莊永裕Chuang, Yung-Yu臺灣大學:資訊工程學研究所謝毓庭Hsieh, Yu-TingYu-TingHsieh2010-06-022018-07-052010-06-022018-07-052008U0001-1806200816345700http://ntur.lib.ntu.edu.tw//handle/246246/184951此篇論文主要的研究，是將影像的背景資訊加入一般物體辨識的流程，以提升其準確率。目前大部分的研究並未將影像的前景物體與背景分開考慮，或者只利用前景的資訊。在這一篇論文中，我們試著加入背景資訊以提過一般物體辨別的準確率。們使用一個偵測使用者感興趣區域（Region of Interest）的方法來將影像前景的物體偵測出來。更進一步地，使用者感興趣區域周圍的背景資訊可以用來加強物體識別。由於同一個種類的物體通常會出現在某些特定的場合，我們將由實驗說明加入背景資訊對一般物體辨識率的提升。一個很有挑戰性的問題是如果將不同的影像特徵合併使用。我們比較了幾個不同的方法在支持向量機（Support Vector Machine）上的表現。實驗結果顯示這些方法在這個問題上的好壞，與他們能否有效運用背景資訊來加提升辨識率。This thesis introduces background information to generic object recognition problem to increase the accuracy. Most of works do not divide images to foreground and background part, or only utilize foreground information. In this thesis, we tried to leverage background information to help object recognition. region of interest (ROI) detector is used to find the foreground object in images. Focusing on foreground object can reduce noisy features from unrelevant background region. Furthermore, the complement area of ROI can be considered as background context. Since objects in a category usually appear in specific context, we will show that adding background clue can improve the recognition accuracy in our experiment.nother challenge problem is how to use different signals together. We compared several methods of feature fusion for machine learning using SVM. Experiment result shows how well these methods can achieve and whether background information benefit them.Acknowledgments iiibstract vist of Figures xiist of Tables xiiihapter 1 Introduction 1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2 Contribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4hapter 2 Related Work 5.1 Feature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.2 ROI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.3 Feature Fusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7hapter 3 Feature Extraction 9iii.1 Grid of Pyramid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.2 Pyramid of Histogram . . . . . . . . . . . . . . . . . . . . . . . . . . 11.3 Pyramid Match Kernel . . . . . . . . . . . . . . . . . . . . . . . . . 11.4 Foreground Representation . . . . . . . . . . . . . . . . . . . . . . . 12hapter 4 Region of Interest 15.1 ROI for Object Detection . . . . . . . . . . . . . . . . . . . . . . . . 15.1.1 Low-level Feature-based Exhaustive Search . . . . . . . . . . 16.1.2 Learning-based Detection with Visual Cue . . . . . . . . . . 17.2 Apply ROI to Classification Problem . . . . . . . . . . . . . . . . . . 20.2.1 Background Representation . . . . . . . . . . . . . . . . . . 21hapter 5 Supervised Learning of Categories 23.1 Support Vector Machine . . . . . . . . . . . . . . . . . . . . . . . . 23.2 Feature Fusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24.2.1 Averaged Kernel . . . . . . . . . . . . . . . . . . . . . . . . 24.2.2 Ensemble Learning . . . . . . . . . . . . . . . . . . . . . . . 24.2.3 Adaptive Grid Search of Weighting . . . . . . . . . . . . . . 26.2.4 Super Kernel Fusion . . . . . . . . . . . . . . . . . . . . . . 27hapter 6 Experiment 29.1 Caltech 101 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29.2 Feature Extraction in ROI . . . . . . . . . . . . . . . . . . . . . . . . 30.3 Feature Fusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30.3.1 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . 31.4 Example of Result Image . . . . . . . . . . . . . . . . . . . . . . . . 32hapter 7 Conclusion 33ibliography 34application/pdf732594 bytesapplication/pdfen-US一般物體辨識背景Generic object recognitionbackgroundthesishttp://ntur.lib.ntu.edu.tw/bitstream/246246/184951/1/ntu-97-R95922017-1.pdf