Hierarchical Multitask Learning for Improved Underwater Recognition on Imbalanced Tasks

Deep learning methods for automatic visual inspection can help detect potential anomalies in real-time operations and reduce the burden of full-length video inspection analysis. Such learning-based methods require, however, huge amounts of diverse labeled examples to be trained on. The lack of variance in examples from less frequent entities (eg. anomalies) makes it difficult for deep models to generalize such tasks, leading to poor performance. We then propose to use multitask learning to improve the generalization capacity, thus the performance of deep classification models at the recognition of less frequent entities in the underwater domain. In particular, we exploit hierarchical relationships between classification tasks to wisely choose which tasks to combine. For the imbalanced task label, our approach is shown to improve multitask performance from 0.74 to 0.89, in terms of average precision. Efficiency is also improved, with inference time being reduced by almost fifty percent, thus making it significantly cheaper to use visual inspection models in realtime operations.