EVALUATION OF SELF-SUPERVISED LEARNING APPROACHES FOR SEMANTIC SEGMENTATION OF INDUSTRIAL BURNER FLAMES

Abstract

In recent years, self-supervised learning has made tremendous progress in closing the gap to supervised learning due to the rapid development of more sophisticated approaches like SimCLR, MoCo, and SwAV. However, these achievements are primarily evaluated on common benchmark datasets. In this paper, we focus on evaluating self-supervised learning for semantic segmentation of industrial burner flames. Our goal is to build an intuition on how self-supervision performs in a scenario relevant for industrial application where training labels and the opportunities for hyperparameter tuning are limited. We demonstrate that self-supervised pre-training can constitute an alternative to the state-of-the-art approach of pre-training on ImageNet. Across all scenarios, the self-supervised approaches are less susceptible to sub-optimal learning rates and achieve higher mean accuracies than ImageNet pre-training, especially when training labels are scarce.