Multi-Branch Supervised Learning on Semantic Segmentation

Autonomous driving is a research and development focus of various countries in recent years, and its popularity cannot be separated from the support of semantic segmentation. Aiming at the problem that the shallow layers of the semantic segmentation model cannot obtain effective and immediate supervision, this paper proposes a multiple branch supervised method, which can simultaneously supervise the shallow layers of model during the training process. In order to increase the information interaction among multiple network layers, a module named “Layer Attention Mechanism” is proposed, which can increase the network's attention to more effective information when the network layers are merged. Based on UNet, experiments in the autonomous driving datasets show that the structure proposed in this paper is better than the UNet3+ network, which improves the accuracy of semantic segmentation and optimizes the effect of semantic segmentation.