Table 4: A quantitative comparison of different methods. The best performance are in bold.

Dataset	DnCNN [9]	RL-Restore [11]	OWAN [12]	Ours
Mild (unseen)	27.51/0.7315	28.04/0.7313	28.33/0.7455	28.54/0.7493
Moderate	26.50/0.6650	26.45/0.6557	27.07/0.6787	27.12/0.6793
Severe (unseen)	25.26/0.5974	25.20/0.5915	25.88/0.6167	25.89/0.6170