Hierarchical Progressive Distillation for Image Super-Resolution of Remote Sensing

Abstract

Remote sensing image super-resolution (SR) fundamentally relies on effective feature representation learning. However, the existing methods that adopt mainstream architecture designs, such as transformer and state space model (SSM), are often restricted by two factors when dealing with the SR of remote sensing images. Initially, to overcome the limitations of shallow networks and achieve sufficient depth in feature extraction, constructing deeper network architectures is inevitable, which increases the model’s parametric complexity and computational cost. On the other hand, as network depth increases continuously, the transmission efficiency of information flow across different hierarchical levels declines markedly. Key features undergo gradual attenuation during successive transformation processes, giving rise to the loss of critical information. We propose a novel hierarchical progressive distillation (HPD) framework to alleviate this predicament, which establishes a systematic hierarchical path for feature refinement. While eliminating redundant computations, it can better preserve semantic information. Subsequently, given the potential for global information loss during distillation, we propose spatial-frequency hybrid attention (SHA) to enhance feature representation. In particular, SHA adaptively captures spatial dependencies of different scales by integrating dynamic window attention (DWA) and multiscale frequency attention (MFA), and optimizes frequency processing to provide global context modeling. A large number of experiments we conducted across multiple public remote sensing datasets show that our method achieves a trade-off among model parameters, computational overhead, and performance, and outperforms existing state-of-the-art methods.