Mitigating hallucination with non-adversarial strategies for image-to-image translation in solar physics

Open Access

Table 1

Performance comparison with respect to generic image reconstruction metrics.

Task	Model	RMSE ↓	PS/N [dB] ↑	SSIM ↑	MS-SSIM ↑
WL-to-CaIIK	Pix2PixCC	16.77 ± 8.01	24.51 ± 3.83	0.72 ± 0.09	0.75 ± 0.12
	Pix2PixCC_w/o-Adv	15.07 ± 7.74	25.56 ± 4.10	0.75 ± 0.07	0.80 ± 0.10
	I2IwFiLM (Ours)	14.44 ± 7.07	25.90 ± 4.12	0.80 ± 0.09	0.80 ± 0.11
AIA304 Å-to-HMI	Pix2PixCC	15.61 × 10⁻³ ± 5.56 × 10⁻³	36.62 ± 2.88	0.96 ± 0.03	0.96 ± 0.03
	Pix2PixCC_w/o-Adv	14.61 × 10⁻³ ± 5.00 × 10⁻³	37.16 ± 2.76	0.94 ± 0.02	0.96 ± 0.02
	I2IwFiLM (Ours)	13.23 × 10⁻³ ± 5.06 × 10⁻³	38.10 ± 2.98	0.96 ± 0.03	0.96 ± 0.03

Notes. Comparison is performed for both whitelight-to-calcium and EUV-to-magnetogram translation tasks. GAN-based model Pix2PixCC is compared with non-adversarial models: Pix2PixCC_w/o-Adv and our proposed I2IwFiLM including input-dependent conditioning.

Current usage metrics show cumulative count of Article Views (full-text article views including HTML views, PDF and ePub downloads, according to the available data) and Abstracts Views on Vision4Press platform.

Data correspond to usage on the plateform after 2015. The current usage metrics is available 48-96 hours after online publication and is updated daily on week days.

Initial download of the metrics may take a while.