【摘要】 针对传统神经网络易受类油膜伪影的影响，导致对海面薄油膜的识别率低、分割精度差等问题，文中提出一种基于UNet的RDW-UNet模型。首先，将UNet的主干提取网络替换为网络结构更深的ResNet50，利用ResNet50更强的特征提取能力与残差结构保留细节信息，解决了深层网络梯度消失的问题；其次，以级联结构将可变形卷积模块与CBAM相结合，提出DCBAM，以适应海面油污形状不规则的特点；最后，针对油污像素占比较低的特点，提出类别敏感型加权交叉熵损失函数，使网络收敛偏向目标类别。实验结果表明，提出的改进RDW-UNet模型相较于UNet，在像素识别精确率、召回率、平均交并比等指标上分别提升了5.94%、11.1%、12.51%。该研究为海面油污识别提供了一种有效方法。更多还原

【Abstract】 Traditional neural networks are prone to interference from oil-like artifacts, which leads to low recognition rate and poor segmentation accuracy for thin oil films on the sea surface. In view of the above, this paper proposes an RDW-UNet model based on the UNet. Firstly, the backbone feature extraction network of the UNet is replaced with the ResNet50 which has a deeper network structure. By leveraging stronger feature extraction capability and residual structure of the ResNet50, the model preserves detailed information and mitigates gradient vanishing in deep networks. Secondly, a cascaded structure combining deformable convolution modules with CBAM(convolutional block attention module) is introduced, termed DCBAM, to adapt to the irregular shapes of marine oil spills. Finally, in view of the low pixel proportion of oil spills, a category-sensitive weighted cross-entropy loss function is proposed to bias network convergence toward the target category. Experiments demonstrate that the pixel recognition precision, recall rate, and mean intersection over union(mIoU) of the proposed RDW-UNet model are improved by 5.94%, 11.1%, and 12.51% respectively, in comparison with those of the original UNet. In a word, this method provides an effective method for oil spill identification on the sea surface.更多还原