A Hybrid Semantic Segmentation Based on Level-Set Evolution Driven by Fully Convolutional Networks

Abstract

Semantic segmentation network is one successful deep architecture for understanding key foregrounds. The fully convolutional networks (FCNs) can learn the regional masks by composing layered representations coarse-to-fine. Nevertheless, these increasing feature maps often fail to improve significantly the discrimination to segmentation details, due to the lack of accurate spatial representations and prior on salient foreground boundaries. This paper presents a hybrid semantic segmentation network based on a differentiable level-set evolving on multiple feature channels, which can be driven by the deep architectures of FCNs. To the level-set iterations, the re-initialization and Gaussian smoothing operations have been integrated to achieve an uniform training procedure. Also, this proposed wrapping level-set layer can be easily generalized to other existing deep segmentation architectures. Experimental results on various open datasets verify that, compared with related baselines, this hybrid framework significantly improves the spatial details and accuracy rates of the segmenting foreground boundaries.