Title Page
Contents
ABSTRACT 11
Ⅰ. INTRODUCTION 13
1. Background 13
2. Goals 17
3. Contributions 18
4. Definition of terms 19
A. Decision mask 19
B. Mask sparsity 19
C. Pixel of high-frequency (PHF) and Pixel of low-frequency (PLF)[이미지참조] 19
D. High parameter convolution (HPC) and low parameter convolution (LPC) 19
E. Spatial convolution 20
F. Point-wise convolution 20
G. Sparse convolution 20
H. Dilation of decision mask 20
Ⅱ. LITERATURE REVIEW 21
1. Overview of image restoration 21
2. Standard evaluation process 24
3. Efficient super-resolution models 26
4. Analysis of image frequency in super-resolution 29
Ⅲ. METHODOLOGY 32
1. Frequency-based decision mask 32
2. Implementing double networks for FLOPs reduction 35
3. Double sparse convolutional operation 37
4. Morphological dilation of decision mask 42
Ⅳ. EXPERIMENTS 46
1. Environment, settings, and standards for comparison 46
2. Results achieved with double sparse convolution 47
3. Preventing performance degradation with dilation 50
4. Generalizing double sparse convolutional operation 53
5. Comparison with other light-weight models 55
6. Visualization of results 58
7. Single model implementation 60
Ⅴ. CONCLUSION 62
1. Conclusion 62
2. Future works 63
REFERENCES 64
국문요약 71
Table 1. Comparative results achieved with different thresholds 47
Table 2. Performance comparison among different dilation kernel sizes 50
Table 3. Quantitative results of super-resolution algorithms 55
Figure 1. A deep learning network for image super-resolution 14
Figure 2. High-frequency and low-frequency within an image 15
Figure 3. Evaluation process in super-resolution 24
Figure 4. Visualization of decision masks 33
Figure 5. The receptive-field problem in double networks 35
Figure 6. Double sparse convolutions for FLOPs reduction 37
Figure 7. Dilating decision masks 43
Figure 8. Increase of mask's sparsity with respect to threshold 44
Figure 9. The baseline model for experiments 47
Figure 10. Visualization of results achieved with different thresholds 48
Figure 11. Visual comparison of results by various dilation kernel sizes 51
Figure 12. Problem of mask dilation when operating with high sparsity 51
Figure 13. Implementation on other models 53
Figure 14. Performance and FLOPs comparison among light-weight models 56
Figure 15. Visual comparison of results generated with different thresholds 58
Figure 16. Visual comparison among state-of-the art models 59
Figure 17. Single and multiple models of various thresholds 60