Presentation Title

Corrosion products classification using self-supervised learning

Presenter Information

vidya bommanapallyFollow

Advisor Information

Mahadevan Subramaniam

Location

MBSC Gallery Room 308 - G

Presentation Type

Oral Presentation

Start Date

4-3-2022 2:00 PM

End Date

4-3-2022 3:15 PM

Abstract

Self-supervised learning has gained popularity in scenarios where large unlabeled datasets with very few annotations are available. Images from fields of medical, engineering rarely have annotations available due to limited expert availability. To preserve the domain information in such situations, self-supervised learning is being studied over transfer learning techniques. In this project, self-supervised learning is used to study biofilm microscopy images to automatically identify if the images contain corrosion products on the metal surface. In self-supervised learning, the model is trained using unlabelled images in a supervised way by providing pseudo-labels for the images. The approach employs augmentation techniques such as rotation and flip to learn representations from the available images. The representations learned by the neural networks are further fine-tuned for classification using few expert annotated images.

This document is currently not available here.

COinS
 
Mar 4th, 2:00 PM Mar 4th, 3:15 PM

Corrosion products classification using self-supervised learning

MBSC Gallery Room 308 - G

Self-supervised learning has gained popularity in scenarios where large unlabeled datasets with very few annotations are available. Images from fields of medical, engineering rarely have annotations available due to limited expert availability. To preserve the domain information in such situations, self-supervised learning is being studied over transfer learning techniques. In this project, self-supervised learning is used to study biofilm microscopy images to automatically identify if the images contain corrosion products on the metal surface. In self-supervised learning, the model is trained using unlabelled images in a supervised way by providing pseudo-labels for the images. The approach employs augmentation techniques such as rotation and flip to learn representations from the available images. The representations learned by the neural networks are further fine-tuned for classification using few expert annotated images.