Zebrafish identification with deep CNN and ViT architectures using a rolling training window

Zebrafish are widely used in vertebrate studies, yet minimally invasive individual tracking and identification in the lab setting remain challenging due to complex and time-variable conditions. Advancements in machine learning, particularly neural networks, offer new possibilities for developing simple and robust identification protocols that adapt to changing conditions. We demonstrate a rolling window training technique suitable for use with open-source convolutional neural networks (CNN) and vision transformers (ViT) that shows promise in robustly identifying individual maturing zebrafish in groups over several weeks. The technique provides a high-fidelity method for monitoring the temporally evolving zebrafish classes, potentially significantly reducing the need for new training images in both CNN and ViT architectures. To understand the success of the CNN classifier and inform future real-time identification of zebrafish, we analyzed the impact of shape, pattern, and color by modifying the images of the training set and compared the test results with other prevalent machine learning models.