Abstract
Detecting ship targets in remote sensing images within complex scenarios faces numerous challenges. The limited feature information of small-scale targets and their random orientation angles often result in missed and false detections. To address these issues, this paper proposes a Multi-Scale Rotated Detection Network (MSRO-Net) for detecting rotated ship targets in remote sensing images. The network adopts a CNN-Transformer hybrid architecture for collaborative feature extraction and integrates our proposed Coordinate-Aware Pyramid Feature Aggregation module (CAPP). This backbone network retains the capability of local feature extraction while also connecting global context and capturing long-range dependencies. The model can gather information from any position in the sequence, extract contextual information of targets at different scales, and enhance global feature representation. Additionally, this paper proposes an Upsampling Feature Reconstruction Pyramid (ARFPN-C), based on Adaptive Rotated Convolution (ARC). This network combines ARC adaptive rotating convolution with a multi-scale feature fusion mechanism to enhance the model's perceptual capability, addressing the issues of limited target feature information and random orientation angles. The proposed algorithm is validated on two public remote sensing image datasets, HRSC2016 and DOTA. The mAP07, mAP12 on the HRSC2016 dataset, as well as the mAP in the ship category of the DOTA dataset, show a significant advantage over other commonly used object detection algorithms, with accuracies of 90.70%, 98.98%, and 89.46%, respectively. These results further validate the accuracy and effectiveness of the proposed MSRO-Net in object detection.