Cascaded hierarchical atrous spatial pyramid pooling module for semantic segmentation

•A new hierarchical structure consisting of multiple levels of atrous convolution layers.•A novel Cascaded Hierarchical Atrous Spatial Pyramid Pooling (CHASPP) network by cascading the hierarchical structures is constructed.•Global features and context information is extensively collected to form a more reasonable and accurate prediction.•The segmentation method outperforms the state of the arts. Atrous Spatial Pyramid Pooling (ASPP) is a module that can collect semantic information distributed in different scopes. However, because of the limited number of sampling ranges of ASPP, much valuable global features and contextual information cannot be sufficiently sampled, which degrades the representation ability of the segmentation network. Besides, due to the sparse distribution of the effective sampling points in the atrous convolution kernels of ASPP, large amount of local detail characteristics are easily discarded. To overcome the above two problems, a new Cascaded Hierarchical Atrous Pyramid Pooling (CHASPP) module, consisting of two cascaded components, is proposed. Each component is a hierarchical pyramid pooling structure containing two layers of atrous convolutions with the aim to densify the sampling distribution. On the foundation of such a hierarchical structure, another same structure is appended to form a cascaded module which can further enlarge the diversity of sampling ranges. Based on this cascaded module, not only rich local detail characteristics can be comprehensively presented, but also important global contextual information can be effectively exploited to improve the prediction accuracy. To demonstrate the performance of our CHASPP module, experiments on the benchmarks PASCAL VOC 2012 and Cityscape are conducted.