Semi-Supervised Semantic Image Segmentation

Supervised deep learning methods have made remarkable strides in various image-based tasks, such as classification, object detection, and semantic segmentation, leveraging large annotated datasets. However, in domains like biomedical imaging, where annotated labels are scarce, these networks often underperform. Semi-supervised learning methods have shown promise in bridging this gap, particularly in image classification, by leveraging regularization techniques to exploit smoothness and cluster assumptions. Despite this progress, semi-supervised semantic segmentation, which requires more laborious data annotation, has not yet achieved comparable success.

Problem:

While semi-supervised learning has advanced in image classification tasks, it has not translated as effectively to semantic segmentation, particularly in scenarios with limited annotated data. The laborious nature of data annotation for semantic segmentation poses a significant challenge, hindering the performance of semi-supervised approaches in this domain. Bridging this gap is crucial for advancing the application of deep learning in biomedical imaging, where obtaining labeled data is exceptionally challenging.

Research Aim: 

The primary objective of this research is to push the boundaries of semi-supervised semantic segmentation, aiming to approach the performance of supervised settings. Building upon recent efforts, the study seeks to explore novel regularization techniques and augmentation strategies to enhance the effectiveness of semi-supervised learning in semantic segmentation tasks. While the challenges and approaches explored are applicable to generic image segmentation tasks, the focus of the applications will be on biomedical imaging for global health, where labeled data scarcity is prevalent.

Outcome So Far:

The research has begun by examining existing semi-supervised learning approaches and identifying areas for improvement specific to semantic segmentation tasks. Initial investigations into novel regularization techniques and augmentation strategies show promising potential for enhancing the performance of semi-supervised semantic segmentation. Ongoing efforts focus on refining these techniques and validating their effectiveness through experimentation, with the ultimate aim of advancing the application of deep learning in biomedical imaging for global health.

Team: Bishesh Khanal and Pratima Upretee

Relevant Publications

Manuscript in preparation.