A Real-time Semantic Segmentation Algorithm Based on Two-branch Network

Abstract:
In recent years, the performance of convolutional neural network has been getting better and better in image semantic segmentation. In order to balance the accuracy and speed of semantic segmentation in the field of autopilot, this paper presents a real-time semantic segmentation algorithm based on two-branch network. Firstly, the model uses a lightweight network module which can be deployed on mobile devices as the detail path branch. It can quickly sample the feature map and obtain a large receptive field, it encodes higher level semantic information at the same time. In addition, a spatial path branch with smaller strides is used to preserve spatial information and obtain higher spatial characteristic images. Finally, an efficient feature fusion network is used to fuse the deep features of the context detail path and the shallow features of the spatial path. The experiment results show that the scheme has good accuracy and real-time performance for the semantic segmentation of urban street scenes.