| 张志超* ** *** ****,王剑* ** ***,章隆兵* ** ***,肖俊华* ** ***.面向目标检测的卷积神经网络优化方法[J].高技术通讯(中文),2022,32(3):227~238 |
| 面向目标检测的卷积神经网络优化方法 |
| Convolutional neural network optimization method for object detection |
| |
| DOI：10.3772/j.issn.1002-0470.2022.03.002 |
| 中文关键词: 卷积神经网络(CNN); 现场可编程门阵列(FPGA); 数据流调度; 目标检测; 加速 |
| 英文关键词: convolution neural network (CNN), field programmable gate array (FPGA), data stream scheduling, object detection, acceleration |
| 基金项目: |
| 作者 | 单位 | | 张志超* ** *** **** | | | 王剑* ** *** | | | 章隆兵* ** *** | | | 肖俊华* ** *** | |
|
| 摘要点击次数: 1337 |
| 全文下载次数: 1740 |
| 中文摘要: |
| 针对星载等功耗受限平台下遥感影像目标检测存在的高准确率、低功耗以及高吞吐量等要求，本文提出了一种面向目标检测的现场可编程门阵列(FPGA)卷积神经网络（CNN）优化方法。采用数据流调度技术以及基于乘法矩阵与前向加法链的卷积计算阵列设计对浮点卷积神经网络模型进行加速。利用该方法在FPGA开发板上实现了浮点卷积目标检测网络，在应用中达到了与原模型一致的准确率，平均准确率为97.59%，吞吐量达到了Titan X的22倍。与同类的FPGA加速浮点卷积方法对比，该方法的吞吐量以及能效比达到了最优。实验数据表明，该方案突破了浮点卷积加速的线速吞吐难点，解决了应用中存在的功耗、准确率以及吞吐量三者制衡的问题。 |
| 英文摘要: |
| Aiming at the requirements of high accuracy, low power consumption and high throughput in remote sensing image object detection on space-borne and other power-constrained platforms, this paper proposes an optimization method of field programmable gate array (FPGA) convolutional neural network (CNN) for object detection. Data stream scheduling technology and convolutional computing array design based on multiplication matrix and forward addition chain are used to accelerate the floating point convolutional neural network model. A floating point convolution object detection network is implemented on FPGA development board by using this method. In application, the accuracy is consistent with the original model, the average accuracy is 97.59%, and the throughput is 22 times that of Titan X. Compared with similar FPGA accelerated floating-point convolution methods, the throughput and energy efficiency ratio of the proposed method are optimal. The experimental results show that the proposed method breaks through the line-speed throughput difficulty of floating-point convolution acceleration, and solves the balance of power consumption, accuracy and throughput in the application. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
