Spatial and Wavelet Attention-Enhanced Super-Resolution for Small Object Detection in Satellite Imagery

Enhancing the detection of small objects in satellite imagery is of paramount importance for applications such as military surveillance and security monitoring. The challenge lies in addressing issues such as low resolution and image noise, which often lead to edge blurring and complicate object detection. This paper investigates super-resolutionenhanced small object detection, particularly for ships in satellite images, employing a transformer-based architecture designed to emphasize and improve edge sharpness. The proposed model eliminates the window attention mechanism, substituting it with spatial and frequency self-attention to reinforce superior super-resolution detail learning, thereby enhancing the model's ability to capture finer details through spatial and frequency enhancements. Furthermore, the model optimizes performance by replacing depthwise-separable convolution, reducing computational complexity without compromising efficiency. Evaluated on the FGSCR dataset, with a specific focus on ship images, the proposed model achieves a notable 0.51 PSNR improvement and a 7% reduction in GFLOPs compared to the baseline SwinIR model. Finally, the proposed model was evaluated on YOLO object detection for practical application.