Improving automated underwater ship hull inspection through incremental learning & uncertainty quantification in deep learning models

Abstract

Rapid technological progress has recently eliminated the necessity of physically deploying humans for manual underwater ship inspections. Instead, Remotely Operated Vehicles (ROVs) can now capture and record videos, which humans can subsequently review. By leveraging annotated data, it becomes possible to train deep learning models to assist in data exploration. Nevertheless, the process of manually inspecting recorded videos remains labor-intensive and time-consuming.

This thesis investigates the potential of training a model on specific types of images to enhance learning efficiency, so the user has to annotate fewer images. Specifically, we investigate the two hypotheses. The first one is that a model's confidence value can be evaluated to give the user indications on which images should be annotated to yield the best performance. The second hypothesis is that a model doesn't have to be trained from scratch every time new training data is added, but only for a fraction of the original epochs.

To enhance accuracy, I plan to utilize a pre-trained deep learning model to identify uncertain images. These images will be reviewed by a human, and the model will be updated using incremental learning. I propose two distinct approaches for determining uncertainty. Separate deep learning models will use each approach to determine the images that will be used for training. I will compare these models to each other and to models trained on images that are classified as not uncertain.

By training on both the new and the old training data, I get results that support my second hypothesis. The findings from the experiments for the second hypothesis were inconclusive.