[Retracted] Deep Image Watermarking to JPEG Compression Based on Mixed-Frequency Channel Attention

Table 3

Results of transparency against other distortions. For revealing the advantage of our proposed method, we make a comparison of PSNR and SSIM with [26, 28] by adjusting strength factor to keep the BER at 0%. Because the experiment of [27] is not done, we do not list it.

Metric	Model	Crop	Crop out	Drop out	GF	Identity
PSNR	HiDDeN [26]	35.20	47.24	42.52	40.55	44.63
	MBRS [28]	32.15	46.77	48.50	40.74	42.81
	Ours	33.01	47.26	49.43	42.30	45.71
SSIM	HiDDeN [26]	—	—	—	—	—
	MBRS [28]	0.7872	0.9910	0.9936	0.9670	0.9740
	Ours	0.8225	0.9924	0.9945	0.9760	0.9867
BER	HiDDeN [26]	0%	3%	0%	0%	0%
	MBRS [28]	0.72%	0%	0%	0%	0%
	Ours	0.29%	0%	0%	0%	0%