Improved Edge Detection Model Based on HED

Abstract

Influenced by the HED network structure, how to further solve the image multi-scale multi-level representation problem and optimize the detection of the overall image. We propose an end-to-end network structure. While enhancing the robustness of backbone network, the BN layer is used to solve the problem of network gradient dispersion, the attention mechanism is added to strengthen the role of deep supervision and final fusion module. The network can learn related semantic features autonomously, obtain richer image information, and solve the problem that HED cannot fully extract features, and high-level and low-level information are simply fused together. Considering the modification on the basis of the HED network structure, the BN layer and the SE structure are added, and the manner of downsampling is modified. The experimental results show that the edge effect from the BSDS300 data set is good, and the running speed is better than the HED model.