In response to the challenge that existing image segmentation algorithms struggle to meet both accuracy and real-time requirements for unstructured road scenes, this paper proposes the Dual Resolution Semantic Segmentation Network (DRSNet) for real-time image segmentation. The network integrates a self-attention mechanism in high-resolution branches to eliminate redundant information and adopts a multi-scale feature map extraction strategy during the feature extraction phase of low-resolution branches, thereby enhancing the recognition of irregular boundaries and classes that are prone to confusion. Additionally, we propose an improved pyramid pooling module to reduce model parameters, thereby achieving a lightweight network architecture. By weighting feature maps, the network fuses bilateral features, mitigating the loss of critical information in unstructured scenes. The algorithm presented in this paper achieves an average pixel accuracy of 77.02% at a speed of 93.77 frames per second on an in-house unstructured dataset, with an average intersection over union ratio of 64.39%. Compared to the Deep Dual Resolution Networks (DDRNet), our approach shows an improvement of 11.23% and outperforms segmentation networks such as Segformer and PSPNet. The experimental results indicate that our model achieves a better balance between segmentation accuracy and inference speed.