Large-scale phase retrieval

eLight

Table 2 Quantitative comparison under the CDP modality (5 modulations)

Algorithm	SNR = 10dB			SNR = 15dB			SNR = 20dB
Algorithm	PSNR	SSIM	TIME	PSNR	SSIM	TIME	PSNR	SSIM	TIME
AP	15.60	0.21	105.76	18.61	0.33	110.73	23.22	0.55	174.98
WF	✘-insufficient measurements			✘-insufficient measurements			✘-insufficient measurements
RWF	✘-insufficient measurements			✘-insufficient measurements			✘-insufficient measurements
AF	13.93	0.19	247.07	17.84	0.33	231.38	23.13	0.60	211.39
TAF	13.40	0.16	257.57	18.14	0.34	225.67	22.71	0.59	213.65
RAF	13.88	0.19	261.59	17.86	0.38	222.38	23.10	0.59	212.09
PLIFT	✘-memory limitation			✘-memory limitation			✘-memory limitation
PLAMP	✘-memory limitation			✘-memory limitation			✘-memory limitation
PMAX	11.08	0.13	295.84	11.36	0.14	300.21	11.66	0.15	296.28
CD	8.69	0.22	357.52	9.47	0.20	321.81	9.78	0.20	264.89
KAC	10.83	0.13	192.44	10.97	0.15	161.48	11.01	0.16	114.75
prDeep	22.67	0.61	301.41	24.42	0.72	282.14	26.85	0.76	380.60
LPR	22.73	0.82	124.80	26.92	0.88	137.33	31.89	0.94	228.42

The Wirtinger flow based (WF, RWF) techniques fail because of insufficient measurements. PLIFT and PLAMP are out of memory. The other methods produce little improvement or consume extremely long running time compared to AP. In comparison, LPR consumes the same level of running time as AP, and obtains the best performance with as much as 8.3dB on PSNR (SNR = 15) and 0.61 on SSIM (SNR = 10)