Active learning in deep convolutional neural networks for image segmentation

Abstract

The sitting position and seat belt orientation of passengers in automobiles can be crucial in the event of a collision. In order to warn passengers of unsafe positions, deep learning models in the form of neural networks can be used to identify the seat belt from image data. Performance of neural networks can be increased by improving the model (model-centric approaches) or by improving the data used to train the model (data-centric approaches). In this thesis we compare the segmentation performance gains from model-centric approaches to data-centric approaches including stratified sampling, data balancing, label error reduction and active learning. No new model architecture was found that improved performance, but the model training time was sped up by four times without performance loss. Stratified sampling, data balancing and label error reduction did not improve performance significantly. In active learning, images to be labeled were selected according to the model’s uncertainty. Several uncertainty metrics were used, all leading to an improvement when using active learning. The best result showed that we achieved 95% and 99% of the best baseline performance using 19% and 23% less data respectively.