Class semantic enhancement network for semantic segmentation

Abstract

Existing semantic segmentation methods favor class semantic consistency by extracting long-range contextual features through multi-scale and attention strategies. These methods ignore the relations between feature channels and classes, which are essential to represent consistent class semantics. To this end, we propose the Class Semantic Enhancement Network (CSENet) which boosts the segmentation performance of a backbone network in a coarse-to-fine manner. CSENet consists of two basic modules – (1) the Class Semantic Channel Graph Module (CSCG) module, which captures inter-dependencies among channels and strengthens the channel-class relation, and (2) the Class Prior Fully Convolution (CP-FC) module, which utilizes the channel-class relation as class priors to refine the segmentation results. Extensive experiments have demonstrated that the proposed CSENet is able to learn discriminative feature representations and achieves state-of-the-art performance on three benchmark datasets, including PASCAL Context, ADE20K and COCO Stuff.