FMR-YOLO: Infrared Ship Rotating Target Detection Based on Synthetic Fog and Multiscale Weighted Feature Fusion

Infrared ship detection has important application value for ensuring navigation safety and real-time monitoring of the sea surface. It is also of great significance in marine intelligent defense and has become an important research branch in the field of computer vision. Affected by the weather at sea and the limitations of infrared cameras, infrared ship images often have the problems of small targets being submerged by noise and low information entropy, which bring great challenges to infrared ship detection. In this article, an infrared ship rotating target detection algorithm FMR-YOLO based on synthetic fog and multiscale weighted fusion is proposed. Our algorithm first corrects the noisy labels of the original dataset due to misclassification and constructs an infrared ship dataset (ISD) containing different concentrations of haze through an improved dark channel prior (DCP) algorithm. Second, in order to avoid the loss of small target features and information as the network deepens, a weighted feature pyramid network (FPN) based on dilated convolution (DWFPN) is proposed. DWFPN weights the fusion of features at different levels based on the attention mechanism to achieve high-quality information interaction. Finally, in view of the large aspect ratio and arbitrary direction of the ship target, rotation detection is introduced to obtain more accurate detection boxes and ship navigation direction information. The experimental results show that compared with the standard YOLOv7, the improved algorithm achieves a mean average accuracy (mAP) of 92.7%, and the recall rate and precision rate are improved by 2.3% and 3%, respectively. Our code and R-ISD dataset are available at: https://github.com/denghuimin1/FMR-YOLO .