Using deep-learning for automatic identification of images of marine benthic macro-invertebrate bycatch: a proof of concept

We applied a deep-learning approach in order to develop a neural network able to detect and identify macro-invertebrate organisms within images of benthos bycatch collected in the Southern Ocean. We used the Faster RCNN architecture and fine-tuning approach. To perform the transfer-learning, we used an annotated dataset of 59,756 images of organisms identified within 1,845 images of lots, covering eleven taxa: Echinodermata, Asteroidea, Arthropoda, Annelida, Chordata, Hemichordata, Cnidaria, Porifera, Bryozoa, Brachiopoda and Mollusca. The resulting network, not yet efficient enough to obtain precise identifications, is able to provide detection and classification of organisms with a good level of accuracy considering the limited quality of the images used for training. We present this study as a proof of concept for teams involved in the management of collections of macro-invertebrate images.