A content-guided attention network (CGANet) was proposed to effectively utilize attention mechanisms and improve the accuracy of lane detection. To enhance the model’s ability to capture contextual information, a content-guided attention (CGA) mechanism was introduced into the model, emphasizing more useful information encoded in the features while reducing the influence of irrelevant information. To reduce the impact of scale differences on model performance, a balanced feature pyramid network (BFPN) was proposed to achieve the balanced fusion of multi-scale features. An ROI (Region of Interest) extractor was introduced to address the issue of missing visual cues. Additionally, the cross-entropy loss was added to the loss function as an auxiliary classification loss to encourage the model to generate clearer probability distributions. Experimental results on multiple lane detection datasets demonstrated that, compared with the cross-layer refinement network (CLRNet) algorithm, the proposed method improves F1 index by 0.65, 0.18 and 0.29 percentage points on CULane, Tusimple and CurveLanes datasets, respectively.