FFNet: Feature Fusion Network for Few-shot Semantic Segmentation

Ya-Nan Wang, Xiangtao Tian, Guoqiang Zhong

Abstract: Semantic segmentation aims at assigning a category label to each pixel in an image. Deep neural networks have achieved many breakthrough research achievements on this task. Nevertheless, there exist two critical bottleneck problems to be solved. First, deep neural networks usually need to be trained on large-scale labeled datasets, which are expensive to obtain or label. Second, traditional semantic segmentation methods are difficult to predict unseen classes after training. To address these problems, few-shot semantic segmentation is proposed, and recent methods have achieved impressive performance. However, many of the existing approaches ignore the semantic correlation between data and fail to generate discriminative features for the semantic segmentation. In this paper, to address the above issue, we propose a feature fusion network (FFNet) for few-shot semantic segmentation to enhance the discriminative ability of the learned data representations. Specifically, a task attention module is devised to learn the semantic correlation between data. Then, a multi-scale feature fusion module is trained to adaptively fuse the contextual information at multiple scale, thus capturing multi-scale object information. To the end, the proposed FFNet experiments conducted on the PASCAL- $$5^i$$ 5 i and COCO- $$20^i$$ 20 i datasets demonstrate the superiority of our proposed FFNet and show its advantage over existing approaches.