Improved Semantic Segmentation of Urban Road Scenes with Deep Dual-Resolution Network

In response to the requirements for efficiency and accuracy in segmentation models in autonomous driving scenarios, an improved urban road semantic segmentation algorithm based on the Deep Dual-Resolution Network (DDRNet) has been proposed. Firstly, a Channel Residual Attention Module (CRAM) based on standard attention mechanisms is introduced at the end of two parallel branches of the DDRNet to focus on important channel features. Secondly, the backbone network incorporates Depthwise Over-parameterized Convolution (DO-Conv) to boost feature extraction performance. Finally, a Deep Aggregation Pyramid Pooling Module (DAPPM) with Hybrid Dilated Convolution (HDC) is designed to capture multi-scale feature information, and Group Normalization (GN) is utilized to mitigate the impact of reduced batch size. The experimental results demonstrate that the proposed method achieves a Mean Intersection over Union (MIoU) of 77.0% on the Cityscapes dataset, representing an improvement of 1.1% over the baseline model. Additionally, the segmentation speed reaches 106 Frames Per Second (FPS), achieving an optimal balance between segmentation accuracy and computational efficiency.