Abstract: In the task of lane line detection, atrous convolution is widely used due to the characteristics of lane lines and the need to obtain a wider receptive field. However, in order to obtain large-scale information, atrous convolution will cause the loss of information near the convolution point. To solve the problems, a lane line detection algorithm was proposed based on multi-scale composite convolution and image segmentation fusion. First, atrous convolution, full convolution and standard convolution of different sizes were combined to compensate for the loss of information caused by atrous convolution. And then, fusing semantic segmentation and instance segmentation, an image segmentation fusion module was arranged to enhance the attention of instance segmentation network to global features. Finally, a weighted cross-entropy loss function was designed to train and optimize the network. The experimental results show that the overall F_1measure of the algorithm in the CULane dataset can achieves 74.9%. Comparing with other algorithms, the overall performance of the proposed detection algorithm is better and improved in various challenging environments.