coordinate the operation of each hardware acceleration unit to complete the hardware acceleration of the whole identification process.
Keywords: Handwritten Recognition CNN Acceleration Unit
引用
No.****
动态公开评议
共计0人参与
勘误表
基于CNN的手写识别芯片关键计算模块硬件加速方案设计
评论
全部评论0/1000