基于UNet++网络与多边输出融合策略的船舶检测模型

船舶检测与识别技术的发展对海上监视及服务工作起重要作用，目前卫星遥感图像船舶目标检测存在背景复杂、船舶尺度变化大等问题，妨碍了海上威胁事件的预测及海上工作效率的提高。提出一种融合多尺度特征信息的目标检测模型，采用UNet++网络进行目标检测提取卫星图像特征，并将全局信息和细粒度信息相融合生成具有高空间精度的中间特征图。在此基础上，使用MSOF策略融合不同语义层次的特征信息，生成最终的检测特征图，以提高船舶目标检测与识别的精度，并通过将二元交叉熵损失函数与Dice系数损失函数结合使用，降低数据集中样本不均衡对模型准确度的影响。基于空客船舶数据集的实验结果表明，该模型能够对遥感图像中的船舶目标进行精准的检测识别，其Dice系数、IOU系数评估值分别为97.3%、96.8%，优于ResNet-34、UNet++等模型。

Ship Detection Model Based on UNet++ Network and Multiple Side-Output Fusion Strategy

The development of ship detection and recognition technology plays an important role in marine surveillance service.At present, there are some problems in ship target detection in satellite remote sensing images, such as complex background and large changes in ship scale, which hinder the prediction of threat events and the improvement of marine operation efficiency.To address the problems of complex background and large changes in ship scale in satellite remote-sensing image ship target detection, a target detection model integrating multiscale feature information is proposed, and the UNet++ network is used for target detection to extract the satellite image features, what's more, global and fine-grained information are fused to generate an intermediate feature map with high spatial accuracy.On this basis, Multiple Side-Output Fusion (MSOF) strategy is used to fuse the feature information of different semantic levels and generate the final detection feature map to improve the accuracy of ship target detection and recognition.Moreover, the influence of the sample imbalance in the dataset on the accuracy of the model is reduced by combining the binary cross-entropy loss function with the Dice coefficient loss function.The experimental results on the Airbus ship dataset show that the model can accurately detect and recognize ship targets in remote-sensing images, and the evaluation values of the Dice coefficient and Intersection Over Union (IOU) coefficient are 97.3% and 96.8%, respectively, which are better than those of ResNet-34, UNet++, and other models.