Robust detection of marine life with label-free image feature learning and probability calibration

Advances in in situ marine life imaging have significantly increased the size and quality of available datasets, but automatic image analysis has not kept pace. Machine learning has shown promise for image processing, but its effectiveness is limited by several open challenges: the requirement for large expert-labeled training datasets, disagreement among experts, under-representation of various species and unreliable or overconfident predictions. To overcome these obstacles for automated underwater imaging, we combine and test recent developments in deep classifier networks and self-supervised feature learning. We use unlabeled images for pretraining deep neural networks to extract task-relevant image features, allowing learning algorithms to cope with scarcity in expert labels, and carefully evaluate performance in subsequent label-based tasks. Performance on rare classes is improved by applying data rebalancing together with a Bayesian correction to avoid biasing inferred in situ class frequencies. A divergence-based loss allows training on multiple, conflicting labels for the same image, leading to better estimates of uncertainty which we quantify with a novel accuracy measure. Together, these techniques can reduce the required label counts ∼100-fold while maintaining the accuracy of standard supervised training, shorten training time, cope with expert disagreement and reduce overconfidence.