In low-altitude remote sensing for real-time inspection, strong parallax, scale variations, and local distortions make the traditional “stitch-then-detect” serial paradigm prone to feature redundancy and error cascading, and make it difficult to unify global and local geometric transformations, which in turn limits speed and weakens robustness. To address this, we propose Fast Stitching and Detection Network (FSDNet), a fast dual-branch stitching and synchronous detection framework built on a shared backbone. The framework adopts a pretrained detection backbone as a unified encoder and embeds an attention-guided local context correlation module in the stitching branch to explicitly regress fine-grained geometric flow fields from shared features. In addition, two collaborative branches are designed for estimating the transformation fields: a global homography branch for coarse alignment and a local thin-plate spline branch for fine alignment. These are combined with transformation-guided detection box rectification and intensity-adaptive fusion to enhance geometric-semantic consistency. Experiments on UDIS-D and Warped AU-AIR demonstrate that, while maintaining high-quality stitching, the proposed method improves Frames per Second (FPS) by about 66% compared with typical serial baselines and achieves superior object detection performance on Warped AU-AIR, validating the efficiency and practicality of the approach.

Influenced by scattering characteristics and imaging geometry, single-polarization SAR remote sensing images suffer from severe speckle noise and significant degradation, leading to texture loss and structural distortion in SAR image super-resolution tasks. To address this issue, this paper proposes a diffusion model-based super-resolution method that integrates frequency-domain processing and structural perception. The method adopts a latent diffusion model as its backbone and introduces a wavelet-guided module and a direction-aware enhancement module to improve performance. The wavelet-guided module performs multi-scale modulation of high-frequency sub-bands through wavelet decomposition and spatially adaptive normalization, thereby dynamically enhancing texture representation and high-frequency reconstruction capability according to the diffusion timestep. The direction-aware enhancement module incorporates multiple adaptive convolutional kernels, embedded with channel attention into the deep residual structures of the encoder to increase the sensitivity of compressed feature maps to structural information. In experiments, a realistic degradation model combining speckle noise and blur kernels is established to closely approximate the actual imaging pipeline. Results demonstrate that the proposed method significantly outperforms existing frameworks across multiple datasets, achieving an average improvement of 8.12% over the best baseline, verifying its effectiveness and advancement in SAR image super-resolution tasks.