Optimizing YOLO Algorithm for Efficient Object Detection in Resource-Constrained Environments

In the realm of airborne remote sensing, it could be challenging to distinguish small objects in aerial pictures. Their limited image quality, varied settings, and faint presence against huge backgrounds make identification more challenging. While many deep learning methods have significant processing resource needs that limit real-time applications, real-time applications such as traffic monitoring, urban planning, and military surveillance depend on their detection. To fulfil real-time demands and enhance the accuracy of small item detection, we introduce RC-YOLOv8, a lightweight and efficient model for aerial photo object detection. As a result, this paper proposes RC-YOLOv8, a lightweight object detection algorithm designed specifically for resource restrictions. The system provides predefined anchor boxes that are flexible and tailored for contexts with limited resources in order to enhance target detection performance. Concurrently, an L1 norm-based pruning technique is developed that greatly reduces the computation and parameter volume of the model without sacrificing accuracy. In terms of performance, RC-YOLOv8 lowers computational needs with GFLOPs falling from 15.9 to 11.2 and parameters by 30.5%, from 7.05 M to 4.9 M. On the DOTA dataset, it outperforms the original YOLOv5, with a 5.5% mAP0.5 improvement, 0.9% greater precision, and 1.4% better recall. When it comes to performance, RC-YOLOv8 is superior to other cutting-edge techniques.