Noise-robust consistency regularization for semi-supervised semantic segmentation

doi:10.1016/j.neunet.2024.107041

Noise-robust consistency regularization for semi-supervised semantic segmentation

2025-04-01

0

PRE

AI

DOI：10.1016/j.neunet.2024.107041

Original

Original request for help

Save

Abstract

En 中文

The essential of semi-supervised semantic segmentation (SSSS) is to learn more helpful information from unlabeled data, which can be achieved by assigning adequate quality pseudo-labels or managing noisy pseudo-labels during training. However, most relevant state-of-the-art (SOTA) methods are mainly devoted to improving one aspect. By revisiting the representative SSSS methods from a robust learning view, this paper discovers that the appropriate combination of multiple noise-robust methods contributes both to assigning sufficient quality pseudo labels and managing noisy labels. Therefore, from five different noise management perspectives, we summarize the reasons why noise-robust techniques can successfully harvest performance gains in SSSS. Subsequently, we present a novel feature perturbation method, multi-view learning strategy, and robust loss function to exploit the advantages of different noise-robust techniques. The outcome of this paper is anew SSSS approach with noise-robust consistency regularization called NRCR that can simultaneously produce adequate quality pseudo-labels and manage noisy pseudo-labels. Abundant experiments on public benchmarks demonstrate the performance superiority of our approach compared with previous SOTA methods and the correctness of our analytical viewpoints. Code is available at https://github.com/zhanghk1996/NRCR.

Keywords:

Semi-supervised semantic segmentation

Robust learning

Consistency regularization

Multi-view learning

Feature perturbation

Noise-robust consistency regularization for semi-supervised semantic segmentation

Abstract

Journal

Organization