Joint Deep Denoising Prior for Image Blind Deblurring

Aiping Yang; Jinbin Wang; Bingwang Yang; Yuqing He

doi:10.3788/AOS201838.1010003

Journals >Acta Optica Sinica >Volume 38 >Issue 10 >Page 1010003 > Article

Acta Optica Sinica
Vol. 38, Issue 10, 1010003 (2018)

Joint Deep Denoising Prior for Image Blind Deblurring

Aiping Yang^*, Jinbin Wang, Bingwang Yang, and Yuqing He

Author Affiliations

School of Electrical and Information Engineering, Tianjin University, Tianjin 300072, China

show less

DOI: 10.3788/AOS201838.1010003 Cite this Article Set citation alerts

Aiping Yang, Jinbin Wang, Bingwang Yang, Yuqing He. Joint Deep Denoising Prior for Image Blind Deblurring[J]. Acta Optica Sinica, 2018, 38(10): 1010003 Copy Citation Text

show less

Fig. 1. Whole framework of algorithm

Download full size

Fig. 2. Pixel gray histograms of (a) clear image and (b) blurred image

Download full size

Fig. 3. Structure of denoising deep convolution neural network

Download full size

Fig. 4. Texture layer and structure layer of image. (a) Original image; (b) texture layer; (c) structure layer

Download full size

Fig. 5. Deblur results of different algorithms. (a) Blur images; (b) method in Ref. [7]; (c) method in Ref. [33]; (d) method in Ref. [16]; (e) proposed method

Download full size

Fig. 6. Clear images. (a) Boys; (b) bridge; (c) paint; (d) face

Download full size

Fig. 7. Blur kernel estimation of different algorithms. (a) True blur kernel; (b) method in Ref. [16]; (c) method in Ref. [7]; (d) method in Ref. [32]; (e) method in Ref. [34]; (f) proposed method

Download full size

Dataset	PSNR /dB		SSIM		t /s
Dataset	Proposedmethod	Method inRef. [19]	Proposedmethod	Method inRef. [19]	Proposedmethod	Method inRef. [19]
BSD68	28.74	29.22	0.8035	0.8278	32	3606
Classic5	29.62	30.38	0.8042	0.8323	2	262
Set12	29.79	30.42	0.8381	0.8617	6	580
Set14	29.34	30.01	0.8099	0.8352	3	719

Table 1. Comparison of performance and time complexity of denoising convolution neural networks

input: blur image B and blur kernel k
I ←B,β ←2λσrepeat: solve for u using Eq. (21)μ ←2λ repeat: solve for g using Eq. (22) solve for z using Eq. (23) solve for x using Eq. (17)μ ←2μ until μ>μ_maxβ ←2βuntil β>β_maxoutput:internediate latent image x→I

Table 2. Iterative algorithm for image deblurring

input: blur image B
initialize I and k with the results from the coarser level;for j=1→5 dosolve for I_S using Eq. (11)solve for k using Eq. (23)solve x→I using iterative algorithm in table 1λ←max $\begin{matrix} \{λ / 1.1,1 \times e^{- 4}\} \end{matrix}$ end foroutput:blur kernel k and internediate latent image x

Table 3. Iterative algorithm for blur kernel estimation

Blur	Image	Method inRef. [16]	Method inRef. [7]	Method Ref. [32]	Method inRef. [34]	Proposedmethod
1	Boys	25.71	24.74	27.11	25.30	30.00
	Bridge	27.74	24.72	24.49	18.97	27.72
	Paint	26.46	23.40	25.10	23.00	29.99
	Face	24.14	25.90	26.36	26.38	29.09
2	Boys	31.21	25.42	28.67	22.79	33.23
	Bridge	28.31	26.41	28.79	20.16	29.40
	Paint	30.99	25.62	28.57	24.79	31.99
	Face	27.80	26.05	28.82	23.64	30.31
3	Boys	27.03	24.22	25.87	18.90	30.84
	Bridge	26.24	22.49	23.48	21.77	28.11
	Paint	22.38	23.76	25.86	18.52	28.46
	Face	24.11	26.60	25.49	25.95	25.83
4	Boys	30.75	26.24	29.03	23.08	32.38
	Bridge	23.39	25.18	26.22	25.64	28.33
	Paint	26.89	25.52	25.98	33.05	30.46
	Face	25.64	23.71	24.79	22.58	29.52

Table 4. PSNR results of different algorithmsdB

Image	Image size /(pixel×pixel)	Blur kernelsize /(pixel×pixel)	Method inRef. [16]	Method inRef. [33]	Method inRef. [34]	Proposedmethod
ECCV	123×126	27×27	101	199	168	30
Roma	593×417	35×35	819	2938	792	138
Cartoon	612×442	19×19	726	2880	249	168
Flower	900×896	35×35	2909	8643	1921	358

Table 5. Time complexity of different algorithmss

Aiping Yang, Jinbin Wang, Bingwang Yang, Yuqing He. Joint Deep Denoising Prior for Image Blind Deblurring[J]. Acta Optica Sinica, 2018, 38(10): 1010003

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information