AsyScale-Net: Fine-Grained Tiny Object Recognition Across Heterogeneous UAV–Satellite Imagery

Abstract

Fine-grained classification of tiny objects in remote sensing (RS) is severely hindered by extreme ground sampling distance (GSD) variance and strict edge-computing constraints. To tackle these challenges, we propose AsyScale-Net. First, by leveraging comprehensive multifidelity imagery from our custom SkyView dataset, the framework employs scale-harmonized resample (SHR) to effectively mitigate cross-scale aliasing artifacts. Second, to optimize feature extraction while minimizing computational overhead, we introduce dilated feature fusion (DFF) and attentive semantic refinement (ASR) to preserve intricate spatial details and actively suppress padding-induced semantic dilution. Crucially, to facilitate practical deployment on edge devices, AsyScale-Net features an asymmetric architecture: it utilizes a dual-branch cross-view scheme during training to encode robust geometric priors, while strictly executing a streamlined single-branch pipeline during inference. Evaluated on the xView and SkyView datasets, the model achieves an accuracy of 86.3% and a Cohen’s kappa of 0.82. Requiring only 22.99 GFLOPs during inference, our framework demonstrates a superior balance between classification precision and computational efficiency for resource-constrained platforms.