臺灣博碩士論文加值系統

現今實現影像的物件識別都是以CNN(Convolutional Neural Network)為架
構在GPU(Graphics Processing Unit)上去做相關運算，速度快，但是運算的
資源消耗大，因此BNN(Binary Neural Network)是將權值(weight)以及激勵函
數(activation)以二值化的形式進行運算，雖然精確度降低，但是大量提升了運算
的速度。
FPGA(Field Programmable Gate Array)在運算二值化網路上更具有優勢，
所以將BNN應用在FPGA上不僅快速且運算所耗費的資源也小很多，雖然在精
準度上還有進步的空間，如果可以應用在邊緣裝置上，將會是個值得研究的領
域。FINN(A Framework for Fast, Scalable Binarized Neural Network Inference)是
一種用BNN網路的向前傳導的架構，只有1.5ms的延遲，非常適合在邊緣裝置上
去做相關應用。
本文探討如何基於FPGA的平台去實現以BNN為架構去完成視覺影像的物件
辨識。
關鍵字: 深度學習、二值化神經網路、機器人視覺、影像分類

Nowadays, the object detection of images is base on CNN(Convolutional Neural
Network) and to do operation on GPU. Although the computing speed is fast,it
takes too much power. Therefore,we provide BNN(Binary Neural Network) which
uses binarize weight and activation. Although reduces some accuracy, it drastically
increase accuracy. FPGA(Field Programmable Gate Array) has speed advantage
for BNN operation.Therefore, we can use BNN on FPGA. Although BNN has lower
accuracy,it has high computing speed and low power consumption.If we apply BNN
on FPGA or other edge device,it would be a new research area.FINN is a forward
conduction structure that base on BNN.It can classify image in 1.5ms.It’s very
suitable for the edge device. In this paper,we proposed a system that combines
BNN and robot vision on FPGA.
keywords: Deep learning 、Binary Neural Network、、Robot vision、Image
classification

書名頁 . . . . . . . . . . . . . . . . . . . . . . . . i
論文口試委員審定書 . . . . . . . . . . . . . . . . . . ii
中文摘要 . . . . . . . . . . . . . . . . . . . . . . . iii
英文摘要 . . . . . . . . . . . . . . . . . . . . . . . iv
誌謝 . . . . . . . . . . . . . . . . . . . . . . . . . v
目錄 . . . . . . . . . . . . . . . . . . . . . . . . . vi
表目錄 . . . . . . . . . . . . . . . . . . . . . . . viii
圖目錄 . . . . . . . . . . . . . . . . . . . . . . . . ix
一、緒論 . . . . . . . . . . . . . . . . . . . . . . . .1
1.1 研究背景 . . . . . . . . . . . . . . . . . . . . . . 1
1.2 研究動機及目的 . . . . . . . . . . . . . . . . . . . 2
1.3 論文架構 . . . . . . . . . . . . . . . . . . . . . . 2
二、相關文獻探討 . . . . . . . . . . . . . . . . . . . . 3
三、研究探討 . . . . . . . . . . . . . . . . . . . . . . 5
3.1 研究方法及架構介紹 . . . . . . . . . . . . . . . . . 5
3.1.1 PYNQ內部架構 . . . . . . . . . . . . . . . . . . . 7
3.2 BNN 介紹 . . . . . . . . . . . . . . . . . . . . . . 8
3.3 FINN 介紹 . . . . . . . . . . . . . . . . . . . . . 10
3.4 實驗環境介紹 . . . . . . . . . . . . . . . . . . . . 13
3.4.1 軟體介紹 . . . . . . . . . . . . . . . . . . . . . 13
vi
3.4.2 硬體環境介紹 . . . . . . . . . . . . . . . . . . . 15
四、研究成果與分析 . . . . . . . . . . . . . . . . . . . . 17
4.1 實驗數據 . . . . . . . . . . . . . . . . . . . . . . 17
4.2 硬體架構 . . . . . . . . . . . . . . . . . . . . . . 22
4.2.1 vivado電路圖 . . . . . . . . . . . . . . . . . . . 24
4.3 CNN訓練 . . . . . . . . . . . . . . . . . . . . . . 28
4.4 PS端設置 . . . . . . . . . . . . . . . . . . . . . . 40
4.5 影像分類結果 . . . . . . . . . . . . . . . . . . . . 41
五、結論 . . . . . . . . . . . . . . . . . . . . . . . . 46
參考文獻 . . . . . . . . . . . . . . . . . . . . . . . . 47

[1] 吳侑霖, “以深度學習結合影像分類實現自動駕駛系統,” 元智大學通訊工程學
系學位論文, 2018.
[2] M. Rastegari, V. Ordonez, J. Redmon, and A. Farhadi, “Xnor-net: Imagenet
classification using binary convolutional neural networks,” European Conference
on Computer Vision, 2016.
[3] S. Han, H. Mao, and W. J. Dally, “Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding,” arXiv
preprint arXiv:1510.00149, 2015.
[4] M. Courbariaux, I. Hubara, D. Soudry, and Y. B. Ran El-Yaniv, “Binarized
neural networks: Training deep neural networks with weights and activations
constrained to +1 or -1,” arXiv:1602.02830, 2016.
[5] S. Zhou, Y. Wu, Z. Ni, X. Zhou, H. Wen, and Y. Zou, “Dorefa-net: Training low bitwidth convolutional neural networks with low bitwidth gradients,”
arXiv:1606.06160, 2016.
[6] R. Zhao, W. Song, W. Zhang, T. Xing, J.-H. Lin, M. Srivastava, R. Gupta,
and Z. Zhang, “Accelerating binarized convolutional neural networks with
software-programmable fpgas,” Int’l Symp. on Field-Programmable Gate Arrays (FPGA), 2017.
[7] Umuroglu, Yaman, Fraser, N. J, G. Gambardella, M. Blott, P. Leong, M. Jahre,
and K. Vissers, “FINN: A Framework for Fast, Scalable Binarized Neural Network Inference, journal = Proceedings of the 2017 ACM/SIGDA International
Symposium on Field-Programmable Gate Arrays, year = 2017, publisher =
ACM,.”
[8] C. Zhang, P. Li, G. Sun, Y. Guan, B. Xiao, and J. Cong, “Optimizing fpgabased accelerator design for deep convolutional neural networks,” In Proc.
ACM/SIGDA ISFPGA, 2015.
47
[9] R. Andri, L. Cavigelli, D. Rossi, and L. Benini, “Yodann: An ultra-low
power convolutional neural network accelerator based on binary weights,”
CoRR,abs/1606.05487, 2016.
[10] H. Alemdar, N. Caldwell, V. Leroy, A. Prost-Boucle, and F. Petrot, “Ternary
neural networks for resource-efficient ai applications,” CoRR,abs/1609.00222,
2016.
[11] SI.Venieris and C.-S. Bouganis, “fpgaconvnet: A framework for mapping convolutional neural networks on fpgas,” 2016 IEEE 24th Annual International Symposium on Field-Programmable Custom Computing Machines (FCCM), 2016.
[12] C. Farabet, C. Poulet, J. Y. Han, and Y. LeCun, “Cnp:an fpga-based processor
for convolutional networks,” IEEE, 2009.
[13] S. K. Esser, P. A. Merolla, J. V. Arthur, A. S. Cassidy, R. Appuswamy, A. Andreopoulos, D. J. Berg, J. L.McKinstry, T. Melano, D. R. Barch, and et al,
“Convolutional networks for fast, energy-efficient neuromorphic computing,”
CoRR, abs/1603.08270, 2016.