Multiscale Feedforward Structure-Based Single Image Rain Removal Algorithm

Zhicheng Jiang; Zhiwei Li; Chen Chen; Jinxiang Zhou; Wuneng Zhou

doi:10.3788/LOP202259.1215001

Journals >Laser & Optoelectronics Progress >Volume 59 >Issue 12 >Page 1215001 > Article

Laser & Optoelectronics Progress
Vol. 59, Issue 12, 1215001 (2022)

Multiscale Feedforward Structure-Based Single Image Rain Removal Algorithm

Zhicheng Jiang¹, Zhiwei Li^1、2、*, Chen Chen¹, Jinxiang Zhou¹, and Wuneng Zhou²

Author Affiliations

¹College of Electronic and Electrical Engineering, Shanghai University of Engineering Science, Shanghai 201620, China

²College of Information Science and Technology, Donghua University, Shanghai 200051, China

show less

DOI: 10.3788/LOP202259.1215001 Cite this Article Set citation alerts

Zhicheng Jiang, Zhiwei Li, Chen Chen, Jinxiang Zhou, Wuneng Zhou. Multiscale Feedforward Structure-Based Single Image Rain Removal Algorithm[J]. Laser & Optoelectronics Progress, 2022, 59(12): 1215001 Copy Citation Text

show less

Fig. 1. SE block structure diagram

Download full size

Fig. 2. MSRAS structure diagrams. (a) MSRAS1; (b) MSRAS2; (c) MSRAS3; (d) MSRAS4

Download full size

Fig. 3. Overall network structure diagram

Download full size

Fig. 4. Comparison of rain removal effects of different structures. (a) Real pictures; (b) rain pictures; (c) ResNet; (d) without MSRAS; (e) ours

Download full size

Fig. 5. Rain removal effect diagram of different algorithms under the composite dataset. (a) Real pictures; (b) rain pictures; (c) DID-MDN; (d) JORDER; (e) RESCAN; (f) ours

Download full size

Fig. 6. Rain removal renderings of different algorithms under real rain maps. (a) Real pictures; (b) DID-MDN; (c) JORDER; (d) RESCAN; (e) ours

Download full size

MSRAS	Filter size	Number of filters	Number of output channels
MSRAS1	$7 \times 7$	16	48
	$5 \times 5$	16
	$3 \times 3$	16
MSRAS2	$7 \times 7$	16	96
	$5 \times 5$	16
	$3 \times 3$	64
MSRAS3	$5 \times 5$	16	96
	$3 \times 3$	16
	$1 \times 1$	64
MSRAS4	$3 \times 3$	16	80
MSRAS4	$1 \times 1$	64	80

Table 1. Parameter setting table of MSRAS

Input size	Filter size	Number of filters	Activation function	Output	Output size
［128，64，64，3］	$3 \times 3$	64	ReLU	$F_{0}$	［128，64，64，64］
［128，64，64，64］	$3 \times 3$	64	ReLU	$F_{1}$	［128，64，64，64］
［128，64，64，64］	$3 \times 3$	64	ReLU	$F_{2}$	［128，64，64，64］
［128，64，64，64］	‒	‒	‒	$F_{3}$	［128，64，64，128］
［128，64，64，128］	$3 \times 3$	64	ReLU	$F_{4}$	［128，64，64，64］
［128，64，64，64］	$3 \times 3$	64	ReLU	$F_{5}$	［128，64，64，64］
［128，64，64，64］	‒	‒	‒	$F_{6}$	［128，64，64，128］
［128，64，64，128］	$3 \times 3$	64	ReLU	$F_{7}$	［128，64，64，64］
［128，64，64，64］	$3 \times 3$	64	ReLU	$F_{8}$	［128，64，64，64］
［128，64，64，64］	‒	‒	‒	$F_{9}$	［128，64，64，384］
［128，64，64，384］	$3 \times 3$	64	ReLU	$F_{10}$	［128，64，64，64］
［128，64，64，64］	$3 \times 3$	64	ReLU	$F_{11}$	［128，64，64，64］
［128，64，64，64］	‒	‒	‒	$F_{12}$	［128，64，64，384］
［128，64，64，384］	$3 \times 3$	16	ReLU	$F_{13}$	［128，64，64，16］
［128，64，64，16］	$3 \times 3$	3	ReLU	$F_{14}$	［128，64，64，3］

Table 2. Network parameter table

Picture	SSIM
Picture	Rain picture	ResNet	Without MSRAS	Ours
Photographer	0.7036	0.8915	0.8722	0.9443
Tiger	0.6968	0.8636	0.9057	0.9569
Cottage	0.7823	0.8932	0.9219	0.9788
Bridge	0.7733	0.8953	0.9032	0.9487
200 composite rain pictures	0.7135	0.8722	0.9063	0.9504
Picture	PSNR /dB
Picture	Rain picture	ResNet	Without MSRAS	Ours
Photographer	19.1911	27.9249	27.4652	33.6541
Tiger	20.0945	26.1495	28.5863	34.0457
Cottage	19.2886	26.3365	30.1487	34.2584
Bridge	18.6967	27.6713	28.9957	32.5586
200 composite rain pictures	18.3256	26.1841	28.9426	33.9953

Table 3. Quantitative values of rain removal effects of different structures

Picture	SSIM
Picture	DID-MDN	JORDER	RESCAN	Ours
Path	0.9069	0.9166	0.9389	0.9392
Model	0.8872	0.9233	0.9258	0.9473
Man	0.9022	0.9341	0.9389	0.9421
Arab	0.8575	0.9178	0.9498	0.9587
Rain100L	0.8821	0.9702	0.9752	0.9821
Rain100H	0.7285	0.7633	0.8726	0.8832
Rain800	0.8858	0.8835	0.8942	0.9289
Rain12000	0.8125	0.8325	0.8432	0.9072
Picture	PSNR/dB
Picture	DID-MDN	JORDER	RESCAN	Ours
Path	29.8813	30.0288	31.8564	32.0873
Model	28.1575	32.573	33.9847	34.2635
Man	31.2462	34.7834	35.1448	33.2516
Arab	27.1485	33.2175	35.5248	35.6874
Rain100L	25.6564	36.1143	36.2145	36.9326
Rain100H	17.8715	25.7412	26.4121	28.9627
Rain800	26.5586	25.9984	30.2486	31.7585
Rain12000	24.2263	25.8896	26.7423	28.9245