Author Affiliations
1School of Physics and Electronic Information Engineering, Qinghai Nationalities University, Xining, Qinghai 810007, China2School of Electrical and Information Engineering, Tianjin University, Tianjin 300072, Chinashow less
Fig. 1. Blurry image, clean image, and edge-weakened image. (a) Blurry image; (b) clean image; (c) edge-weakened image learned by PNet
Fig. 2. Structure of proposed network
Fig. 3. Diagram of DNet subnet generator for image deblurring
Fig. 4. Dense residual block
Fig. 5. Diagram of PNet subnet discriminator (PatchGAN) for image deblurring
Fig. 6. Results of image deblurring of compared methods on test dataset of GOPRO. (a) Blurry images; (b) method in Ref. [11]; (c) method in Ref. [13]; (d) method in Ref. [15]; (e) method in Ref. [16]; (f) ours
Fig. 7. Results of image deblurring of compared methods on dataset of K?hler. (a) Blurry images; (b) method in Ref. [11]; (c) method in Ref. [13]; (d) method in Ref. [15]; (e) method in Ref. [16]; (f) ours
Fig. 8. Results of deblurring of compared methods for real blurred images. (a) Blurry images;(b) method in Ref. [11]; (c) method in Ref. [13];(d) method in Ref. [15];(e) method in Ref. [16];(f) ours
Fig. 9. Visual results of subnetworks on GOPRO test set. (a) Blurry input; results of (b) w/o content, (c) w/o edge, (d) w/o adv, (e) w/o PNet, and (f) proposed method
Method | GOPRO | Köhler |
---|
PSNR | SSIM | PSNR | SSIM |
---|
Method in Ref. [11] | 27.2778 | 0.8187 | 21.2371 | 0.6490 | Method in Ref. [13] | 28.3225 | 0.8588 | 21.2335 | 0.6525 | Method in Ref. [15] | 25.2363 | 0.7773 | 20.8507 | 0.6340 | Method in Ref. [16] | 27.8086 | 0.8564 | 19.0843 | 0.5838 | Ours | 29.2278 | 0.8779 | 21.2987 | 0.6544 |
|
Table 1. Quantitative evaluation results of proposed method and compared methods on GOPRO and K?hler datasets
Method | PSNR | SSIM |
---|
w/o PNet | 28.8856 | 0.8687 | Ours | 29.2278 | 0.8779 |
|
Table 2. Quantitative evaluation results on dataset of GOPRO with different subnetworks
Method | PSNR | SSIM |
---|
w/o content | 26.5778 | 0.8034 | w/o edge | 28.4260 | 0.8418 | w/o adv | 28.5863 | 0.8513 | Ours | 29.2278 | 0.8779 |
|
Table 3. Quantitative evaluation results on dataset of GOPRO for different loss functions
Method | FLOPs /109 | Average time /s |
---|
Method in Ref. [11] | 4.12 | 1300.0 | Method in Ref. [13] | 1760.04 | 8.1 | Method in Ref. [15] | 678.29 | 1.1 | Method in Ref. [16] | 411.34 | 0.7 | Ours | 628.03 | 1.3 |
|
Table 4. Quantitative evaluation results of proposed method and compared methods on dataset of GOPRO