Transfer Learning for Constellation Recognition: A Comparative Study of Lightweight CNNs on Small-Scale Astronomical Image Data
Abstract
Constellation classification under severe data constraints remains an underexplored small-data computer vision problem, particularly in relation to lightweight architectures and augmentation strategy. This paper presents an exploratory benchmark comparing four approaches, a baseline CNN, MobileNetV2, EfficientNetB0, and HOG+SVM, on a manually curated set of 210 synthetic constellation images spanning seven classes. Using 5-fold cross-validation with confidence interval and statistical significance analysis, together with a held-out test split, the study evaluates performance with and without standard geometric and photometric augmentation. The results show that pretrained transfer learning models outperform the baseline CNN and HOG+SVM on this dataset, while augmentation has architecture-dependent effects: EfficientNetB0 achieved the strongest mean cross-validation performance without augmentation, whereas MobileNetV2 demonstrated comparatively stronger held-out generalization. Grad-CAM analysis further suggests that pretrained models may rely on both constellation geometry and background or rendering-style cues, raising important considerations regarding robustness and generalizability. Because the dataset was assembled from heterogeneous web sources without systematic provenance records, the conclusions should be interpreted as dataset-specific evidence rather than a general statement about constellation classification.
Authors: Kelzang Yuden, Ziaullah Momand, Pornchai Mongkolnam, Debajyoti Pal