Algorithm evaluation demonstrates superior detail recovery and structural preservation in remote sensing imagery, highlighting the power of multi-level attention diffusion models.
Key Points
To develop a fast-sampling conditional diffusion framework that mitigates prior detail loss, global modeling deficiencies, and high-noise degradation in remote sensing image super-resolution.
Constructed a multi-dilated residual attention block to capture prior features from low-resolution inputs and counteract interpolation-induced detail loss.
Integrated a dual-attention mechanism into the denoising network to enhance long-range spatial context and adaptive channel feature weighting.
Applied a residual-domain v-prediction strategy in place of standard noise prediction to restore fine structural details under high-noise conditions.
DAV-Diff demonstrated superior structural and textural detail restoration over existing state-of-the-art super-resolution methods.
Ablation studies and computational sensitivity tests confirmed that the multi-level attention modules and v-prediction strategy substantially improved sampling efficiency and ground-object fidelity.