Introduction
Reliable identification of individual fish is a fundamental requirement for aquaculture research and production, particularly when individual-level monitoring is needed for biomass estimation, health assessment, selective breeding, and growth tracking. However, in many aquaculture settings, individual fish are still monitored through group-based observations because conventional identification approaches often rely on physical tagging. These methods are labour-intensive, time-consuming, impractical for large-scale systems, and in some cases infeasible because handling or tagging may interfere with experimental conditions, disrupt normal production, or negatively affect fish welfare.
Although several non-invasive image-based identification approaches have been proposed, many of them have mainly been evaluated under controlled or short-term experimental conditions and are not yet sufficiently validated under realistic longitudinal aquaculture scenarios [1]. Moreover, many existing approaches depend on having a relatively large number of images per individual, which is often unrealistic in practical aquaculture experiments. In real monitoring conditions, only a limited number of images may be available for each fish, while appearance can also change over time due to growth, health status, pose variation, and differences in imaging conditions. This limited-image-per-individual setting therefore remains a major challenge for reliable fish identification.
Recent advances in computer vision and deep learning provide a promising basis for overcoming these limitations by enabling image-based identification of individual fish without physical contact. In this context, methods that can maintain high identification accuracy when only a small number of images per fish are available are particularly important. Image-based individual identification therefore offers a non-invasive alternative to conventional tagging methods, enabling remote, rapid, and cost-effective in situ monitoring, while also requiring careful consideration of ethical issues such as animal welfare.
Materials and Methods
In this study, a prototype-based one-shot individual identification framework was developed for early juvenile Oncorhynchus mykiss. A ConvNeXt backbone network, pre-trained on a longitudinal image dataset of rainbow trout, was used as the main feature extractor and enhanced with a multi-stage feature fusion strategy to obtain discriminative visual representations from fish images.
In this framework, the pre-trained model encoded each fish image into a latent embedding space, where individual-specific visual patterns were represented as compact identity descriptors. Prototype representations were generated for each individual using a single reference image, consistent with the one-shot identification setting. To improve prototype robustness, the reference image was further expanded using synthetic augmentation. During inference, query images collected after a temporal gap at subsequent sampling sessions were compared with the stored prototypes using cosine similarity. Identity was assigned according to the prototype with the highest cosine similarity score.
The proposed method was evaluated using a long-term red mark syndrome (RMS) monitoring dataset consisting of 138 individual rainbow trout. For each individual fish, single images of the left and right lateral sides were photographed under controlled imaging conditions in a photo tent. This dataset was collected over 99 days, which includes five sampling events of photos throughout the duration of the experiment. During inference, new observations from later sessions were matched against prototypes generated from previously observed samples. Identification was evaluated by comparing the predicted identity of each query image with its known ground-truth identity. In addition, predictions from the left and right lateral sides were assessed jointly. A final individual-level prediction was considered correct only when both lateral views were assigned to the same identity and this identity matched the ground-truth label. This evaluation design enabled assessment of the framework's ability to maintain individual recognition under longitudinal monitoring conditions with temporal variation in fish appearance.
Results and Discussion
The proposed one-shot identification framework was evaluated using Rank-1 accuracy across consecutive photo-sampling sessions. For each query image collected in a later session, identity prediction was performed by comparing its embedding with prototype representations generated from the previous session. Left and right lateral images of each fish were first evaluated separately and then jointly assessed to confirm individual identity. A prediction was considered correct only when the identities retrieved from both lateral views were consistent and matched the known ground-truth identity. Ground-truth labels were verified using recorded individual information and expert visual assessment. Between all sequential sampling sessions, this method achieved an average 97.71% Rank-1 accuracy. This confirms reliable identification performance despite changes in fish appearance, including growth-related morphological variation and lesion-associated changes caused by red mark syndrome [2]. The outcome results indicate that the proposed one-shot framework can reliably identify individual fish despite changes in appearance. These findings indicate its potential use in aquaculture research and production, particularly for health monitoring, growth tracking, and welfare-oriented management.
References
[1] Ziaei, M. M., Urban, J., & Cisar, P. (2025) Image-Based Individual Fish Identification as a Substitute for Invasive Methods: State-of-the-Art and Research Gaps. Reviews in Aquaculture. https://doi.org/10.1111/raq.70078
[2] Ziaei, M. M., Orioles, M., Schmidt, J. G., Koliada, I., Urban, J., & Cisar, P. (2026). Can non-invasive methods be used for early detection of fish skin pathology? A red mark syndrome study. Aquacultural Engineering, 114, 102708. https://doi.org/10.1016/j.aquaeng.2026.102708