Multiscale Nonlocal Neural Network for Single Image Deraining

Fig. 6. Derained results on Test1200 and Test1000 synthetic dataset. (a) Rain image; (b) DerainNet; (c) DID-MDN; (d) UMRL; (e) proposed method; (f) ground truth

Download full size

Fig. 7. De-rained results on real-world scenarios. (a) Rain image; (b) DerainNet; (c) DID-MDN; (d) UMRL; (e) proposed method

Download full size

Table 1. Parameter settings of Conv-LSTM module
View table
Table 1. Parameter settings of Conv-LSTM module
Layer Parameter
Conv-LSTM Input gate Conv (3×3, 32) Stride is 1 Sigmoid
Forget gate Conv (3×3, 32) Stride is 1 Sigmoid
Cell state Conv (3×3, 32) Stride is 1 Sigmoid
Output gate Conv (3×3, 32) Stride is 1 Sigmoid

Table 2. Quantitative comparation results of five models on Test1200
View table
Table 2. Quantitative comparation results of five models on Test1200
Parameter Model 1 Model 2 Model 3 Model 4 Proposed method
PSNR /dB 28.76 29.28 26.14 28.95 30.16
SSIM 0.87 0.90 0.84 0.88 0.92

Table 3. Quantitative evaluation results in terms of PSNR and SSIM

View table

Table 3. Quantitative evaluation results in terms of PSNR and SSIM

Dataset	Rain image	DerainNet	DDN	DID-MDN	UMRL	Proposed method
Dataset	PSNR /dB SSIM	PSNR /dB SSIM	PSNR /dB SSIM	PSNR /dB SSIM	PSNR /dB SSIM	PSNR /dB SSIM
Test1200	21.15 0.77	22.07 0.84	27.33 0.90	27.95 0.91	29.42 0.91	30.16 0.92
Test1000	19.31 0.77	19.73 0.83	25.63 0.88	26.08 0.90	26.47 0.91	26.88 0.92