R-YOLOv5: Rotated Anchor Boxes with Angular Prediction for Real-Time Ship Detection in Remote Sensing Imagery

Existing object detection networks mostly focus on horizontal detection, struggling to handle ship targets with large aspect ratios and dense arrangements, thus failing to support subsequent ship model identification. To address this, this paper presents R-YOLOv5, an inclined bounding box detection model based on an improved YOLOv5 algorithm. By introducing angle parameters to define inclined anchor boxes and adding a rotation angle prediction channel, the model can learn and predict these boxes. Additionally, the model is optimized by replacing the activation function with the superior HardSwish function and adopting the soft-NMS algorithm instead of the original NMS algorithm. Experiments show that R-YOLOv5 effectively detects ship targets using inclined angle anchor boxes, capturing their aspect ratio, direction, and contour. With high accuracy and real-time inference speed, it can handle ship target detection in remote sensing images, laying a solid foundation for ship tracking and identification tasks.