SSFCN-DCRF: Semantic Segmentation for Biomedical Image

Abstract

In this paper, we present an end-to-end solution for the task of semantic biomedical image segmentation. We use skip stride Fully Convolutional Networks to output segmentation map. To further improve the performance of the semantic segmentation, we use the dense conditional random field (DCRF) layer to fine tune the segmentation map. We also apply separable operation to the Fully Convolutional Networks to get a lightweight network. We evaluate our network on Kaggle Lung CT Dataset and UCSB Bio-Segmentation Benchmark. The results show that our network achieves state of art performance on semantic segmentation for biomedical Image.