Collaborative Image Annotation for Deep Learning

Reducing the need for large initial training data sets by including a human domain expert in the loop when training classification models from annotated data.

 

Availability of annotated (labelled) images is crucial to train a Deep Learning (DL) model that can classify and predict objects and events under water correctly. Existing approaches to developing a DL classifier model often require large amounts of training data which is a resource consuming task. Recently, technical approaches have been proposed to address and reduce the need for large initial training data sets by interactively incorporating a human domain expert in the loop when training classification models from annotated data. These approaches (Figure 1) allow a human expert to manually correct or revise predictions with low confidence scores, and the initial model is then trained with the revised data.

Figure 1: Bauer et al., eXplainable Cooperative Machine Learning with NOVA, KI - Künstliche Intelligenz, January 2020
Figure 1: Bauer et al., eXplainable Cooperative Machine Learning with NOVA, KI - Künstliche Intelligenz, January 2020

Research Topic

This MSc thesis aims at developing a human-in-the-loop approach for collaborative annotation and semantic segmentation of subsea images.

Methods

  • Identify Deep Learning and Cooperative Machine Learning approaches and algorithms for semi-automatic labelling of objects and events (Figure 2).
  • Develop a framework that will address and reduce the need for large initial training data sets by interactively incorporating a human domain expert in the loop when training classification models from annotated data.
  • Train an initial model that will serve as a basis for further development.
  • Implement the framework in the current software prototype (Seekuence).
  • Evaluate the approach by using a relevant use case pilot in a current SINTEF innovation project (LIACi).


Figure 2: Islam et al., Semantic segmentation of underwater imagery: Dataset and benchmark. arXiv preprint arXiv:2004.01241 (2020)

Expected Results and Learning Outcome

  • Approach for human-in-the-loop annotation of images in a subsea context.
  • Software prototype.
  • Evaluation of prototype with real business use cases.
  • Publication

Recommended prerequisites:

Basic knowledge of Machine Learning and Computer Vision techniques will be considered an advantage.

Publisert 7. mai 2021 00:44 - Sist endret 7. mai 2021 00:44

Omfang (studiepoeng)

60