Transfer Learning of Image Classification Networks in Application to Dolphin Whistle Detection

Dolphin whistle detection is an important and multipurpose but time-consuming task. The ability to automate and streamline this process can be invaluable for future research in marine studies and other fields that aim to utilise these signals. When dealing with underwater acoustics, a large obstacle to overcome is the abundance of noise and interfering sounds, natural and anthropogenic alike. In this paper, we apply successful image classification networks to two separate datasets containing dolphin whistles with the goal of determining an effective method to conduct automated detection with minimal interference from a manual operator regardless of environment. We further investigate the impacts of shrinking the dataset size and performing parameter freezing on the networks at hand. Networks are assessed by their detection accuracy and achieve performances comparable to those in existing works, the best being 96.7%, thus proving the effectiveness of these pre-trained image classification models.