To address the low robustness of lane detection algorithms in diverse scenarios, we propose a lane detection framework that integrates traditional image processing techniques with deep learning algorithms. Firstly, an image preprocessing method is developed that concatenates the original three-channel RGB image with an enhanced single-channel image, creating a four-channel input for the subsequent network. This method effectively enhances lane features and enriches the diversity of training samples. Secondly, optimizations are applied to the feature extraction backbone network of the Ultra Fast structure-aware deep Lane Detection (UFLD) method, focusing on refining its stem structure, residual structure, and convolutional block stacking ratio. Furthermore, the multi-scale feature maps from the backbone network's output are fused, and a three-branch attention mechanism is introduced, thereby enhancing the capture of detailed and spatial information of lane lines. The optimized lane detection network (OPT-UFLD) demonstrates comprehensive improvements in F1 scores across diverse scenarios on the CULane dataset compared to the original model, achieving an overall F1 score increase of 1.3%.