超解析度方法與系統設計比較研究;Comparative Study for Super-Resolution Methods and System Design

NCUIR > College of Electrical Engineering & Computer Science > Graduate Institute of Computer Science and Information Engineering > Electronic Thesis & Dissertation > Item 987654321/75306

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/75306

Title:	超解析度方法與系統設計比較研究;Comparative Study for Super-Resolution Methods and System Design
Authors:	張翔珳;Chang, Hsiang-Wen
Contributors:	資訊工程學系
Keywords:	超解析度;SRCNN;APNN;Bicubic spline;比較研究;Super resolution;SRCNN;APNN;Bicubic spline;Comparison Study
Date:	2017-11-30
Issue Date:	2018-01-16 11:04:05 (UTC+8)
Publisher:	國立中央大學
Abstract:	在視訊監控、視覺檢測領域，傳統攝影機因為解析度不足，造成監控品質、檢測率降低，而超解析度方法能夠超越攝影機物理極限，將原低解析度影像插補轉換為高解析度影像，增加系統應用的效能。本論文以三種基於不同原理的超解析度方法Bicubic spline、APNN、SRCNN作為比較的對象，以人臉辨識，條碼辨識來評估和量化監控品質、檢測率的改善程度，以及針對演算法實作後成本進行全面評估，比較每個演算法在人眼視覺品質、記憶體使用量、即時性、適合硬體化、相對耗電量、硬體資源的成本，以提供嵌入式系統開發在選擇超解析度方法時，能夠有量化數據做為參考。從實驗結果得知效果方面，超解析度方法能夠改善系統的辨識率，其中以SRCNN改善幅度最大，超解析度方法也進一步降低資料傳輸頻寬，在成本方面，Bicubic spline有少量資源和計算速度快的特性，適合實作於嵌入式軟硬體，APNN需要相對較多的資源，適合實作於嵌入式硬體，SRCNN需要龐大的硬體資源，故目前僅適合實作於GPU平台，其中，本論文對SRCNN硬體部分做硬體最佳化，在邏輯閘數量和精準度做取捨，達到演算法近似計算，本論文找到兩組解，將SRCNN常數權重乘法器硬體化簡，一組為降低0.79%的邏輯閘，平均特徵差增加0.000120，一組為降低12.2%的邏輯閘，平均特徵差增加0.264911。;In the field of video surveillance and vision detection, traditional cameras have low video quality and detection rates, owing to their low resolution. To increase the effectiveness of system applications, the super-resolution method was developed to convert low-resolution images into high-resolution images. In this paper, three super-resolution methods (Bicubic spline, APNN, and SRCNN) are compared and evaluated through two experiments to assess their performance in areas such as face and barcode recognition. The properties of each algorithm, including human vision quality, memory usage, execution time, hardware complexity, relative power consumption, and hardware resources, are also discussed. According to the experimental results, system analyzers can choose the appropriate super-resolution method for embedded system development. Our results show that SRCNN has the greatest improvement of recognition rate among the super-resolution methods. In terms of cost, the Bicubic spline method is suitable for embedded software and hardware applications due to its low cost and high speed. APNN can be applied in embedded hardware applications because of its low resource usage. Due to the high resource usage of SRCNN, it is only suitable for certain GPUs. Therefore, we use a genetic algorithm to obtain a trade-off between hardware cost and accuracy to compute an approximation for the SRCNN hardware. We found two approximated results for SRCNN: a 0.79% decrease in hardware cost and an increase of 0.000120 in average feature difference or a 12.2% decrease in hardware cost and an increase of 0.264011 in average feature difference.
Appears in Collections:	[Graduate Institute of Computer Science and Information Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	281	View/Open

社群 sharing

Loading...