指導教授：貝蘇章臺灣大學：電信工程學研究所陳立恆Chen, Li-HengLi-HengChen2014-11-302018-07-052014-11-302018-07-052014http://ntur.lib.ntu.edu.tw//handle/246246/264267如何準確的評估影像的品質在影像處理的研究及應用扮演著舉足輕重的腳色，傳統測量影像品質的方法是比較原始影像與失真影像的峰值信噪比(PSNR)，但近年來許多研究發現PSNR所測量出來的影像失真度和人眼視覺系統(HVS)有所不符，換句話說，我們使用的PSNR所測量出來的影像品質好壞和人眼所感知的影像品質是不一致的，因此我們需要發展新的影像品質測量演算法來解決這個問題。 首先我們會以一些淺顯易懂的例子來說明PSNR不符合HVS的事實，接著會介紹一些其他人提出的方法，如SSIM, FSIM, MSVD等。最後提出我們利用機器學習的方法結合已存在的不同頻帶的影像品質特徵和色彩特徵，訓練出和人眼視覺系統有高度相關性的模型。此外，我們也試著以Saliency map 來改善現有的影像品質估測方法，對此我們也做了一系列的實驗。The objective image quality assessment (IQA) plays a key role in the development of various multimedia applications. The release of new IQA dataset (TID2013) challenges the wildly used 2D IQA metrics (e.g. PSNR and SSIM) since they cannot handle the diversity of distortion types. In this thesis, we propose a machine learning approach IQA model with features extracted from different frequency band (DOG features). The color distortion is also considered in our system. The effectiveness of our IQA system is verified by comparing with subjective score on the available databases. The experimental results show the high consistency between MOS score and our metric.口試委員會審定書 ........................................................................................................... # 誌謝 .................................................................................................................................. ii 中文摘要 ..........................................................................................................................iv ABSTRACT .....................................................................................................................vi CONTENTS .................................................................................................................. viii LIST OF FIGURES ........................................................................................................ xii LIST OF TABLES ..........................................................................................................xvi Chapter 1 Introduction .............................................................................................. 1 Chapter 2 Related Image Quality Assessment......................................................... 3 2.1 Introduction..................................................................................................... 3 2.1.1 Subjective score and Objective score .................................................... 3 2.1.2 Performance evaluation ......................................................................... 5 2.2 Image quality assessment in mean square error sense .................................. 10 2.2.1 Peak signal to noise ratio (PSNR) ....................................................... 11 2.2.2 Problems of the mean square error based image quality assessment .. 12 2.3 Structure similarity image quality assessment (SSIM) ................................. 17 2.3.1 Introduction to structure similarity index ............................................ 17 2.3.2 Properties of SSIM index .................................................................... 20 2.3.3 Image quality assessment using SSIM index ...................................... 20 2.3.4 Comparison and experimental results ................................................. 23 2.3.5 Other applications of SSIM ................................................................. 25 2.3.6 Multi-scale structural similarity .......................................................... 26 2.4 SVD based image quality assessment........................................................... 28 2.4.1 Singular value decomposition ............................................................. 28 2.4.2 Image quality metric with singular values (M-SVD) .......................... 30 2.4.3 Image quality metric with singular vectors (V-SVD) ......................... 32 2.5 Feature similarity image quality assessment ................................................ 36 2.5.1 Structure information of FSIM index .................................................. 36 2.5.2 Framework of FSIM metric ................................................................ 41 Chapter 3 Machine learning based IQA model ..................................................... 45 3.1 Machine learning based image quality assessment ...................................... 45 3.1.1 System overview ................................................................................. 45 3.1.2 Features for image quality assessment ................................................ 46 3.1.3 Support Vector Regression .................................................................. 52 3.1.4 Implementation details ........................................................................ 53 3.1.5 Experimental results on TID2013 database ........................................ 60 3.1.6 Experimental results on other databases ............................................. 64 3.2 Image quality metric with visual saliency consideration .............................. 66 3.2.1 Saliency map ....................................................................................... 66 3.2.2 Visual saliency considered image quality metrics............................... 68 Chapter 4 IQA metric model with multi band extracted features ....................... 75 4.1 Difference of Gaussian ................................................................................. 75 4.1.1 Difference of Gaussian ........................................................................ 75 4.1.2 Multi band decomposition by DOG .................................................... 76 4.1.3 Perfect reconstruction in DOG decomposition ................................... 79 4.2 Feature extraction ......................................................................................... 80 4.2.1 Multi band feature extraction .............................................................. 80 4.2.2 Feature extraction with mixture metric module .................................. 85 4.3 Experimental results ..................................................................................... 85 4.3.1 Experimental results on TID2013 database ........................................ 86 4.3.2 Experimental results on other databases (LIVE and CSIQ) ................ 89 4.3.3 Experimental results on color distorted subset (TID2013) ................. 97 Chapter 5 Conclusion and future work ............................................................... 101 5.1 Conclusion .................................................................................................. 101 5.2 Future work ................................................................................................. 101 REFERENCE ................................................................................................................ 1039431063 bytesapplication/pdf論文公開時間：2016/07/22論文使用權限：同意有償授權(權利金給回饋本人)人眼視覺系統影像品質評估結構相似性支持向量機thesishttp://ntur.lib.ntu.edu.tw/bitstream/246246/264267/1/ntu-103-R01942037-1.pdf