Real-Time Semantic Segmentation of Road Scenes via Hybrid Dilated Grouping Network

Article Real-Time Semantic Segmentation of Road Scenes via Hybrid Dilated Grouping Network Yan Zhang 1, Xuguang Zhang 1,*, Deting Miao 1, and Hui Yu 2 1 School of Communication Engineering, Hangzhou Dianzi University, Hangzhou 310018, Zhejiang, China 2 cSCAN, University of Glasgow, G12QB, United Kingdom * Correspondence: [email protected] Received: 10 January 2024 Accepted: 23 April 2024 Published: 25 March 2025 Abstract: Real-time semantic segmentation is a critical step for various real-world application scenarios such as autonomous driving systems. How to achieve a high accuracy while keeping a high inference speed has become a challenging issue for real-time semantic segmentation. To tackle this challenge, we propose a Hybrid Dilated Grouping Network (HDGNet) for real-time semantic segmentation of outdoor scenes in this study, which not only improves the accuracy of image segmentation, but also considers the inference speed. To reduce model parameters to speed up inference, we propose to use factorization convolution to replace ordinary two-dimensional convolution. However, simply reducing the amount of model parameters may sacrifice segmentation accuracy. We thus further introduce dilated convolution to extract multi-scale spatial information. The HDG module is constructed by combining factorization convolution and dilated convolution, which not only reduces the model parameters and improves the model inference speed, but also extracts local and more contextual information. And furthermore, to enhance the feature expression ability of the network, we introduce a channel attention mechanism to capture the information interaction between channels. After obtaining the shallow features and deep high-level semantic information, we design the skip layer connections to fuse the feature branches from different stages to improve the segmentation accuracy. The experiments conducted on the widely used datasets show that the proposed model achieves superior real-time performance over existing methods but using significantly fewer model parameters.