An improved model based on YOLOv8 for tiny object detection in the maritime environment

To address the issue of low accuracy in maritime tiny target detection, this paper presents an optimized maritime tiny target detection model based on YOLOv8. Firstly, the Spatial tiny target detection model based on YOLOv8. Firstly, the Spatial and Channel Reconstruction Convolution (SCConv) and the Cascaded Group Attention Mechanism (CGA) are introduced to optimize the original C2f module in the backbone of YOLOv8. SCConv helps reduce spatial and channel redundancy, thereby enhancing the model’s efficiency without incurring additional computational costs. Meanwhile, CGA is conducive to enhancing the focus on significant targets, resulting in an efficient improvement of the model’s detection performance. Secondly, the the focus on significant targets, resulting in an efficient improvement of the model’s detection performance. Secondly, the P2-tiny target detection layer and the GhostConv module are adopted to enhance the model’s ability to detect tiny targets and reduce the number of parameters in the model. Finally, G-ELAN reduce the number of parameters in the model. Finaliy, G-ELAN is introduced to replace the C2f module in the neck to deepen the network depth and enhance the model’s feature extraction capability. The improved model is trained and validated on the SeaDronesSee Object Detection V2 dataset. The results indicate SeaDronesSee Object Detection $V 2$ dataset. The resuits indicate that our model improves the precision and mAP mA<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">50</inf> by 3.9% and $\mathbf{8. 5 \%}$, respectively, compared to the YOLOv8 base model. 8.5%, respectively, compared to the YOLOv8 base model. Compared with other newly released YOLO models, our approach also exhibits good performance in the maritime tiny target detection task.