Huang Jianwei, Wang Kangbo, Hou Yue, Wang Jiahe
College of Power Engineering, Naval University of Engineering, Wuhan 430033, China.
Maritime College, Fujian Chuanzheng Communications College, Fuzhou 350007, China.
Sensors (Basel). 2024 Dec 26;25(1):65. doi: 10.3390/s25010065.
Arbitrary-oriented ship detection has become challenging due to problems of high resolution, poor imaging clarity, and large size differences between targets in remote sensing images. Most of the existing ship detection methods are difficult to use simultaneously to meet the requirements of high accuracy and speed. Therefore, we designed a lightweight and efficient multi-scale feature dilated neck module in the YOLO11 network to achieve the high-precision detection of arbitrary-oriented ships in remote sensing images. Firstly, multi-scale dilated attention is utilized to effectively capture the multi-scale semantic details of ships in remote sensing images. Secondly, the interaction between the spatial information of remote sensing images and the semantic information of low-resolution features of ships is realized by using the cross-stage partial stage. Finally, the GSConv module is introduced to minimize the loss of semantic information on ship features during transmission. The experimental results show that the proposed method has the advantages of light structure and high accuracy, and the ship detection performance is better than the state-of-the-art detection methods. Compared with YOLO11n, it improves 3.1% of mAP@0.5 and 3.3% of mAP@0.5:0.95 on the HRSC2016 dataset and 1.9% of mAP@0.5 and 1.3% of mAP@0.5:0.95 on the MMShip dataset.
由于遥感图像存在高分辨率、成像清晰度差以及目标尺寸差异大等问题，任意方向的船舶检测变得具有挑战性。现有的大多数船舶检测方法难以同时满足高精度和高速度的要求。因此，我们在YOLO11网络中设计了一个轻量级且高效的多尺度特征扩张颈部模块，以实现对遥感图像中任意方向船舶的高精度检测。首先，利用多尺度扩张注意力有效地捕捉遥感图像中船舶的多尺度语义细节。其次，通过使用跨阶段局部模块实现遥感图像的空间信息与船舶低分辨率特征的语义信息之间的交互。最后，引入GSConv模块以最小化船舶特征在传输过程中语义信息的损失。实验结果表明，该方法具有结构轻量和精度高的优点，船舶检测性能优于现有最先进的检测方法。在HRSC2016数据集上，与YOLO11n相比，其mAP@0.5提高了3.1%，mAP@0.5:0.95提高了3.3%；在MMShip数据集上，mAP@0.5提高了1.9%，mAP@0.5:0.95提高了1.3%。
