二值卷積神經網路硬體加速器設計與實作;Design and Implementation of Hardware Accelerator for Binarized Convolutional Neural Network

NCU Institutional Repository > 資訊電機學院 > 資訊工程研究所 > 博碩士論文 > Item 987654321/86604

請使用永久網址來引用或連結此文件: http://ir.lib.ncu.edu.tw/handle/987654321/86604

題名:	二值卷積神經網路硬體加速器設計與實作;Design and Implementation of Hardware Accelerator for Binarized Convolutional Neural Network
作者:	王政晏;Wang, Jheng-Yan
貢獻者:	資訊工程學系
關鍵詞:	神經網路硬體加速器;二值卷積神經網路;Hardware Accelerator for Neural Network;Binarized Convolutional Neural Network
日期:	2021-08-03
上傳時間:	2021-12-07 13:01:09 (UTC+8)
出版者:	國立中央大學
摘要:	在解決影像識別的問題上，深度學習取得了很好的成果，其中卷積神經網路是最具代表性的模型，也成為現今解決影像辨識問題的主流方法，然而其龐大的運算量與記憶體佔用，使得神經網路模型難以運用在硬體資源受限的邊緣裝置上。為了解決上述問題，有許多新的模型被提出來，後來甚至發展出二值卷積神經網路，大幅降低了深度神經網路的硬體資源需求。隨著近幾年智慧物聯網的興起，為了提供邊緣裝置即時且有效解決影像識別問題的方法，增強邊緣裝置的能力，本論文以ReActNet二值卷積神經網路模型作為參考，並透過階層式模組化的設計方法，設計了一個具有彈性架構的二值卷積神經網路硬體加速器，且在硬體設計上使用管線化技術，以提升神經網路的推論速度，同時使用8位元的定點運算取代浮點數運算，減少硬體資源的使用量。根據實驗結果，硬體加速器的單一影像辨識速度是使用圖形處理器的伺服器的27倍，雖然辨識率差了2.71%，但整體的耗電量較低，是伺服器的0.06倍。本論文提出之二值卷積神經網路硬體加速器不僅具有彈性架構，且在速度上具有即時性，能夠很好的提供硬體資源受限的裝置解決影像識別問題的能力。;In solving the image recognition problem, deep learning has achieved good results, among which the convolutional neural network is the most representative model and has become the mainstream approach to solve the image recognition problem nowadays. However, the huge computation and memory consumption make it difficult to use the neural network model on the edge devices with limited hardware resources. To solve these problems, many new models have been proposed, and later even binary convolutional neural networks have been developed to significantly reduce the hardware resource requirements of deep neural networks. With the emergence of Artificial Intelligence of Things in recent years, to provide a real-time and effective solution to the image recognition problem and enhance the capability of the edge devices, this paper uses ReActNet as a reference for the binary convolutional neural network model and designs a binary convolutional neural network hardware accelerator with a flexible architecture through a hierarchical modular design approach. In addition, the pipelining technique is used in the hardware design to improve the inference speed of the neural network, and 8-bit fixed-point computing is used instead of floating-point computing to reduce the hardware resource usage. According to the experimental results, the single image recognition speed of the hardware accelerator is 27 times faster than that of the server using graphics processor. Although the recognition rate is 2.71% worse than that of the server, the overall power consumption is lower than that of the server by 0.06 times. The binary convolutional neural network hardware accelerator proposed in this paper not only has a flexible architecture, but also has real-time speed, which can provide a good ability to solve image recognition problems for devices with limited hardware resources.
顯示於類別:	[資訊工程研究所] 博碩士論文

文件中的檔案:

檔案	描述	大小	格式	瀏覽次數
index.html		0Kb	HTML	84	檢視/開啟

在NCUIR中所有的資料項目都受到原著作權保護.

社群 sharing

資料載入中.....