Abstract: Image semantic segmentation is an important part of image understanding, which is applied to automatic driving. In this paper, we use the Pascal VOC 2012 data set and ResNet101 as the basic network.We propose semantic segmentation algorithm based on separable dilated convolution and improved normalization method to solve the problem of information loss andslow speed Firstly, we combineseparable convolution and dilated convolution to extract the last three layers' output of ResNet101.Compared with standard dilatedconvolution, separable dilatedconvolution accelerates the training, validation and prediction of the network. Then, in the semantic segmentation, the instance normalization method is applied and compared with the application batch normalization to verify the effectiveness of batch normalization. Finally, two normalization methods combining batch normalizationand instance normalizationare proposed to improve the effect of semantic segmentation. This method is tested in Pascal VOC 2012 data set. The results show thatour methodaccelerates the training, validation and prediction of the network. Thehighest mean intersection over union ofthis method in Pascal VOC 2012 data set is 80.62%.