A Novel Semantic Segmentation Algorithm Using a Hierarchical Adjacency Dependent Network

Recent semantic segmentation networks mainly focus on how to fuse multi-level features from classification networks to improve segmentation accuracy. Some researches evenly emphasize the correlation of pixels in a global region, such as conditional random field (CRF). However, the strong correlation feature of pixels in a limited region is less considered in the previous researches and the remedy ability of the correlation of local pixels in semantic segmentation is severely ignored. To deal with this problem, we introduce a hierarchical adjacency dependent network (HadNet), in which an adjacency dependency module (ADM) is constructed by calculating and utilizing the impact fact of the pixel in different directions to classify the pixel. We explored the correlation of adjacent pixels and feature coverage in different feature levels to improve the segmentation accuracy. We evaluate our method on the popular Pascal VOC 2012 test set, and achieve a comparable result of mIOU accuracy of 79.8 with the state of art methods, such as DeepLabv3 and Exfuse. Further, we discuss and analyze the data distribution of COCO dataset for deeply understanding the feature correlation and coverage in semantic segmentation.