Abstract:To address the issues of high redundancy, high computational cost, and low accuracy in current real-time semantic segmentation methods, this study proposes a novel real-time semantic segmentation approach based on deep semantic guidance and attention fusion. The proposed architecture employs MobileNetV3 as its backbone network, enhanced with an innovative deep dual-branch parallel structure. The semantic branch guides the correction of pixel points in the spatial branch, enhancing its detail features without increasing the parameter count. Additionally, an attention fusion module with multi-scale parallel branches is employed to achieve real-time computation and to improve segmentation accuracy through cross-spatial information aggregation. The method achieves inference speeds of 81.2 f/s and 147.6 f/s on the Cityscapes and CamVid datasets, respectively, with MIoU scores of 75.2% and 77.4%, and a parameter count of just 3.4M. Experimental results demonstrate that the proposed method effectively balances segmentation accuracy and speed while maintaining a small parameter count.