Semantic segmentation of high-resolution remote sensing images is a research hotspot in automatic feature extraction. In recent years, deep convolutional neural networks have achieved a lot of results in the direction of semantic segmentation, but the complexity of the remote sensing image feature information itself limits the network to play. This paper introduces multi-scale parallel atrous convolution to help the network capture contextual information on a larger scale. Without increasing the parameters, the network shows good recognition ability for both small and large targets. The extraction of semantic-level features will inevitably lead to loss of resolution. The method in this paper discards the pooling operation and uses atrous convolution to maintain the resolution. The output segmentation results use fully connected conditional random fields to fuse image-level features to compensate for detailed location information, To improve the semantic segmentation capabilities of the network on remote sensing images. The learning strategy of the network adopts the cosine annealing method to adjust the learning rate. This paper introduces the cosine annealing strategy with increasing period to stabilize the network learning process and obtain a suitable number of local optimal solutions. The integration of the local optimal solutions further improves the network"s pixel classification ability. The experimental results on the Gaofen Image Dataset show that the method in this paper can effectively identify objects with complex structures, and its overall accuracy and kappa coefficient are better than the commonly used semantic segmentation models.