Panoramic imaging is increasingly critical in UAVs and high-altitude surveillance applications. In addressing the challenges of detecting small targets within wide-area, high-resolution panoramic images, particularly issues concerning accuracy and real-time performance, we have proposed an improved lightweight network model based on YOLOv8. This model maintains the original detection speed, while enhancing precision, and reducing the model size and parameter count by 10.6% and 11.69%, respectively. It achieves a 2.9% increase in the overall mAP@0.5 and a 20% improvement in small target detection accuracy. Furthermore, to address the scarcity of reflective panoramic image training samples, we have introduced a panorama copy-paste data augmentation technique, significantly boosting the detection of small targets, with a 0.6% increase in the overall mAP@0.5 and a 21.3% rise in small target detection accuracy. By implementing an unfolding, cutting, and stitching process for panoramic images, we further enhanced the detection accuracy, evidenced by a 4.2% increase in the mAP@0.5 and a 12.3% decrease in the box loss value, validating the efficacy of our approach for detecting small targets in complex panoramic scenarios.
Keywords: YOLOv8; panoramic detection; reflective panorama; small object detection.