Lightweight Vision Models for Galaxy Morphology Classification: A Comparative Study of ResNet, MobileNet, EfficientNet, and MobileViT on Galaxy10 DECaLS

Abstract

Automated galaxy morphology classification now rests on parameter-heavy convolutional backbones, whose cost becomes a design constraint as survey volumes grow. This paper compares three lightweight architectures (MobileNetV2, EfficientNetB0, and MobileViT-Extra Small (XS)) against a ResNet50 baseline on the Galaxy10 Dark Energy Camera Legacy Survey (DECaLS) benchmark of 17,736 images in ten classes, holding data splits, augmentation, optimizer, loss, and stopping rule fixed so that the backbone is the only variable. The lightweight models give up between 2.07 and 2.55 percentage points of support-weighted accuracy while removing 82.9% to 91.8% of the parameters and 42% to 59% of the per-image inference latency, which makes them 5.7 to 11.8 times more efficient per parameter. Two results run against expectations. The hardest class is not the rarest one. In a redshift-stratified comparison, the only attention-based backbone falls furthest behind the baseline on exactly the smallest and faintest galaxies, so the advantage reported for Vision Transformers does not carry over to a lightweight hybrid. For survey-scale processing, the trade is worth making; for a catalogue of a few thousand galaxies, it is not.