多深度特征增强与顶层信息引导的边缘检测网络

Edge Detection Network with Multi-Depth Feature Enhancement and Top-Level Information Guidance

针对现有边缘检测网络在复杂自然场景下的检测结果仍存在边缘缺失、噪声过多等问题,提出多深度特征增强与顶层信息引导的边缘检测网络.首先,采用UNet++作为主干网络提取不同深度的特征,并通过特征叠加使不同尺度的边缘更加完整;然后,在每个卷积层的侧输出后添加特征增强模块,通过引入空洞卷积增大感受野,增强多尺度信息;最后,设计顶层信息引导模块,通过在跳跃连接中引入高层的语义特征增强边缘检测效果.实验结果表明,在BSDS500,NYUDv2和Multicue这3个数据集上进行训练均取得了较好的效果,其中,BSDS500数据集上的ODS,OIS和AP指标分别达到了0.821,0.839和0.869,整体上高于现有边缘检测网络,且噪声少,主观效果也更接近真值.

The existing edge detection networks still have problems such as missing edges and excessive noise in complex natural scenes.Therefore,an edge detection network with multi-depth feature enhancement and top-level information guidance is proposed.First,UNet++is used as the backbone network to extract features of different depths,and the edges of different scales are made more complete by feature superposition.Then,a feature enhancement module is added after the side output of each convolution layer to increase the receptive field and enhance the multi-scale information by introducing the dilated convolution.Finally,a top-level information guidance module is designed to enhance the edge detection effect by introducing top-level semantic features into jump connection.The experimental results show that training on the three datasets of BSDS500,NYUDv2 and Multicue has achieved good results.On the BSDS500 dataset,the ODS,OIS and AP of this network reach 0.821,0.839 and 0.869 respectively,which is generally higher than the existing edge detection networks.Moreover,the result has less noise and the subjective effect is closer to the ground truth.