2025MNRAS.540..128C

Nearby extremely metal-poor galaxies (XMPs) allow us to study primitive galaxy formation and evolution in greater detail than is possible at high redshift. This work promotes the use of convolutional neural networks (CNNs) to efficiently search for XMPs in multiband imaging data based on their predicted N2 index (N2 $\equiv \log$$[\mathrm{N}\, {{\small II}}]\, \lambda 6585$/H $\alpha$). We developed a sequential characterization pipeline, composed of three CNN procedures: (i) a classifier for metal-poor galaxies, (ii) a classifier for XMPs, and (iii) an N2 predictor. The pipeline is applied to over 7.7 million Sloan Digital Sky Survey (SDSS) DR17 imaging data without SDSS spectroscopy. The predicted N2 values are used to select promising candidates for observations. This approach was validated by new observations of 45 candidates with redshifts less than 0.065 using the 2.54 m Isaac Newton Telescope and the 4.1 m Southern Astrophysical Research Telescope between 2023 and 2024. All 45 candidates are confirmed to be metal poor, including 28 new discoveries. There are 18/45 galaxies lacking detectable $[\mathrm{N}\, {{\small II}}]\, \lambda 6585$ lines ($\mathrm{ S/N}< 2$); for these, we report $2\sigma$ upper limits on their oxygen abundance. Our XMPs have estimated oxygen abundances of $7.1\le$$12+\log \left(\mathrm{O/H} \right)$$\le 8.7$ ($2\sigma$ upper limit), based on the N2 index, and 21 of them with estimated metallicity $< 0.1\,\mathrm{ Z}_{\odot }$. Additionally, we identified 4 potential candidates of low-metallicity AGNs at ${\lesssim} 0.1\,\mathrm{ Z}_{\odot }$. Finally, we found that our observed samples are mostly brighter in the g band compared to other filters, similar to blueberry galaxies, resembling green pea galaxies and high-redshift Ly $\alpha$ emitters.