EFNet: Enhancement-Fusion Network for Semantic Segmentation

Semantic segmentation is a challenging and important task in computer vision. Convolutional neural networks (CNNs) have demonstrated their outstanding performances on such dense classification tasks. Most recent segmentation networks mainly focus on feature extraction for one single input image, while paying little attention to facilitating the segmentation by image manipulation or enhancement. In this paper, we design an enhancement-fusion network (EFNet), which aims at enhancing an input image for more diversified features to boost the following task of pixel-wise labeling. Specifically, the enhancement modules are trained to produce multiple enhanced images. Afterwards, the fusion module selectively attends on such images and fuses them to yield one new image. The proposed EFNet can be directly and flexibly integrated as an auxiliary network with state-of-the-art semantic segmentation networks, while maintaining the end-to-end training manner. Extensive results on benchmark datasets corroborate that the combination of the EFNet and the CNN-based semantic segmentation networks significantly improves the segmentation performance compared with the original segmentation networks. (c) 2021 Elsevier Ltd. All rights reserved.

Automated summary

Semantic segmentation, as a challenging task in computer vision, has been effectively improved by the design of an enhancement-fusion network (EFNet). EFNet enhances input images to provide diversified features for pixel-wise labeling. Experimental results demonstrate that the combination of EFNet and CNN-based segmentation networks significantly enhances segmentation performance.