Due to their complex geometric and visual features, curved lane lines are more challenging to detect than straight lane lines. Addressing the common issue of insufficient detection accuracy in existing curved lane detection algorithms, this study proposes a deep learning-based method for detecting curved lane lines using multi-scale feature fusion and composite structural loss optimization, aiming at efficiently and accurately extracting and recognizing curved lanes. In image preprocessing, this paper employs a region-specific cropping strategy, preserving the key areas of the image through selected cropping ratios based on the relative position of lane lines within the image. This method effectively reduces environmental interference and enhances the model’s prediction speed. Based on the optimized image input, a deep learning model is constructed, integrating a ResNet34 backbone, feature pyramid network (FPN), and dynamic convolution module, using multi-scale feature fusion technology to significantly improve the accuracy of lane line detection. To further optimize detection performance, the paper introduces a novel composite structural loss function that combines position and shape losses. This function not only optimizes the estimation of lane line positions, but also enhances shape continuity in curves. Tested on the CULane dataset under curved scenarios, the proposed method achieves an $F_1 $ score of 85.54, demonstrating its high accuracy and robustness.