Channel Reorganization for Few-Shot Segmentation

Abstract

Few-shot segmentation methods are often modeled as two branch convolutional neural networks, namely support branch and query branch. Existing methods often rely too much on support images. They ignore the power of the query image and fail to fully learn the information of the query image. In addition, we all know that convolution extracts information features by fusing spatial and channel features in local receptive fields. However, most of the existing methods extract information by fusing spatial features, ignoring the role of channel features in information extraction. To address the issues, we propose a new semantic segmentation module based on channel reorganization graph convolution network (CRGCN). First, we construct the graph structure according to the channel features, then screen the beneficial structures based on motif, and finally use GCN to recombine the channel features. This can sufficiently mine the potential relationship between the features of the query image. Experiments on PASCAL − 5ⁱ and Fss-1000 datasets show that our proposed method is superior to the baseline and state-of-the-art method.