DISCUSSION
Here we present the utility of a new software package, SealNet, a complete, automated pipeline to non-invasively identify individual seals in photographic images. We describe a novel face detector GUI trained to detect harbor seal faces and the development of a new neural network to recognize individual seal faces. Following validation of the technique, we use SealNet in a preliminary study to explore site-fidelity of harbor seals in the Casco Bay, Maine region of the northwestern Atlantic coast. Our initial validation analyses confirm the efficiency and accuracy of our facial recognition technology in the photo identification of an economically important coastal marine mammal.
Our trained face detector had a precision of 85%, and a recall of 87% despite having very little restrictions on the position of the seals within the photo with the only limitation that both eyes of the seals are visible. This feature enables the successful use of SealNet software to conduct studies on animals in the wild without having the animals looking directly at the camera. The precision and recall of our detector could be increased by restricting the possible angle and pose of the seal, but this would limit the number of photos that meet such requirements in field studies.
In a direct performance comparison of the recognition task, SealNet performs better than PrimNet on average at all ranks with up to 18% improved classification accuracy at rank-1 for closed-set and 9% for open-set identification with SealNet. It is also important to note that our model performs consistently well as our database increases in size. The consistent performance of our model demonstrates that SealNet generalizes well (i.e., overfitting is not an issue).
For our recognition software, we have achieved a remarkable accuracy in both close-set (rank-1: 88% and rank-5: 96%) and open-set (rank-1 and rank-5: 93%), but there is still room for improvement. FaceNet which was trained on more than 3 million images of almost 10,000 unique human individuals, achieved an accuracy of almost 100%. Other biometric models for chimpanzees and lemurs also yield an accuracy around 92%–94%. Therefore, with a larger dataset with more photos per seal individual, it is possible that we can further improve our accuracy.