Multi-Scale Remote Sensing Ship Target Detection in Complex Backgrounds

In complex backgrounds, the features of multi-scale ship targets may be disturbed by background noise, leading to inaccurate feature extraction, and feature conflicts may occur during the fusion of features at different scales. To address these issues, this paper takes Oriented R-CNN as the baseline and makes multiple improvements to the backbone and neck networks. Firstly, ResNet50 is replaced with ConvNeXt Tiny to enhance the model's feature extraction capability in complex backgrounds. Secondly, by introducing Spatial Pyramid Pooling - Fast (SPPF) and the Efficient Multi-scale Attention mechanism (EMA), the EMAB-SPPF module is constructed. Combining ConvNeXt Tiny with the EMAB-SPPF module, the MSFE-ConvNeXt is built to improve the multi-scale feature extraction capability of the model in complex backgrounds. Finally, Multi-Scale Feature Enhancement and Fusion module (MSFEF) is developed and introduced into the neck network to form the MSFEF-FPN module, which is used for feature enhancement and improving the model's multi-scale feature fusion ability. Experiments conducted on the processed HRSC2016 and FGSD2021 datasets show that the modified model achieves improvements in mean Average Precision (mAP) of 3.3% and 3.8%, respectively. Experimental results on the datasets demonstrate that the proposed model improvements achieve significant gains compared to the baseline model.