Few-Shot Image Segmentation Based on Dual Comparison Module and Sequential k-Shot Integration

Abstract

Few-shot image segmentation intends to segment query images (test images) given only a few support samples with annotations. However, previous works ignore the impact of the object scales, especially in the support images. Meanwhile, current models only work on images with the similar size of the object and rarely test on other domains. This paper proposes a new few-shot segmentation model named DCNet, which fully exploits the support set images and their annotations and is able to generalize to the test images with unseen objects of various scales. The idea is to gradually compare the features from the query and the support image, and refine the features for the query. Furthermore, a sequential k-shot comparison method is proposed based on the ConvGRU to integrate features from multiple annotated support images. Experiments on Pascal VOC dataset and X-ray Security Images demonstrate the excellent generalization performance of our model.

Keywords

• Few-shot learning
• Image segmentation
• Dual comparison module
• Convolutional-gated recurrent unit