Deep co-training for semi-supervised image segmentation

• Published in: Pattern recognition Vol. 107; p. 107269
• Main Authors: Peng, Jizong, Estrada, Guillermo, Pedersoli, Marco, Desrosiers, Christian
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.11.2020
• Subjects: Co-training; Deep learning; Ensemble learning; Image segmentation; Semi-supervised learning; Deep learning; Co-training; Image segmentation; Ensemble learning; Semi-supervised learning
• ISSN: 0031-3203; 1873-5142
• DOI: 10.1016/j.patcog.2020.107269

•Semi-supervised semantic segmentation based on an ensemble of deep learning models.•All models are trained jointly in a co-training setting.•Coherence among models is enforced by minimizing the Jensen-Shannon divergence of the probabilities distributions on the unlabeled samples.•Diversity among models is preserved by enforcing similarity between the prediction of a model on an unlabeled sample and the adversarial prediction of another model on the same sample.•Results on ACDC dataset and SCGM dataset show the capability of our model to outperform previous ensemble approaches by a significant margin. In this paper, we aim to improve the performance of semantic image segmentation in a semi-supervised setting where training is performed with a reduced set of annotated images and additional non-annotated images. We present a method based on an ensemble of deep segmentation models. Models are trained on subsets of the annotated data and use non-annotated images to exchange information with each other, similar to co-training. Diversity across models is enforced with the use of adversarial samples. We demonstrate the potential of our method on three challenging image segmentation problems, and illustrate its ability to share information between simultaneously trained models, while preserving their diversity. Results indicate clear advantages in terms of performance compared to recently proposed semi-supervised methods for segmentation.