Emerging semi-supervised learning methods have enabled great progress in segmentation tasks. However, popular semi-supervised segmentation models use constraints that are not strict. In this paper, we propose a new method, multi-scale cross pseudo-supervision, that introduces higher constraints by multi-scale information to improve the quality of pseudo-labels. Specifically, we extend the backbone segmentation network by adding a multi-scale feature pyramid at the decoder to extract multi-scale information. In addition, to further enhance the consistency on multiple scales, we perform perturbation operations on the original input image. Experiments show that our method achieves excellent segmentation performance on both sonar and ISIC2016 datasets. The performance gain benefits from two techniques—multi-scale constraints and perturbation consistency. And the proposed method alleviates the annotation pressure for image segmentation in real-world human-centric applications.

新兴的半监督学习方法使分割任务取得了巨大进步。然而，流行的半监督分割模型使用的约束并不严格。在本文中，我们提出了一种新方法，多尺度交叉伪监督，通过多尺度信息引入更高的约束来提高伪标签的质量。具体来说，我们通过在解码器处添加多尺度特征金字塔来提取多尺度信息来扩展主干分割网络。此外，为了进一步增强多个尺度上的一致性，我们对原始输入图像进行扰动操作。实验表明，我们的方法在声纳和 ISIC2016 数据集上均实现了出色的分割性能。性能增益受益于两种技术——多尺度约束和扰动一致性。所提出的方法减轻了现实世界中以人为中心的应用中图像分割的注释压力。