Improved YOLOv8n Models for Object Detection in Remote Sensing Images

In recent years, numerous efficient object detectors have emerged in computer vision. However, applying these models to remote sensing images remains challenging due to complex backgrounds, high object scale variation, and the difficulty of detecting small objects. Many existing approaches prioritize accuracy but overlook the trade-off between precision and efficiency, limiting their usability in resource-constrained environments. To address these challenges, we propose a lightweight model based on the original YOLOv8n, culminating in the Transformer with Bi-directional and Coordinate Attention-YOLOv8n (TBC-YOLO8n) model. Our approach introduces a weighted bidirectional feature pyramid network that adaptively assigns importance to different feature maps, enhancing multi-scale object detection. Additionally, we integrate transformer blocks to improve global feature representation and capture long-range dependencies, alongside a coordinate attention block before each detection head to enhance localization by incorporating spatial information. Unlike previous YOLO variants, our method optimizes feature fusion strategies and integrates advanced attention mechanisms specifically tailored for remote sensing images, effectively addressing small-object detection, large-scale variations, and complex backgrounds. Compared to the original YOLOv8n, our proposed TBC-YOLO8n model improves mAP@50 from 90.2% to 94.5%, with a slight increase in inference time from 3.0 ms to 4.0 ms. Experimental results confirm the robustness and effectiveness of our model, making it a practical choice for real-time remote sensing applications with limited computational resources.