Semi-supervised Defect Segmentation with Uncertainty-aware Pseudo-labels from Multi-branch Network

Abstract

Semi-supervised learning methods have recently gained considerable attention for training deep learning networks with limited labeled samples and additional large label-free samples. Consistency regularization and pseudo-labeling methods are among the most widely used semi-supervised learning methods. However, unreliable pseudo labels will largely limit the model’s performance when learning from unlabeled images. To alleviate this problem, we propose uncertainty-rectified pseudo labels generated from dynamically mixing predictions of multiple decoders with a shared encoder network for semi-supervised defect segmentation. We estimated the uncertainty as the prediction discrepancy between the average prediction and the output of each decoder head. The estimated uncertainty then guides the consistency training as well as the pseudo-label-based supervision. The proposed method achieved significant performance improvement over the fully supervised baseline and other state-of-the-art semi-supervised segmentation methods on similar labeled data proportions. We also performed an extensive ablation study to demonstrate that the proposed method performs well under various setups.