Remote sensing images of ships is characterized by small target sizes, complex backgrounds, and significant attitude changes. Traditional ship detection algorithms focus on improving detection accuracy while neglecting model size and real-time performance, thereby limiting their practical application on resource-constrained devices. A lightweight Rotated Fusion Detection Network RFDNet adapted to remote sensing ship images is proposed to address the above problems. Considering that the remote sensing ship images are taken at a long distance, resulting in small target sizes and rich background information in the images, ACFNet is designed to improve the detection accuracy of small ship targets by fully extracting local feature information and global spatial sensing information. To avoid accuracy degradation when detecting ship targets with different attitudes, a rotating bounding box loss function is introduced, which utilizes the orientation information of rotating targets for obtaining a more accurate bounding box regression loss, thereby improving the detection performance of ship targets rotating in any direction; for the problem of increasing parameter counts brought about by increasing the accuracy of the model, a lightweight convolution is introduced into the feature fusion part, which combines the convolution, the depth separable convolution, and the channel blending to reduce the number of parameters in the model. Through comparative and ablation experiments, it has been demonstrated that RFDNet achieved mAPs of 97.63% and 81.63% on the HRSC2016 and DOTA datasets, respectively, while reducing the model parameters to 2.99M. This not only effectively improved detection accuracy but also realized a lightweight model design, providing a new insight for the application of remote sensing ship detection algorithms to resource-constrained devices.