An enhanced label-free pipeline for classifying galaxy morphology from single-band survey images. A convolutional autoencoder denoises each image and an adaptive polar-coordinate transform imposes rotational invariance; a pre-trained ConvNeXt network encodes the image into a 2048-d vector that PCA compresses to 1500-d; and a bagging-based multi-model voting scheme runs three clustering algorithms (BIRCH, k-means, hierarchical agglomerative), keeping only samples on which at least two agree. Using large-model encoding plus PCA, the pipeline cuts the required cluster groups from 100 to 20 on ~100k HST I-band COSMOS galaxies, classifies about half of them, and its five resulting morphological categories reproduce the expected trends in Sersic index, Gini, M20, concentration, and MID parameters.
This paper presents an enhanced unsupervised machine-learning (UML) pipeline for classifying galaxy morphology from single-band survey images without labels, upgrading the UML stage of the authors' U…