Lane detection is one of the critical tasks to ensure safe driving and intelligent decision-making for vehicles. However, there are still many challenges to accurately extract lanes in complex real scenes, such as illumination variations, nighttime environments and non-standard roads. Recent advancements in model evaluation have introduced stricter metrics, particularly a high Intersection-over-Union (IoU) threshold, to rigorously assess detection performance and improve the precision and reliability of lane extraction. Inspired by the human cognitive process of perceiving and reasoning lanes during driving, this paper proposes a Multi-scale Dynamic Weighted Lane Feature Network, which comprises two core components: Lane Feature Fusion Module (LFM) and Shared Clear Focus Attention Module (CFAM). The LFM utilizes the CFAM-Residual Block (CRBlock) and Sigmoid-based reweighting strategy to adaptively extract and fuse multi-scale lane features. The Shared CFAM enhances feature representation and significantly reduces parameter redundancy. Extensive experiments on both CULane and TuSimple datasets demonstrate that in comparison to prior approaches, the proposed model exhibits superior lane detection performance in complex real-world scenes, particularly demonstrating significant advantages under a high IoU threshold. Our code has been publicly released for further research and applications: https://github.com/EADMO/DLFNet .