边缘信息引导多级尺度特征融合的显著性目标检测方法

针对基于FCN和U型网络架构的深度学习显著性目标检测方法提取的显著性图存在边界不清晰和结构不完整的问题，文中提出了一种基于边缘信息引导多级尺度特征融合网络(EGMFNet)。EGMFNet使用多通道融合残差块(RCFBlock)以嵌套的U型网络架构作为主干模型。同时，在网络的较低层级引入具有边缘信息引导的全局空间注意力模块(EGSAM)以增强空间特征及边缘特征。此外，在损失函数中引入了图像边界损失，用于提升显著性图的质量并在学习过程中保留更加清晰的边界。在四个基准数据集上进行实验，实验结果表明，文中方法的F值较典型方法提升1.5%、2.7%、1.8%和1.6%，验证了EGMFNet网络模型的有效性。

Salient object detection method based on multi-scale feature-fusion guided by edge information

In this paper, an Edge-information Guided Multi-scale Feature-fusion Network (EGMFNet) is proposed to solve the problems of unclear boundary and incomplete structure of saliency map extracted by deep learning saliency target detection method based on FCN and U-shaped network architecture. EGMFNet uses Residual muti-Channel Fusion Block (RCFBlock) and uses a nested U-shaped network architecture as the backbone model. At the same time, an Edge-information Guided Global Spatial Attention Module (EGSAM) is introduced at the lower level of the network to enhance spatial features and edge features. In addition, image boundary loss is introduced into the loss function, which is used to improve the quality of saliency map and keep clearer boundaries in the learning process. Experiments on four benchmark data sets show that the F values of the proposed method are increased by 1.5%, 2.7%, 1.8% and 1.6% compared with typical methods, which verifies the effectiveness of EGMFNet network model.