基于多模态遥感影像的边缘感知引导显著性检测

Edge aware guidance saliency detection based on multi-modal remote sensing image

摘要:

针对多模态遥感影像显著性检测鲁棒性差和检测精确度不佳等问题，提出一种基于多模态边缘感知引导的显著性检测方法，该方法主要由多模态遥感影像显著检测主干网络、跨模态特征共享模块和边缘感知引导网络构成。通过在特征提取主干网络中加入跨模态特征共享模块，使得不同模态间特征通过共享交互实现协同增强，并且抑制具有缺陷的特征信息。基于边缘感知引导网络，通过边缘图监督模块来检测边缘特征的有效性，从而生成准确边界。在3种显著目标检测遥感图像数据集上进行实验，平均的Fβ、平均绝对误差(MAE)、Sm分数分别为0.917 6，0.009 5和0.919 9。实验结果表明，提出的多模态边缘感知引导网络(MEGNet)适用于在多模态场景中进行显著性检测。

Abstract:

To address the problems of poor robustness and poor detection accuracy of multimodal remote sensing image saliency detection, this paper proposes a method based on the novel and efficient Multi-modal Edge aware Guidance Network(MEGNet), which mainly consists of a salient detection backbone network for multi-modal remote sensing images, a cross-modal feature sharing module and an edge aware guidance network. First of all, a Cross-modal Feature Sharing Module(CFSM) is used during feature extraction for remote sensing image pairs, which encourages different modalities to complement each other in the feature extraction process and suppresses the influence of defective feature data from different modalities. Secondly, based on the Edge Aware Guidance Network(EAGN), the effectiveness of edge features is detected through the edge map supervision module and the final salient detection map will have clear boundaries. Finally, experiments are carried out on three kinds of saliency objects detection remote sensing image datasets. The average Fβ, Mean Absolute Error(MAE) and Sm scores are 0.917 6, 0.009 5 and 0.919 9, respectively. The experimental results show that the proposed MEGNet is suitable for saliency detection in multi-modal scenes.