Deep learning-based image semantic segmentation techniques have made great strides in recent years. However, they still need large amounts of finely annotated image data, and generalizing the model from known classes to unknown ones remains a challenge. Most of the work on few-shot semantic segmentation techniques deals with the support set by directly utilizing images and masks for feature fusion. That tends to make the model not pay enough attention to the less-sample category, leading to missed detections. To alleviate this problem, in this paper, we propose the Region Select Enhancement Network, a novel structural model composed of base and meta learner, based on the perspective of metric learning and data enhancement. We employ an additional base learner to individually recognize targets within the base class, utilizing the recognition results of the base class as background-guided features for the final target. We then effectively fuse the outputs of the base learner and meta learner to produce accurate target images. Notably, unlike the common meta learner, we add a separate target category selection enhancement branch to the meta learner, augmenting the target features with known information from the support set. This further reduces background interference, thereby improving the model’s generalization ability. We conducted experiments on Cityscapes- 3ⁱ, a few-shot outdoor dataset constructed from labeled images in the Cityscapes dataset, to validate the effectiveness of our method.
Wang et al. (Mon,) studied this question.