Xue Wentao, He Maozheng, Zhang Yincheng, Ye Hui
College of Automation, Jiangsu University of Science and Technology, Zhenjiang 212100, China.
Sensors (Basel). 2023 Aug 8;23(16):7027. doi: 10.3390/s23167027.
The efficiency and accuracy of ship detection is of great significance to ship safety, harbor management, and ocean surveillance in coastal harbors. The main limitations of current ship detection methods lie in the complexity of application scenarios, the difficulty in diverse scales object detection, and the low efficiency of network training. In order to solve these problems, a novel multi-target ship detection method based on a decoupled feature pyramid algorithm (DFPN) is proposed in this paper. First, a feature decoupling module is introduced to separate ship contour features and position features from the multi-scale fused features, to overcome the problem of similar features in multi-target ships. Second, a feature pyramid structure combined with a gating attention module is constructed to improve the feature resolution of small ships by enhancing contour features and spatial semantic information. Finally, a feature pyramid-based multi-feature fusion algorithm is proposed to improve the adaptability of the network to changes in ship scale according to the contextual relationship of ship features. Experiments on the multi-target ship detection dataset showed that the proposed method increased by 6.3% mAP and 20 FPS higher than YOLOv4, 7.6% mAP and 36 FPS higher than Faster-R-CNN, 5% mAP and 36 FPS higher than Mask-R-CNN, and 4.1% mAP and 35 FPS higher than DetectoRS. The results demonstrate that the DFPN can detect multi-target ships in different scenes with high accuracy and a fast detection speed.
船舶检测的效率和准确性对于沿海港口的船舶安全、港口管理和海洋监测具有重要意义。当前船舶检测方法的主要局限性在于应用场景的复杂性、不同尺度目标检测的困难以及网络训练的低效率。为了解决这些问题，本文提出了一种基于解耦特征金字塔算法（DFPN）的新型多目标船舶检测方法。首先，引入特征解耦模块，从多尺度融合特征中分离出船舶轮廓特征和位置特征，以克服多目标船舶中特征相似的问题。其次，构建结合门控注意力模块的特征金字塔结构，通过增强轮廓特征和空间语义信息来提高小船舶的特征分辨率。最后，提出基于特征金字塔的多特征融合算法，根据船舶特征的上下文关系提高网络对船舶尺度变化的适应性。在多目标船舶检测数据集上的实验表明，所提方法比YOLOv4的平均精度均值（mAP）提高了6.3%，帧率提高了20帧每秒；比Faster-R-CNN的mAP提高了7.6%，帧率提高了36帧每秒；比Mask-R-CNN的mAP提高了5%，帧率提高了36帧每秒；比DetectoRS的mAP提高了4.1%，帧率提高了35帧每秒。结果表明，DFPN能够在不同场景下高精度、快速地检测多目标船舶。
