Video surveys are commonly used to monitor the abundance and distribution of managed species to support management. However, considerable effort, time, and cost are required for human review and automated fish species recognition provides an effective solution to remove the bottleneck of post-processing. Implementing fish species detection techniques for underwater imagery is a challenging task. In this work, we present the Multiple Instance Active-learning for Fish-species Recognition (MI-AFR), which is formulated as an object detection-based approach to perform localization and classification of fish species. It can select the most informative fish images from unlabeled sets by estimating the uncertainty of unlabeled images by using adversarial classifiers trained on labeled sets. Moreover, we have analyzed the improved performance of MI-AFR by considering different backbone networks as a trade-off between speed and accuracy. For experiments, we have used the fine-grained and large-scale reef fish dataset obtained from the Gulf of Mexico – the Southeast Area Monitoring and Assessment Program Dataset 2021 (SEAMAPD21). The experimental results illustrate that the superiority of the proposed method can establish a solid foundation for active learning in fish species recognition, especially with a small number of labeled sets.
|