A Lightweight Network for Semantic Segmentation of Road Images Based on Improved DeepLabv3+

Abstract

Aiming at the problem that the existing semantic segmentation algorithms have a large number of model parameters and are difficult to be used on mobile devices, a lightweight network for semantic segmentation of road images based on improved DeepLabv3+ is proposed. The algorithm improves DeepLabv3+, uses the lightweight MobileNetv2 as the backbone network to reduce the complexity of the network model, and extracts low-level feature details at different depths; uses a Distinctive Atrous Spatial Pyramid Pooling (DASPP) to enhance the multiscale semantic information to obtain rich global contextual semantic information by feature extraction; after introducing an attention mechanism to enhance feature learning, a Detail-preserving Feature Fusion Network (DFFN) is used for feature fusion. Experiments are carried out on the Cityscapes dataset, and the model parameters of the improved DeepLabv3+ algorithm are 2.75M, which is 42.87M less than that of DeepLabv3+. The road image semantic segmentation model proposed in this paper can greatly reduce the number of model parameters while segmenting images well.