Aggregating multi-scale contextual features from multiple stages for semantic image segmentation
Dingchao Jiang, Hua Qu, Jihong Zhao, Jianlong Zhao, Meng-Yen Hsieh · Connection Science · 2021
Semantic segmentation plays a vital role in image understanding. Recent studies have attempted to achieve precise pixel-level classification by using deep networks that provide hierarchical features. These methods are trying to effectively utilise multi-level features that are extracted from the data and precisely reconstruct some characteristics of objects that are lost in producing high-level features. In this paper, we propose a multi-scale context U-net (MSCU-net) for semantic image segmentation. This network uses a multi-scale context block (MSCB) to aggregate multi-level features and employs the CRF layer to explicitly model the dependencies among pixels. This network significantly outperforms other state-of-the-art methods on both the PASCAL VOC 2012 and Cityscapes datasets.