To enhance segmentation accuracy and robustness for complex road structures under severe occlusion and fragmentation in high-resolution remote sensing images, this study proposes MSAR-Net—an attention refinement network integrating multi-scale residual learning and efficient dual-kernel perception. The network adopts an encoder-decoder architecture, embedding a Multi-Scale Residual Learning module within the LinkNet50 backbone. This module partitions input features and processes them through heterogeneous-kernel cascading to achieve cross-scale semantic fusion, expanding receptive fields while preserving feature consistency. An efficient Dual-Kernel Attention Module is designed to jointly model spatial and channel dependencies via dual parallel convolutions with adaptive fusion mechanisms, strengthening long-range contextual modeling capabilities. During training, a combined Dice loss and BCE loss mitigates class imbalance and enhances recognition of thin road structures. Evaluated on three benchmark remote sensing road datasets (Massachusetts, DeepGlobe, and CHN6-CUG), MSAR-Net achieves F1 scores of 0.802 5, 0.782 3, and 0.781 7, with IoU values of 0.670 1, 0.642 4, and 0.641 6 respectively. Key metrics surpass state-of-the-art methods by average margins of 0.72% (F1) and 1.20% (IoU). Ablation studies and large-scale image tests confirm superior generalization capability and topological preservation in complex urban scenarios with dense occlusions. The framework effectively mitigates limitations of conventional methods regarding insufficient receptive fields and road fragmentation while maintaining computational efficiency, providing a precise, stable, and generalizable solution for high-resolution remote sensing road extraction.