請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/68316
標題: | 稀疏表示應用於衛星影像全色態銳化之方法 Pan-sharpening of Multispectral Images Using Sparse Representation |
作者: | Yu-Hsuan Lin 林妤宣 |
指導教授: | 徐百輝(Pai-Hui Hsu) |
關鍵字: | 全色態銳化,衛星影像融合,稀疏表示,聯合字典訓練, Pan-sharpening,Image fusion,Sparse representation,Joint dictionary learning, |
出版年 : | 2020 |
學位: | 碩士 |
摘要: | 多數遙測衛星受限於光譜分光成像技術及有限的儲存空間,僅能分別提供具 高空間解析度的全色態影像,以及具光譜資訊的多光譜影像。全色態銳化的目的 即在於結兩種影像各自的優勢,另外產製一具有高空間解析度的多光譜影像,除 了可提升視覺上效果外,亦可增加後續應用的可用性。在過去已有許多針對全色 態銳化的演算法被提出,但均存有不同層次的問題,有些可能造成光譜失真,有 些則產生錯誤的空間資訊,如假影或殘影。近年基於稀疏表示及壓縮感測理論的 影像超解析技術,在提升影像解析度取得相當不錯的成果,因而被擴充應用至全 色態銳化中,其主要以來源影像與銳化後影像間的關係模型為基礎,配合稀疏表 示理論重建銳化後影像,然此類方法除需掌握影像間的關係模型外,更因直接產 製銳化後影像的計算複雜程度高,多含銳化效率不彰之問題。 本研究為克服上述問題,提出一結合稀疏表示與細節注入的全色態銳化方法。 分別以全色態影像紋理及平滑程度與多光譜相當之全色態影像紋理,作為高、低 解析度先驗資訊,學習其間的對應關係,建構在相同稀疏數係下能分別對應至高 低解析度紋理之成對字典,配合 Feature Sign Searching 最佳化演算法,經由字典 抽換的方式,逐步恢復各區塊影像高解析度紋理,最終注入經升取樣之多光譜影 像中,以達全色態銳化目標。本研究除能有效提升銳化效率外,更加入相鄰區塊 影像紋理一致性的約束條件,及自動化產製平滑化全色態影像,同時提升紋理重 建成果及流程自動化程度。以 WorldView-2 及 WorldView-3 的衛星影像,與多種 經典銳化方法比較,並利用數個空間及光譜品質指標評估銳化之成效,經實驗證 實提出之方法能更好的保有光譜資訊,並提供相當完整的空間細節資訊。 Pan-sharpening is a fusion technique used to increase the spatial resolution of the multispectral image while simultaneously preserving its spectral information. Considerable research has been devoted to remote image fusing technique during the past years such as IHS, PCA and wavelet transform etc. However, these traditional methods suffer from the problem of spectral distortion or causing ringing and aliasing problem in fused images. Recently, in the field of image processing a technique called super resolution has been proposed and proved to effectively improve the resolution of dynamic or static images. Numerous studies have brought this concept to pan- sharpening. Among many popular methods based on super resolution concept, the result of pan-sharpening with sparse representation (SR) is superior than others. Most of these SR based methods directly reconstruct the whole image through SR theory by defining reconstruct model between source images and sharpened image, which performance depend on model accuracy and suffer from poor time efficiency. Considering the above problems, a pan-sharpening method base on sparse representation theory and detail injection is proposed to both accelerate fusion efficiency and improve fusion result. Using the texture from Pan and its lower resolution version as high- and low-resolution priori information to train a pair of coupled texture dictionaries. With respect to the low-resolution texture dictionary, the feature sign searching (FSS) algorithm is used to resolve each patch's sparse coefficient. Through dictionary replacement, the high-resolution texture can be represented by the same coefficient and finally inject to up-sampling MS. In addition to improving efficiency, the proposed method also added adjacent patch constraints and automated production of low-resolution Pan for the dictionary training, while improving both reconstruction accuracy results and process automation. The proposed method is compared with several conventional methods using spatial and spectral quality measurement indices. The experimental results demonstrate that the proposed algorithm gives less spectral distortion and preserves spatial details of the source image in final product. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/68316 |
DOI: | 10.6342/NTU202003759 |
全文授權: | 有償授權 |
顯示於系所單位: | 土木工程學系 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
U0001-1708202014561300.pdf 目前未授權公開取用 | 14.43 MB | Adobe PDF |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。