Comparison of Convolutional Neural Network Architectures for Sea Turtle Individual's Recognition

Identification of individual turtles for population studies usually uses fin tags or other physical markers. Since this method is not very practical, we are studying different convolutional neural networks (CNN) to identify individual images of turtles by using transfer learning networks AlexNet, GoogleNet, VGG-19 Net and ResNet50. Sea turtles, unlike other animals, have unique facial patterns, making them excellent prospects for feature recognition. This study examined 1426 images of right facial scutes from 20 classes of turtles. Experiments with high-quality, low-quality and a mixture of both image qualities were conducted to test the performance of different CNNs. The ResNet50 technique achieved 95.3% accuracy with a mixed dataset, and AlexNet obtained the highest accuracy with high-quality image dataset (97.74%). The VGG19 network, on the other hand, performed well for dataset containing low-quality images with 91.82% accuracy. Experiments have shown different results for each data set. However, considering the uncertainty in a real underwater environment where the captured image quality is sometimes high and sometimes low, the ResNet50 network addresses this task-related problem as it has achieved the highest accuracy for the mixed dataset.