Dual Residual Denoising Network Based on Hybrid Attention

Haitao Yin; Hao Deng

doi:10.3788/LOP202158.1410002

Journals >Laser & Optoelectronics Progress >Volume 58 >Issue 14 >Page 1410002 > Article

Laser & Optoelectronics Progress
Vol. 58, Issue 14, 1410002 (2021)

Dual Residual Denoising Network Based on Hybrid Attention

Haitao Yin^* and Hao Deng

Author Affiliations

College of Automation and College of Artificial Intelligence, Nanjing University of Posts and Telecommunications, Nanjing, Jiangsu 210023, China

show less

DOI: 10.3788/LOP202158.1410002 Cite this Article Set citation alerts

Haitao Yin, Hao Deng. Dual Residual Denoising Network Based on Hybrid Attention[J]. Laser & Optoelectronics Progress, 2021, 58(14): 1410002 Copy Citation Text

show less

Fig. 1. Network structure of DnCNN

Download full size

Fig. 2. Structure of ResNet block and SE-ResNet block. (a) Traditional ResNet block; (b) SE-ResNet block

Download full size

Fig. 3. Structure of non-local block

Download full size

Fig. 4. Structure of traditional residual network

Download full size

Fig. 5. Structure of dual residual network

Download full size

Fig. 6. Structure of HDDNet network model

Download full size

Fig. 7. Structure of D-block module

Download full size

Fig. 8. Structure of ResNet block. (a) Traditional ResNet block; (b) simplified ResNet block

Download full size

Fig. 9. Structure of DNL-block module

Download full size

Fig. 10. Test images

Download full size

Fig. 11. Denoised images of different algorithms on Starfish (σ=30). (a) Original image; (b) noisy image; (c) DnCNN; (d) FDnCNN; (e) FFDNet; (f) IRCNN; (g) DuRN; (h) HDDNet

Download full size

Fig. 12. Denoised images of different algorithms on Butterfly (σ=50). (a) Original image; (b) noisy image; (c) DnCNN; (d) FDnCNN; (e) FFDNet; (f) IRCNN; (g) DuRN; (h) HDDNet

Download full size

Fig. 13. Denoised images for pepper & salt noise (noise density is 10%). (a)--(g) Pepper & salt noise images; (h)--(n) denoised images

Download full size

Module	Conv-A kernel size	Conv-B kernel size	Dilation
DNL-block1	5	3	1
D-block1	7	5	1
D-blcok2	7	5	2
D-block3	11	7	2
D-block4	11	5	1
DNL-block2	11	7	3

Table 1. Parameters of DNL-block and D-block modules

Noise level	Method	Cameraman	House	Peppers	Starfish	Butterfly	Airplane	Parrot	Average
σ=15	DnCNN	32.59	34.99	33.24	32.13	33.25	31.67	31.88	32.82
	FDnCNN	32.59	35.13	33.17	32.07	33.14	31.67	31.85	32.80
	FFDNet	32.37	35.05	33.02	31.95	32.92	31.55	31.79	32.66
	IRCNN	32.53	34.88	33.21	31.96	32.98	31.66	31.88	32.73
	DuRN	32.29	34.73	33.02	32.01	32.95	31.62	31.80	32.63
	HDDNet	32.59	34.96	33.21	32.15	33.22	31.69	31.86	32.81
σ=30	DnCNN	29.28	32.29	29.85	28.28	29.42	28.19	28.59	29.42
	FDnCNN	29.43	32.59	29.92	28.37	29.43	28.21	28.70	29.52
	FFDNet	29.28	32.57	29.87	28.34	29.39	28.13	28.65	29.46
	IRCNN	29.28	32.19	29.79	28.14	29.22	28.14	28.62	29.34
	DuRN	28.99	32.26	29.53	28.35	29.24	28.08	28.52	29.28
	HDDNet	29.43	32.43	29.98	28.56	29.52	28.23	28.65	29.53
σ=50	DnCNN	27.26	29.96	27.35	25.6	26.83	25.83	26.42	27.04
	FDnCNN	27.35	30.25	27.41	25.66	26.84	25.81	26.59	27.13
	FFDNet	27.34	30.36	27.41	25.68	26.92	25.79	26.57	27.14
	IRCNN	27.16	29.90	27.33	25.48	26.66	25.78	26.48	26.97
	DuRN	26.91	30.14	27.13	25.71	26.73	25.87	26.35	26.98
	HDDNet	27.33	30.35	27.37	25.89	26.97	25.90	26.47	27.18

