Abstract: Traditional semantic segmentation tasks usually need a lot of labeled data for training,and it is difficult to generalize to new categories.Few-shot segmentation aims to segment new categories of target objects from query images using a small number of annotated supporting images.Due to the limited supporting image data,how to extract representative guidance information from limited support images is an important challenge for few-shot segmentation task.In order to solve this problem,a few-shot segmentation method based on multi-scale prototype hierarchical matching is proposed in this paper.Firstly,the middle-level and high-level features of the query image and the support image are obtained through the residual network ResNet.In order to further extract the rich context feature information of the target object,the extracted middle-level features are fed into the pyramid pooling module for multi-scale feature extraction.Based on the idea of prototype learning,middle-level features and high-level features are layered to generate prototypes and matched to obtain the final predicted segmentation mask.Experiments are carried out on the PASCAL-5ⁱ dataset and experimental results show that the mIoU of the proposed method achieves 66.7% in 1-way 5-shot setting,which is 11% and 4.8% higher than the current mainstream PANet and PFENet models,respectively,demonstrating the effectiveness and advanced nature of the method.

Key words: Few-shot segmentation, Muti-scale, Semantic segmentation, Prototype learning, ResNet