In advanced driving assistance system, lane detection is an significant component of traffic scene analysis. In the past few years, as the rapid development of deep learning, more and more lane detection methods based on deep learning have emerged. But most methods are expensive and time-consuming. In this paper, a novel lane detection method, called the LSNet, is proposed to detect the lane. In the proposed LSNet, it adopts the consecutive depth-wise separable convolution in the feature encoding stage. Furthermore, the upsampling and dawnsampling result on each layer is fused to raise the decoding accuracy in the feature decoding stage. In addition, extensive experiments on the TuSimple lane dataset are conducted to evaluate the robustness of the proposed method. From the experiments, it demonstrates that the proposed LSNet is superior to other lane detection methods, and can effectively enhance the detection rate while maintaining a better accuracy of lane detection.