Publication

Abstract : Aerial images captured by drones usually have the characteristics of low resolution, small object size, and high background clutter, so it is difficult to detect objects accurately. In this paper, an end-to-end framework based on super resolution image enhancement and the latest object detection is presented as a solution to alleviate the problems. VisDrone dataset images are firstly enhanced by the Enhanced Super-Resolution Generative Adversarial Network (ESRGAN) to increase spatial quality, restore intricate details, and augment textural information. Subsequently, the super-resolved images are passed to a YOLOv8 object detector, which is an anchor-free detection head with CSP backbone and PAN-FPN neck that performs multi-scale feature extraction and precise localization. Experimental results show that the proposed ESRGAN + YOLOv8 pipeline greatly enhances the detection results, especially for small or far away objects such as pedestrians, bicycles, and motorcycles. The results outperform those of the baseline YOLOv8 models, both rote trained on the original images, showing explicit improvements in mean Average Precision (mAP) and recall, which demonstrate the benefit of the joint application of image enhancement and object detection after the image enhancement step. The proposed architecture is a feasible solution for real time UAV surveillance, traffic monitoring, and crowd analysis, where achieving a high detection accuracy is essential even if the associated image resolution is limited. This work also paves the way for future investigations on the integration of lightweight super-resolution models and high-speed detection networks in aerial scenarios.