Enhancing Object Detection in Aerial Images Using Transformer-Based Super-Resolution

Enhancing the resolution of aerial images is an important task in military and scientific applications. Sufficient image quality is essential for robust object detection. In this paper, we propose using transformer-based super-resolution techniques to increase object detection accuracy in aerial im-agery, specifically focusing on the Dataset for Object Detection in Aerial Images (DOTA) dataset. We utilize the Hybrid Attention Transformer for Image Restoration (HAT-L) architecture as a transformer model for super-resolution and analyze its influence on object detection performance, especially using the HAT- L model pre-trained on the ImageNet dataset. We integrate the YOLOv8 (You Only Look Once) OBB model, pre-trained on the DOTA dataset, to assess the effectiveness of our approach in enhancing object detection capabilities. Our results highlight the benefits of combining the HAT-L archi-tecture for super-resolution with the YOLOv8 OBB model for object detection tasks. We achieve a Peak Signal-to-Noise Ratio (PSNR) of 37.847 and a Structural Similarity Index (SSIM) value of 0.903 for super-resolved images on the DOTA validation set using the HAT- L architecture. Additionally, our integrated approach yields a mean Average Precision (mAP) of 0.809 at IOU 0.5 and 0.656 at IOU 0.5-0.95 using the YOLOv8x OBB model on the super-resolved images. Our findings contribute valuable insights into the effectiveness of transformer models for image enhancement and object detection tasks in remote sensing applications.