为了高效地对城市景观等复杂场景进行分割解析，本文结合高分辨率网络（HRNet），通过金字塔池化模块（Pyramid pooling module，PPM）补充全局上下文信息，提出了一个高分辨率场景解析网络。首先，以HRNet为基干特征提取网络，并利用空洞可分离卷积改进其大量使用的残差模块，在减少参数量的同时提高了对于多尺度目标的分割能力；其次，利用混合空洞卷积框架设计了多级空洞率，在稠密感受野的同时减小了网格问题的影响；然后，设计了多阶段的连续上采样结构以改进HRNetV2简单的后融合机制；最后，使用改进的可适应不同图像分辨率的金字塔池化模块聚合不同区域的上下文信息获得高质量的分割图。在城市景观数据集（CityScapes）上仅以16.4 Mbit的参数数量实现了83.3% MIOU的精度，在Camvid数据集也取得了良好的效果，实现了更加可靠、准确、低计算量的基于语义分割的场景解析方法。

In order to efficiently segment and analyze complex scenes such as urban landscapes, this paper combines the high-resolution network (HRNet) and supplements the global context information through the pyramid pooling module, and proposes a high-resolution scene analysis network. Firstly, HRNet was used as the backbone feature extraction network, and the atrous separable convolution was used to improve its widely used residual module, so as to reduce the amount of parameters and improve the segmentation ability of multi-scale targets. Secondly, the mixed cavity convolution framework was used to design the multi-level cavity rate, which can dense the receptive field and reduce the influence of the grid problem. Then, a multi-stage continuous up-sampling structure was designed to improve the simple post fusion mechanism of HRNetV2. Finally, the improved pyramid pooling module which can adapt to different image resolutions was used to aggregate the context information of different regions to obtain high-quality segmentation images. The accuracy of 83.3% MIOU is achieved with only 16.4 Mbit parameters on the CityScapes urban landscape dataset, and good results are also achieved on the Camvid dataset. A more reliable, accurate, and low-computing scene analysis method based on semantic segmentation has realized.