Fine-Grained Object Detection in Remote Sensing Images Based on Improved Yolov8

Remote sensing targets are of various sizes with complex backgrounds. The detection accuracy and recognition ability of small targets are limited, which easily leads to false detection and missed detection, and the multi-category detection performance is unstable, which brings challenges to refined target detection. In order to improve the multi-target detection accuracy of remote sensing images and reduce detection errors, this paper proposes a remote sensing fine-grained target detection model based on improved YOLOv8. Based on YOLOv8, this model uses full-dimensional dynamic convolution ODConv to build a new C2f module, which improves the utilization of multi-dimensional information in the kernel space and reduces the amount of calculation. At the same time, the multi-head self-attention mechanism MHSA is introduced into the backbone network to focus on target features effectively, improve the performance of more fine-grained small target detection, and suppress background interference. Experimental results based on the FAIR1M million-level remote sensing image fine-grained target detection data set show that compared with models such as R3Det, S2ANet, F-RCNN, ReDet, ROI Transformer, and O-RCNN, the improved YOLOv8 has better detection accuracy and model efficiency. The performance has been significantly improved, and mAP has increased by 8.25%, 7.93%, 5.28%, 3.94%, 2.83% and 2.66% respectively. Compared with the original YOLOv8 algorithm, mAP has increased by 1.94%, and the number of parameters has been reduced by 7.4%, which better balances detection accuracy and computational complexity and meets the needs of real-time detection.