Please use this identifier to cite or link to this item:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/74453
Title: | 基於深度學習之超解析度即時系統架構設計 Efficient Deep Learning based Super-Resolution Real-Time System |
Authors: | Po-Wei Yen 嚴柏煒 |
Advisor: | 簡韶逸(Shao-Yi Chien) |
Keyword: | 超解析度,深度學習, Super-Resolution,Deep Learning, |
Publication Year : | 2019 |
Degree: | 碩士 |
Abstract: | 影像超解析度的目標,是將輸入的低解析度影像,產生出含有銳利邊緣和豐富細節的高解析度影像。由於現今顯示器進步的速度飛快,影像超解析度漸漸成為了一項不可或缺的技術。
影像超解析度是一個困難的非單一解問題,因為一張低解析度影像,可能由多張不同的高解析度影像產生。為了解決這個問題,以深度學習為基底的單一影像超解析度演算法,透過數以萬計來自外部的低解析度對高解析度圖塊配對,學習其中的對應關係,而達到了出色且先進的成果和表現。 在本篇論文中,我們提出了一個改良版深度學習網路,在所有參數量都為12K的網路中,表現是最好的,及基於這個網路的即時超解析度系統架構。此網路首先優化了Information Distillation Network,利用sub-pixel convolution讓整個網路只透過3×3卷積就能實現。高效率的架構則是透過Winograd快速卷積完成。該系統通過Synopsys 32奈米技術的驗證,在200MHz的頻率運作,可達到每秒輸出60張1920×1080 (full HD)影像。有了我們提出的及時超解析度系統,就能將超解析度實作在顯示面板的驅動晶片上,改善顯示系統上的影像問題。 Image super-resolution aims to generate good quality high-resolution images with sharp edges and rich details information from input low-resolution images. Due to the rapid growth of the display resolution on edge devices, super-resolution has become an essential technique in many applications. Super-resolution is a well-known ill-posed problem since a single low-resolution image could be generated from more than one high-resolution images. To solve this problem, deep learning-based single image super-resolution methods have achieved outstanding performance and gained state-of-the-art results, by learning an end-to-end mapping function from millions of low-resolution (LR) to high-resolution (HR) image pairs. We proposed an improved deep-learning architecture that generates highest quality image among other 12K parameters networks as well as a real-time deep- learning based super-resolution system. The proposed architecture inferences a deep-learning model that generate high- resolution images from low-resolution images. This real-time super-resolution system makes it possible to integrate super-resolution operation into display driver carrying out computational photography in display system. The system is verified with Synopsys 32nm educational design kit technology which achieves output resolution of 1920 × 1080 full-HD resolution at 60 frames per second (fps). The proposed system is working at a clock frequency of 200 MHz. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/74453 |
DOI: | 10.6342/NTU201902878 |
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-108-1.pdf Restricted Access | 16.55 MB | Adobe PDF |
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.