Cascaded hierarchical atrous spatial pyramid pooling module for semantic segmentation

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