<p>Efficient and accurate semantic segmentation is particularly important in scene parsing. Despite the significant progress made by deep convolutional neural networks in semantic segmentation, lightweight models still struggle to achieve a balance between real-time performance and accurate boundary segmentation, particularly for complex and small objects in road scenes. To address these challenges, a lightweight pixel-wise boundary information refinement network is proposed for road scene segmentation. First, a multi-scale boundary information extraction module (MBIE) is proposed for deep pixel feature refinement. Second, a low-level boundary information extraction module (LBIE) is proposed for shallow feature refinement extraction. Finally, a feature reuse fusion module (FRF) is designed for boundary pixel information fusion. The MBIE module introduces a channel separation pyramid and cross-scale self-attention to capture long-range dependencies. The LBIE module uses dilated self-attention to enhance low-level feature semantics. The FRF module employs dual-attention mechanisms to fuse semantic and spatial information effectively. Experimental results on CamVid and Cityscapes datasets show that the proposed method achieves a mIoU of 74.61% and 69.33%, respectively, outperforming state-of-the-art scene segmentation methods in terms of segmentation accuracy while maintaining real-time performance.</p>