A deep learning algorithm and software for photo identification of the Indo-Pacific humpback dolphin (Sousa chinensis)

Deep neural networks have been increasingly used to identify individual animals in ecological studies by learning and distinguishing their naturally occurring marks or features. Traditional individual animal recognition requires prior knowledge and experience, which can be time-consuming and inefficient. In this paper, a distinctive deep learning framework that automatically reidentifies individual Indo-Pacific humpback dolphins (Sousa chinensis) from photos was proposed. For most dolphin species with a dorsal fin, this feature is reliably used to identify and distinguish individuals in studies that require distinction between members of a group or population. Feature cutting and background removing strategies were added to allow a focus on local information. Knowledge distillation was also applied to improve the robustness of the framework. Additionally, an automatic dolphin recognition software suite for cetologists that may reduce the amount of effort and time required to manually confirm individual dolphin ID from photographs had been developed. In the end, the effectiveness of applying this deep neural network approach for individual Indo-Pacific humpback dolphin recognition had been demonstrated.