Wavelet Sparse Coding-Based Lightweight Networks for Image Superresolution

Table 3

Comparisons of WLSR with public benchmark scores of CNNs. We use float32 WLSR and Luminance () channel while computing PSNR (The bold parts indicate the best performance).

Models	Parameters	Scale	Set5	Set14	BSD100	Urban100
Bicubic	—	×3	30.41	27.55	27.22	24.47
FSRCNN	25K	×3	33.16	29.43	28.60	26.48
VDSR	668K	×3	33.66	29.77	28.82	27.14
ESPCN	31K	×3	33.13	29.42	28.50	26.41
XLSR	22K	×3	33.42	29.73	28.55	26.71
WLSR	22K	×3	33.67	29.97	28.84	26.94
Bicubic	—	×4	28.42	26.00	25.26	23.06
FSRCNN	25K	×4	30.71	27.70	26.97	24.61
VDSR	668K	×4	31.35	28.02	27.29	25.18
ESPCN	31K	×4	30.90	27.73	27.06	25.07
XLSR	22K	×4	31.19	28.04	27.11	25.37
WLSR	22K	×4	31.46	28.16	27.30	25.41