Enhanced Multi-Scale Ship Detection Through Joint Multi-Source Image Training

In order to improve the detection performance of traditional single-source remote sensing image ship detection methods in terms of anti-interference capability and multi-scale targets. We develop an improved algorithm based on YOLOv8, called Sea-YOLOv8, and obtain the multi-source remote sensing image detection model SAR-YOLOv8 through migration learning.First, Efficient Multi-Scale Convolution (EMSConv) is used to improve the ability of the C2f module in capturing the features of ships at different scales by using different sizes of convolution kernels. scale ship features using different sized convolution kernels. Secondly, Large Separable Kernel Attentio (LSKAtt) is applied to the SPPF module to decompose the 2D convolution kernel into horizontal and vertical 1D kernels, which improves the computational efficiency while maintaining the high efficiency image processing capability. Experimental results on the SeaShips and SSDD datasets show that Sea-YOLOv8 has a good performance in terms of R, mAP, model parameters and FPS have achieved satisfactory results. Finally, the optimal transfer learning strategy was determined through layer-by-layer freezing, training and validation, which led to the construction of the multi-source remote sensing image recognition model SAR-YOLOv8.