跨层多尺度特征融合的边缘检测模型

Cross-layer Multi-scale Feature Fusion Model for Edge Detection

充分利用多尺度信息对于提高不同尺度对象的边缘检测十分重要,因此提出一种跨层多尺度特征融合的边缘检测模型。使用残差网络作为模型的主干网络,为了增大模型的感受野,在最后一个阶段使用扩张卷积,同时在每个Bottleneck模块中添加了全局注意力模块。此外,使用多尺度融合模块对特征图提取更准确的边缘,使用跨层融合模块将高层特征和低层特征进行融合。在BIPED数据集,BSDS500数据集和NYUDv2数据集上进行评估,并在BIPED数据集上实现了0.866的ODS-F值和0.871的OIS-F值,比在BIPED数据集上的最新技术分别提高了0.7%和0.4%。

Making full use of multi-scale information is very important to improve the edge detection of objects at different scales,so a cross-layer multi-scale feature fusion model for edge detection is proposed.The residual network is used as the back⁃bone network,in order to increase the receptive field of the model,dilated convolution is used in the last stage,and a global atten⁃tion module is added to each Bottleneck module.In addition,a multi-scale fusion module is used to extract more accurate edges of the feature map,and a cross-layer fusion module is used to fuse high-level features and low-level features.The method is evaluated on the BIPED dataset,BSDS500 dataset and NYUDv2 dataset,and it achieves ODS F-measure of 0.866 and OIS F-measure of 0.871,0.7%and 0.4%higher than current state-of-the-art on BIPED respectively.