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

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.