Machine learning technique for morphological classification of galaxies from SDSS. IV. Visual inspection vs CNN for merging, irregular, edge-on, barred, ringed, and with dust lanes galaxies at 0.02<z<0.1
Context. Convolutional neural networks (CNNs) are widely used for automated galaxy morphological classification in large surveys. However, projection effects, image artefacts, and intrinsic degeneracies limit reliable identification of detailed features, requiring large-scale visual validation. Aims. To visually inspect SDSS galaxies at 0.02 < z < 0.1 classified by a CNN as merging, irregular, edge-on, barred, ringed, or dust-lane galaxies; assess CNN completeness and failure modes; construct visually verified morphological catalogues; and determine nuclear activity types via BPT diagrams. Methods. We visually inspected all galaxies assigned by the CNN to six morphological classes: merging (2,574), irregular (9,432), edge-on (17,000), barred (6,000), ringed (13,882), and dust-lane (588), regardless of CNN probability. Refined samples were cross-matched with Galaxy Zoo 2; remaining galaxies were classified here for the first time. Nuclear activity was determined from SDSS DR17 spectra using Hααα, H\b{eta}β\b{eta}, [O III]λ$5007, and [N II] λ$6583 line ratios. Results. We present catalogues of 612 merging, 9,372 irregular, 16,822 edge-on, 575 dust-lane, 811 barred, and 2,150 ringed galaxies. CNN misclassifications stem primarily from projection effects, foreground stars, faint tidal features, and irregular star-forming structures. We characterise nuclear activity types for edge-on, barred, ringed, and dust-lane galaxies, finding systematic differences in LINER-like and composite fractions across subsamples. Five strong polar ring galaxy candidates were identified. Conclusions. Visual validation remains essential for refining CNN-based classifications. The resulting datasets support morphological studies, investigations of galaxy structure and secular evolution, and provide robust training samples for future machine learning models.