Co-Seg: An Image Segmentation Framework Against Label Corruption
This addresses the issue of label corruption in image segmentation for researchers and practitioners, offering an incremental improvement by enhancing robustness in noisy datasets.
The paper tackles the problem of neural networks overfitting to corrupted labels in image segmentation by proposing Co-Seg, a framework that collaboratively trains networks to filter and correct noisy labels, achieving results comparable to supervised learning with noise-free labels in experiments.
Supervised deep learning performance is heavily tied to the availability of high-quality labels for training. Neural networks can gradually overfit corrupted labels if directly trained on noisy datasets, leading to severe performance degradation at test time. In this paper, we propose a novel deep learning framework, namely Co-Seg, to collaboratively train segmentation networks on datasets which include low-quality noisy labels. Our approach first trains two networks simultaneously to sift through all samples and obtain a subset with reliable labels. Then, an efficient yet easily-implemented label correction strategy is applied to enrich the reliable subset. Finally, using the updated dataset, we retrain the segmentation network to finalize its parameters. Experiments in two noisy labels scenarios demonstrate that our proposed model can achieve results comparable to those obtained from supervised learning trained on the noise-free labels. In addition, our framework can be easily implemented in any segmentation algorithm to increase its robustness to noisy labels.