Detection-oriented diffusion super-resolution for autonomous remote sensing perception

Abstract

Low-resolution remote sensing imagery often degrades downstream object detection by blurring object boundaries and suppressing class-discriminative textures, particularly for small and densely distributed targets under sensing and transmission constraints. Although Remote Sensing Image Super-Resolution (RSISR) has achieved substantial progress, most existing methods are optimized primarily for visual reconstruction fidelity. Consequently, higher PSNR or SSIM does not necessarily translate into better downstream detection performance. To address this mismatch, we develop a detection-oriented super-resolution framework built upon a residual diffusion-based super-resolution branch and guided by downstream detection supervision. A frozen YOLOv11 detector is introduced during training to provide task-level guidance, encouraging the super-resolution model to recover structures relevant to object localization and classification. In addition, a Learnable Canny Edge Module (LCEM) is incorporated to enhance contour-sensitive structures that are critical for aerial object detection. Experiments on the DOTA benchmark show that the proposed framework improves mAP50 from 0.5655 for low resolution inputs to 0.5899, while maintaining competitive reconstruction fidelity. Ablation studies further indicate that detector-guided supervision is the main source of the mAP50 improvement, while LCEM provides complementary structural refinement. These results suggest that, for autonomous remote sensing perception, image enhancement should be optimized and evaluated as part of the downstream sensing pipeline rather than as an isolated visual restoration task.