Fig. 1. An example of stationary wavelet transform using db2 in internal dataset. (a) Original image; (b)-(d) horizontal, vertical, diagonal high frequency of 1-level SWT; (e)-(g) horizontal, vertical, diagonal high frequency of 2-level SWT
Fig. 2. Demonstration of overlapping separating
Fig. 3. Framework of proposed super-resolution method
Fig. 4. Visual comparison of 3×upscaling reconstruction effect. (a) Original image; (b) bicubic interpolation; (c) method in Ref. [5]; (d) method in Ref. [6]; (e) SRCNN; (f) proposed method (with haar)
Window size /(pixel×pixel) | PSNR | SSIM | RLNE | EPI |
---|
2×2 | 28.4266 | 0.9939 | 0.0232 | 1.3792 | 3×3 | 28.5170 | 0.9941 | 0.0228 | 1.3618 | 5×5 | 28.3959 | 0.9935 | 0.0234 | 1.3691 | 7×7 | 28.2219 | 0.9933 | 0.0239 | 1.3858 |
|
Table 1. Comparison of different patch sizes (3×upscaling, dictionary size 500, clustering number 3, wavelet base db2)
Dictionarysize | PSNR | SSIM | RLNE | EPI |
---|
2000 | 28.6238 | 0.9945 | 0.0222 | 1.4045 | 1000 | 28.6330 | 0.9945 | 0.0222 | 1.3779 | 500 | 28.5170 | 0.9941 | 0.0228 | 1.3618 | 100 | 28.2422 | 0.9928 | 0.0242 | 1.3657 |
|
Table 2. Comparison of different dictionary sizes (3×upscaling, patch size 3 pixel×3 pixel, clustering number 3, wavelet base db2)
Clusteringnumber | PSNR | SSIM | RLNE | EPI |
---|
3 | 28.6330 | 0.9945 | 0.0222 | 1.3779 | 4 | 28.5408 | 0.9941 | 0.0226 | 1.3657 | 5 | 28.5996 | 0.9943 | 0.0223 | 1.3807 |
|
Table 3. Comparison of different clustering numbers (3×upscaling, patch size 3 pixel×3 pixel, dictionary size 1000, wavelet base db2)
Index | haar | db2 | db3 | db4 | sym2 | sym3 | sym4 | sym6 | sym8 | coif1 | coif2 | dmey |
---|
PSNR | 28.7574 | 28.6330 | 28.5366 | 28.6411 | 28.5444 | 28.6124 | 28.5836 | 28.5155 | 28.4995 | 28.5814 | 28.603 | 28.3491 | SSIM | 0.9949 | 0.9945 | 0.9941 | 0.9943 | 0.9941 | 0.9943 | 0.9945 | 0.9942 | 0.9941 | 0.9945 | 0.9941 | 0.9938 | RLNE | 0.0215 | 0.0222 | 0.0227 | 0.0221 | 0.0226 | 0.0223 | 0.0224 | 0.0228 | 0.0228 | 0.0224 | 0.0223 | 0.0237 | EPI | 1.3583 | 1.3657 | 1.3728 | 1.3576 | 1.3690 | 1.3561 | 1.3675 | 1.3671 | 1.3777 | 1.3701 | 1.3575 | 1.3766 |
|
Table 4. Comparison of different wavelet bases (3×upscaling, patch size 3 pixel×3 pixel, dictionary size 1000, clustering number 3)
Algorithm | MRI | DCE-MRI |
---|
PSNR | SSIM | EPI | RLNE | PSNR | SSIM | EPI | RLNE |
---|
Bicubic interpolation | 30.1714 | 0.8817 | 0.4475 | 0.0348 | 32.4535 | 0.8854 | 0.5003 | 0.0094 | Ref. [5] | 31.2093 | 0.8945 | 0.5570 | 0.0274 | 33.8091 | 0.8923 | 0.5718 | 0.0070 | Ref. [6] | 31.5505 | 0.9025 | 0.5641 | 0.0254 | 34.2530 | 0.9059 | 0.6290 | 0.0062 | Ref. [7] | 31.7256 | 0.8984 | 0.5820 | 0.0330 | 34.0285 | 0.8698 | 0.6237 | 0.0148 | db2 | 31.7225 | 0.9040 | 0.5721 | 0.0244 | 34.3888 | 0.9066 | 0.6343 | 0.0060 | haar | 31.7220 | 0.9036 | 0.5740 | 0.0245 | 34.5002 | 0.9072 | 0.6343 | 0.0058 | sym4 | 31.7052 | 0.9041 | 0.5705 | 0.0244 | 34.3643 | 0.9066 | 0.6356 | 0.0061 | db2_IBP | 31.8756 | 0.9089 | 0.6141 | 0.0238 | 34.6785 | 0.9135 | 0.6831 | 0.0056 | haar_IBP | 31.8918 | 0.9088 | 0.6147 | 0.0240 | 34.5985 | 0.9104 | 0.6897 | 0.0057 | sym4_IBP | 31.8480 | 0.9089 | 0.6155 | 0.0237 | 34.4396 | 0.9094 | 0.6897 | 0.0060 |
|
Table 5. Comparison of 3×upscaling reconstruction effect of different methodsdB