林巍聳臺灣大學：電機工程學研究所方俊雄Fang, Chun-HsiungChun-HsiungFang2007-11-262018-07-062007-11-262018-07-062007http://ntur.lib.ntu.edu.tw//handle/246246/53105本論文的主旨是用演算模型仿製人類視覺之意圖視注覺和自動辨識功能。「意圖視注覺演算模型」可以從影像萃取意圖特徵並找出注目點，色彩意圖和目標物意圖用意圖向量和拓墣關係聯合表示，藉由匹配光刺激數據和意圖向量萃取意圖特徵，再藉由模糊推論法計算意圖特徵的拓墣關係吻合度找出注目點。「合成式仿射不變函數」可以把經過各種仿射轉換的二維形體變成唯一性的表示，本質上，「合成式仿射不變函數」是用「合成式特徵信號」所推導出來的參數化輪廓信號。輪廓信號的「合成式特徵信號」可以用小波轉換、傅力葉轉換或餘弦轉換推演出來，特點是「合成式仿射不變函數」所採用的「合成式特徵信號」必須涵蓋輪廓信號中沒有平移成份的全部成份信號，因此在自動辨識的程序中，「合成式仿射不變函數」可以準確地代表原始的輪廓信號。經由實驗數據比較數種仿射不變函數的不變性、代表性、雜訊耐受性和辨識正確率的結果，「合成式仿射不變函數」確定最為傑出，其中又以採用二元小波轉換法取得四個「合成式特徵信號」的「合成式仿射不變函數」最適合實際應用。This dissertation addresses the computational implementation of intentional attention and automated recognition in the human vision. An Intention-oriented Computational Visual Attention (ICVA) model is proposed to extract intentional features from images and eventually to find out attentive spots. An intention about favorite colors or target objects is expressed as a combination of preference vectors and topological relationships. Intentional features are extracted from the optical excitation data by matching the preference vectors. Attentive spots are found by fuzzy inference to fulfill the topological relationships. Synthesized Affine Invariant Function (SAIF) is derived to uniquely represent two-dimension shapes subject to affine transformations. SAIF is essentially a parameterized contour signal derived with the Synthesized Feature Signals (SFSs). It is shown Wavelet, Fourier, or Fourier cosine transforms can be used to deduce SFSs from a contour signal. Specifically, SAIF uses SFSs involving every component except translation of the contour signal. Hence, SAIF has minimum information loss in representing the contour signal for automated recognition. The invariance property, representative property, robustness against noise and correctness in the shape recognition of several significant SAIFs are compared experimentally. It is shown SAIFs outperform other affine invariant functions and the SAIF using four SFSs calculated with the dyadic wavelet transform is the most practical implementation.誌謝 i 摘要 iii Abstract v Table of Contents vii List of Figures xi List of Tables xv List of Abbreviation xvii Chapter 1 Introduction 1 1.1 Background 1 1.1.1 Human Vision 2 1.1.2 Computer Vision 7 1.1.3 Literature review 13 1.2 Motivation 24 1.3 Main Contributions 29 1.4 Dissertation Organization 30 Chapter 2 Visual Attention 33 2.1 Chapter Introduction 33 2.1.1 Chapter outline 34 2.2 The Neurobiology of Human Visual System 35 2.2.1 Overview of the human visual system 36 2.2.2 The eye and the optic chiasm 37 2.2.3 The thalamus 45 2.2.4 The visual cortex 48 2.2.5 The pathways in the visual cortex 52 2.2.6 The Superior Colliculus (SC) and the Frontal Eye Field (FEF) 54 2.2.7 The human visual attention 55 2.3 Psychophysical Models of Visual Attention 57 2.3.1 Feature integration theory of attention 58 2.3.2 Guided Search 60 2.3.3 Other models 63 2.4 Computational Models of Visual Attention 66 2.4.1 Koch & Ullman’s model 66 2.4.2 Milanese’s model 68 2.4.3 Itti’s model 70 2.4.4 Discussion on bottom-up and top-down features 74 2.5 Chapter Summary 77 Chapter 3 Intention-Oriented Computational Visual Attention 81 3.1 Chapter Introduction 