PFMNet: Few-Shot Segmentation with Query Feature Enhancement and Multi-Scale Feature Matching

The datasets in the latest semantic segmentation model often need to be manually labeled for each pixel, which is time-consuming and requires much effort. General models are unable make better predictions, new categories of information that have never been seen before, than few-shot has emerged. However, still faced up with two challenges. One inadequate exploration conveyed high-level features, other inconsistency segmenting objects at different scales. To solve these problems, we proposed a prior feature matching network (PFMNet). It includes novel modules: (1) Query Feature Enhancement Module (QFEM), makes full use support set enhance query feature, (2) multi-scale module (MSFMM), increases probability multi-scales objects. Our method achieves an intersection over union average score 61.3% one-shot 63.4% five-shot segmentation, surpasses state-of-the-art results by 0.5% 1.5%, respectively.