Please use this identifier to cite or link to this item:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/78028
Title: | 視訊壓縮感知演算法與硬體加速器設計 Design of Algorithms and Hardware Accelerators for Video Compressive Sensing |
Authors: | Ying-Yu Tseng 曾纓喻 |
Advisor: | 盧奕璋(Yi-Chang Lu) |
Keyword: | 視訊壓縮感知,硬體設計,一階原始對偶演算法, Video Compressive Sensing,Hardware Design,First Order Primal Dual, |
Publication Year : | 2017 |
Degree: | 碩士 |
Abstract: | 視訊壓縮感知能有效的提高拍攝視訊的最高幀率。本論文將以8倍幀率的視訊壓縮感知為研究主軸,針對視訊壓縮感知重建速度與品質進行探討,加速視訊重建速度至每秒30幀以上供使用者即時預覽,並進一步將視訊壓縮感知推廣到彩色視訊,提出一套重建品質更好的彩色視訊壓縮感知流程。
首先我們測定現有的兩種先備字典視訊壓縮感知重建演算法,從中挑選出品質較好的高斯混合模型視訊重建演算法進行加速,嘗試以演算法、程式設計和平行程式設計的方式加速重建流程,然而重建流程的運算量龐大,資料相依性高,使得傳統運算系統在執行時硬體效率不佳,無法達到即時預覽的幀率。因此我們進一步設計多層平行化結構的硬體加速視訊重建,基礎單元為視訊重建子單元,在雙時鐘域下,以深度管線化與平行化等硬體設計技巧進行設計達到220倍的加速,以TSMC 90 nm製程實現,硬體尺寸為3.15 mm^2,運作頻率設計為400 MHZ,功耗為368.89 mW。視訊重建子單元還能在不增加資料傳輸頻寬下進一步平行化為最多25倍的子單元組。4倍平行的子單元組每秒能重建42幀300x400的視訊,供使用者即時預覽。 不同於其他研究者將彩色視訊壓縮感知簡化為三個獨立的壓縮感知問題,我們提出了一套彩色視訊壓縮感知流程。彩色視訊壓縮感知流程中,我們提出一個CCGVF-D3演算法交換各顏色通道的資訊,搭配先備字典有效的提昇重建的彩色視訊品質,對一般的視訊能有1~3.5 dB的峰值信噪比增益。 Video Compressive Sensing can increase the highest frame rate of video streams efficiently. In this thesis, we set our target of video compressive sensing to 8 times the frame rate. Balancing on both time and quality, our design achieves the video reconstruction rate of 30 fps to enable instant user preview. Furthermore, we extend the video compressive sensing to color video compressive sensing and propose a new color video compressive sensing flow for better video qualities. First, we test two currently available dictionary-base video compressive sensing algorithms. We choose to accelerate Gaussian-mixture-model-base video reconstruction algorithm since it provides better quality. However, the huge amount of computation and dependency of data degrade the utilization rate of the computing systems when implemented by software. Therefore, we propose a multilevel-parallel hardware design for video reconstruction acceleration. The basic module, video reconstruction subunit, is implemented subunit with a two-clock domain scheme as well as deep pipeline and parallel techniques. Using TSMC 90 nm technology, each subunit takes about 3.15 mm^2 in area, and its power consumption is 368.89 mW when operating at 400 MHz. The hardware can achieve 220 times speed up when compared to the original software version. Subunits can be further grouped into a super module of 25-parallel subunits group without increasing I/O bandwidth. The 4-parallel subunit group system can achieve 42 fps reconstruction rate required for instant user preview. Different from other researches which simplify the color video compressive sensing to a problem of 3 independent compressive sensing channels, we propose a new color video compressive sensing flow. In our flow, we propose a CCGVF-D3 algorithm exchanging information between channels. Accompanied with pre-learned dictionary, CCGVF-D3 algorithm can improve color video quality by 1~3.5 dB PSNR gain. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/78028 |
DOI: | 10.6342/NTU201700157 |
Fulltext Rights: | 有償授權 |
Appears in Collections: | 電子工程學研究所 |
Files in This Item:
File | Size | Format | |
---|---|---|---|
ntu-106-R03943052-1.pdf Restricted Access | 20.96 MB | Adobe PDF |
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.