基于改进YOLOv5的雾霾环境下船舶红外图像检测算法

从监控图像中准确检测船舶对于港区水域船舶交通智能监管具有重要意义。为解决雾霾条件下传统YOLOv5目标检测算法对船舶红外图像检测准确率低、小目标特征提取能力弱等问题，提出了基于Swin Transformer的改进YOLOv5船舶红外图像检测算法。为扩大原始数据集的多样性，综合考虑船舶红外图像轮廓特征模糊、对比度低、抗云雾干扰能力强等特点，改进算法提出基于大气散射模型的数据集增强方法；为增强特征提取过程中全局特征的关注能力，改进算法的主干网络采用Swin Transformer提取船舶红外图像特征，并通过滑动窗口多头自注意力机制扩大窗口视野范围；为增强网络对密集小目标空间特征提取能力，通过改进多尺度特征融合网络（PANet），引入底层特征采样模块和坐标注意力机制（CA），在注意力中捕捉小目标船舶的位置、方向和跨通道信息，实现小目标的精确定位；为降低漏检率和误检率，采用完全交并比损失函数（CIoU）计算原始边界框的坐标预测损失，结合非极大抑制算法（NMS）判断并筛选候选框多次循环结构，提高目标检测结果的可靠性。实验结果表明:在一定浓度的雾霾环境下，改进算法的平均识别精度为93.73%，平均召回率为98.10%，平均检测速率为每秒38.6帧；与RetinaNet、Faster R-CNN、YOLOv3 SPP、YOLOv4、YOLOv5和YOLOv6-N算法相比，其平均识别精度分别提升了13.90%、11.53%、8.41%、7.21%、6.20%和3.44%，平均召回率分别提升了11.81%、9.67%、6.29%、5.53%、4.87%和2.39%。综上，所提的Swin-YOLOv5s改进算法对不同大小的船舶目标识别均具备较强的泛化能力，并具有较高的检测精度，有助于提升港区水域船舶的监管能力。 

A Ship Detection Algorithm for Infrared Images under Hazy Environment based on an Improved YOLOv5 Algorithm

Accurately detecting ships from surveillance images is crucial for intelligent ship traffic surveillance around port waters. To address the issues of low accuracy and capability of small target feature extraction from traditional YOLOv5 object detection algorithms from the infrared images under hazy weather, an improved YOLOv5 algorithm based on Swin Transformer is proposed. To expand the diversity of the original dataset, the improved algorithm considers the characteristics of ship infrared images with strong resistance to cloud and fog interference but blurred image contour features and low contrast, and enhances the dataset based on an atmospheric scattering model. To enhance the algorithm's attention to global features during feature extraction, the backbone network of the improved algorithm uses Swin Transformer to extract ship infrared image features and expands the window view range using a multi-head self-attention mechanism controlled by a sliding window. To enhance the capability of extracting spatial features of dense small targets, a multi-scale feature fusion Path Aggregation Network (PANet) is improved by adding a bottom-up feature sampling module and a coordinate attention (CA) mechanism, in order to capture the position, direction, and cross-channel information of small target ships. To reduce false negatives and false positives, a complete intersection over union loss function (CIoU) is used to calculate the coordinate prediction loss of the original bounding box and combined with the non-maximum suppression algorithm (NMS) to judge and filter candidate boxes in a multi-loop structure to improve the reliability of object detection. Study results show that under certain concentrations of haze, the average recognition accuracy, recall rate, and detection rate of the improved algorithm is 93.73%, 98.10%, and 38.6 frames per second, respectively. Compared with the following algorithms: RetinaNet, Faster R-CNN, YOLOv3 SPP, YOLOv4, YOLOv5, and YOLOv6-N, the average recognition accuracy of the proposed algorithm is improved by 13.90%, 11.53%, 8.41%, 7.21%, 6.20%, and 3.44% respectively; and the average recall rate is improved by 11.81%, 9.67%, 6.29%, 5.53%, 4.87%, and 2.39%, respectively. The proposed Swin-YOLOv5s algorithm has a strong generalization ability for ship target recognition of different sizes and has a high detection accuracy, which helps to improve the surveillance capability of ships around port waters.