81 3.1.1 Chapter outline 82 3.2 Intention for Visual Attention 82 3.3 Model of Intention-Oriented Computational Visual Attention 88 3.3.1 Overview of the model 89 3.3.2 The feature extraction process 91 3.3.2.1 The color space 92 3.3.2.2 The chromatic opponent bases 95 3.3.2.3 The habitual features in the ICVA model 97 3.3.2.4 The intentional feature in the ICVA model 102 3.3.3 The conspicuity mapping process 108 3.3.3.1 The details of the conspicuity mapping process109 3.3.3.2 Conspicuity maps 119 3.3.4 The fuzzy classification process 123 3.3.5 The attention image 126 3.4 Experimental Results 128 3.4.1 Experiment 1: Color-oriented visual attention 128 3.4.2 Experiment 2: Object-oriented visual attention130 3.4.3 Experiment 3: M-objects-oriented visual attention 132 3.5 Comparison between ICVA model and CVA model 134 3.6 Chapter Summary 136 Chapter 4 Synthesized Affine Invariant Functions for 2D Shape Recognition 139 4.1 Chapter Introduction 139 4.1.1 Chapter outline 142 4.2 State of the Art of Affine Invariant Functions 143 4.3 Affine Transformation and Affine Invariants 148 4.4 Wavelet-Based Synthesized Affine Invariant Functions 150 4.4.1 Wavelet-based synthesized feature signals 151 4.4.2 Formulation of Wavelet-based Synthesized Affine Invariant Functions 155 4.4.2.1A WSAIF with two pairs of WSFSs 155 4.4.2.2 A WSAIF with three pairs of WSFSs 156 4.5 Fourier-Based Synthesized Affine Invariant Functions 158 4.5.1 Fourier-based synthesized feature signals 158 4.5.2 Formulation of Fourier-based Synthesized Affine Invariant Functions... 160 4.6 Cosine-Based Synthesized Affine Invariant Functions 162 4.6.1 Cosine-based synthesized feature signals 162 4.6.2 Formulation of Cosine-based Synthesized Affine Invariant Functions... 163 4.7 Characterizing SFSs by Specifying the Weight Vector 165 4.8 Performance of SAIFs in Applications of Shape Recognition 167 4.8.1 The Root Mean Square errors (RMS) and the Local Invariance Measure (LIM) 169 4.8.2 The Representative Index (RI) 169 4.8.3 Discrimination function of shape recognition 170 4.9 Experiment I – Comparison between SAIFs and Previous AIFs 171 4.9.1 Comparing CSAIF with FAIF 171 4.9.1.1 Experiment I-1: Performance of the CSAIF-based shape recognition 171 4.9.1.2 Experiment I-2: Robustness of the CSAIF-based shape recognition 173 4.9.2 Comparing WSAIFs with previous AIFs 175 4.9.2.1 Experiment I-3: Invariance property of WSAIFs 176 4.9.2.2 Experiment I-4: Representative property of WSAIFs 178 4.9.2.3 Experiment I-5: Correctness of shape recognition with WSAIFs and previous AIFs 180 4.10 Experiment II – Comparisons between SAIFs 182 4.10.1.1 Experiment II-1: Invariance property of SAIF with different weighting types 184 4.10.1.2 Experiment II-2: Representative property of SAIF 187 4.10.1.3 Experiment II-3: Performance of SAIF in the correctness of shape recognition 188 4.11 Chapter Summary 191 Chapter 5 Conclusions 195 Bibliography 203 Publication List 2173912954 bytesapplication/pdfen-US演算視覺視注覺意圖仿射不變函數形狀辨識小波轉換傅力葉轉換餘弦轉換computational visionvisual attentionintentionaffine invariant functionshape recognitionwavelet transformFourier transformcosine transformthesishttp://ntur.lib.ntu.edu.tw/bitstream/246246/53105/1/ntu-96-F91921049-1.pdf