Title Page
Contents
초록 8
Abstract 9
Ⅰ. Introduction 10
Ⅱ. Related work 17
2.1. Alignment-based methods 17
2.2. Pixel rejection-based methods 17
2.3. CNN-based methods 18
2.4. Transformer-based methods 19
2.5. Deformable Convolution Networks 19
Ⅲ. Proposed Method 21
3.1. Network Architecture 21
3.2. Deformable Fusion Block 23
3.2.1. Deformable Convolution Layer 23
3.2.2. Deformable Alignment Module 26
3.2.3. Depth-Wise Attention Module 27
3.3. Loss Function 29
Ⅳ. Experimental Results 30
4.1. Dataset and Implementation Details 30
4.1.1. Datasets 30
4.1.2. Evaluation metrics 33
4.1.3. Optimization details 33
4.2. Comparison with Other Methods 34
4.2.1. Datasets with Ground Truth 34
4.2.2. Datasets without Ground Truth 36
4.3. Ablation Study 39
4.4. Comparison on the Computational Cost 43
Ⅴ. Conclusion 44
Bibliography 45