Table 2. PSNR values of different methods

Noise level	Method	Cameraman	House	Peppers	Starfish	Butterfly	Airplane	Parrot	Average
σ=15	DnCNN	0.9131	0.8855	0.9121	0.9146	0.9501	0.9077	0.9049	0.9126
	FDnCNN	0.9132	0.8870	0.9119	0.9135	0.9503	0.9080	0.9047	0.9127
	FFDNet	0.9118	0.8877	0.9112	0.9126	0.9491	0.9074	0.9045	0.9120
	IRCNN	0.9113	0.8831	0.9107	0.9123	0.9477	0.9064	0.9039	0.9108
	DuRN	0.9089	0.8834	0.9153	0.9205	0.9487	0.9071	0.9078	0.9131
	HDDNet	0.9137	0.8858	0.9117	0.9147	0.9505	0.9081	0.9050	0.9128
Noise level	Method	Cameraman	House	Peppers	Starfish	Butterfly	Airplane	Parrot	Average
σ=30	DnCNN	0.8500	0.8518	0.8609	0.8453	0.9025	0.8511	0.8425	0.8577
	FDnCNN	0.8593	0.8542	0.8648	0.8463	0.9071	0.8537	0.8463	0.8617
	FFDNet	0.8599	0.8542	0.8652	0.8457	0.9073	0.8537	0.8467	0.8618
	IRCNN	0.8530	0.8474	0.8561	0.8412	0.8990	0.8475	0.8427	0.8553
	DuRN	0.8482	0.8507	0.8634	0.8541	0.9020	0.8494	0.8474	0.8593
	HDDNet	0.8611	0.8538	0.8661	0.8502	0.9082	0.8545	0.8455	0.8628
σ=50	DnCNN	0.8077	0.8185	0.8090	0.7722	0.8513	0.7978	0.7952	0.8074
	FDnCNN	0.8107	0.8245	0.8137	0.7747	0.8553	0.7986	0.7995	0.8110
	FFDNet	0.8138	0.8273	0.8164	0.7750	0.8585	0.7997	0.8004	0.8130
	IRCNN	0.8028	0.8159	0.8044	0.7675	0.8454	0.7953	0.7953	0.8038
	DuRN	0.7954	0.8211	0.8082	0.7800	0.8451	0.7954	0.7921	0.8053
	HDDNet	0.8142	0.8293	0.8144	0.7812	0.8575	0.8012	0.7971	0.8136

Table 3. SSIM values of different methods

Noise level	Method	Cameraman	House	Peppers	Starfish	Butterfly	Airplane	Parrot	Average
σ=15	DnCNN	5.99	4.54	5.55	6.31	5.55	6.66	6.49	5.87
	FDnCNN	5.99	4.47	5.60	6.36	5.62	6.65	6.52	5.89
	FFDNet	6.14	4.51	5.70	6.44	5.76	6.74	6.56	5.98
	IRCNN	6.03	4.60	5.57	6.44	5.72	6.66	6.49	5.93
	DuRN	6.20	4.68	5.70	6.40	5.74	6.69	6.55	5.99
	HDDNet	5.98	4.56	5.57	6.30	5.57	6.64	6.51	5.87
σ=30	DnCNN	8.76	6.19	8.20	9.82	8.62	9.93	9.48	8.72
	FDnCNN	8.61	5.98	8.14	9.73	8.61	9.91	9.36	8.62
	FFDNet	8.76	6.00	8.19	9.76	8.65	10.00	9.42	8.68
	IRCNN	8.76	6.27	8.26	9.99	8.82	9.99	9.46	8.74
	DuRN	9.06	6.22	8.51	9.75	8.80	10.05	9.57	8.85
	HDDNet	8.61	6.09	8.14	9.52	8.52	9.89	9.41	8.60
σ=50	DnCNN	11.05	8.10	10.94	13.32	11.62	13.04	12.18	11.46
	FDnCNN	10.94	7.83	10.87	13.29	11.60	13.07	11.95	11.36
	FFDNet	11.08	7.74	10.87	13.26	11.50	13.09	11.97	11.36
	IRCNN	11.18	8.15	10.96	13.57	11.85	13.11	12.10	11.56
	DuRN	11.51	7.94	11.22	13.21	11.75	12.98	12.28	11.55
	HDDNet	10.97	7.74	10.91	12.95	11.44	12.93	12.11	11.29

Table 4. RMSE values of different methods

Haitao Yin, Hao Deng. Dual Residual Denoising Network Based on Hybrid Attention[J]. Laser & Optoelectronics Progress, 2021, 58(14): 1410002

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information