Learning what healthy fish look like
Self-supervised and annotation-efficient methods locate fish anatomy and identify unusual appearance, wounds, deformation, fin or eye damage, and behavioural changes.
Explore WP1 →Research Council of Norway · Project 344022
cAIge
Computer Vision and Artificial Intelligence based Salmon Identification and automated long-term welfare assessment in aquaculture net-pens
cAIge develops computer vision and artificial intelligence for non-invasive identification of individual salmon and automated, long-term welfare assessment in aquaculture net-pens.
The challenge
Monitoring welfare at the individual-fish level
Welfare monitoring in commercial aquaculture is difficult: large populations, changing underwater conditions, occlusion, and natural variation all limit what can be observed. cAIge investigates technology that can follow welfare indicators objectively, at individual and group level, over time. The project also explores self-learning approaches—including self-supervised and unsupervised learning—that can discover useful visual patterns in largely unlabelled underwater data.
A non-invasive fish health journal
The project combines visual biometric re-identification with automatic detection and scoring of welfare indicators. Repeated observations can then be connected with environmental measurements and farming operations to improve understanding of the conditions that support healthy fish.
This is enabling research: the goal is to develop and validate methods that can eventually support better decisions, earlier detection of problems, and more sustainable aquaculture operations.
Research areas
From underwater video to long-term welfare insight
The research moves from extracting information in underwater video to validating its biological meaning and usefulness in full-scale salmon farming.
Recognising individuals without tags
Visual biometrics—including natural melanin patterns, body regions, and learned representations—are studied for robust recognition across viewpoints, cameras, and time.
Explore WP2 →Connecting observations and conditions
Individual and population-level welfare observations are related to environmental data and production operations to reveal useful correlations and contributors to healthy fish.
Explore WP3 →Selected WP3 publication ↗
Testing in realistic environments
Methods are evaluated against biological expertise and tested with aquaculture data, including challenging underwater recordings from full-scale facilities.
Explore WP4 →Behaviour analysis example
Following mouth-opening behaviour over time
This short underwater example illustrates how video and computer vision can capture repeated mouth movements. Such measurements can support automatic estimation of respiration-related behaviour without handling the fish.
Student research
Master's theses connected to cAIge
These projects form a growing research portfolio around salmon identification, welfare indicators, underwater sensing, and annotation-efficient AI. The theses are supervised collaboratively by researchers across the cAIge team.
2026
AI enhanced wound analysis for salmon: Few-shot segmentation for wound and anomaly detection
Develops a few-shot pipeline that uses DINOv3 features, prototype-based segmentation, CRF refinement, and temporal filtering to identify salmon anatomy and open wounds while reducing the need for manually annotated training data.
Hans Kristian Lorentzen
Supervisor: Annette Stahl
Co-supervisor: Christian Schellewald
Draft / current work
Contrastive learning for salmon re-identification
Evaluates closed-set, open-set, and cross-camera recognition for 2,015 salmon identities from industrial net-pens. The work compares CLIP- and DINO-based models and investigates domain-invariant training for robust recognition.
Christian Li Sivertsen & Jonathan Kvalvaag Dysvik
Supervisors: Rudolf Mester, Annette Stahl, Christian Schellewald, and Espen Uri Høgstedt
Draft / current work
3D reconstruction of fish in aquaculture scenes
Develops a refraction-aware stereo pipeline for dense reconstruction of salmon. The work traces calibration, correspondence, and shape-refinement errors to support reliable size and surface measurement in underwater conditions.
Christian Nielsen
Supervision: Annette Stahl, Rudolf Mester, Christian Schellewald, and Espen Uri Høgstedt (draft information)
Draft / current work
Synthetic data for computer vision in aquaculture
Creates a synthetic underwater stereo dataset with labels for depth, detection, segmentation, anatomy, and wounds. Models trained on the generated data are evaluated for transfer to real aquaculture images.
Olaf Talmo
Supervisors: Annette Stahl, Christian Schellewald, and Rudolf Mester
Unsupervised AI for fish welfare
Combines automatic video segmentation, DINO features, contamination-aware memory banks, and temporal filtering to flag visible anomalies in real salmon tracks while reducing dependence on detailed training annotations.
Ole Magnus Lærum
Supervisor: Annette Stahl
Co-supervisors: Christian Schellewald, Rudolf Mester, and Espen Uri Høgstedt
2025
Unsupervised anomaly detection for fish appearance monitoring
Transfers techniques from industrial visual inspection to underwater salmon imagery. The project builds real-world datasets and detects and localises unusual visual regions without requiring anomaly examples for training.
Eline Karlsen
Supervisor: Annette Stahl
Co-supervisors: Christian Schellewald, Rudolf Mester, and Espen Berntzen Høgstedt
Self-supervised patch features for anatomical segmentation
Uses frozen DINOv2 features to segment salmon body parts from only one annotated support image. The approach substantially reduces the manual work needed to produce detailed aquaculture datasets.
Jens Oskar Ramm-Pettersen
Supervisor: Annette Stahl
Co-supervisors: Rudolf Mester, Christian Schellewald, and Espen Berntzen Høgstedt
3D trajectory reconstruction and swimming speed estimation of salmon in sea cages using stereo vision
Combines stereo imaging, fish-head detection, multi-object tracking, and 3D trajectory estimation to analyse individual and group swimming speed as a potential basis for behavioural welfare assessment.
Mari Hetlesæter
Supervisors: Annette Stahl, Rudolf Mester, Christian Schellewald, and Espen Berntzen Høgstedt
Transformer-based architectures for Atlantic salmon re-identification
Develops specialised Transformer architectures that represent natural melanin spot patterns. The study evaluates both closed-set and open-set recognition and compares its models with established vision architectures.
Torstein Korten & Paal Markus Bjørnstad
Supervisors: Rudolf Mester, Annette Stahl, Christian Schellewald, and Espen Berntzen Høgstedt
2024
Assessing re-identification capabilities of salmon body parts
Builds a pipeline for fish detection, body-part detection, and individual recognition. Experiments compare how informative the thorax, fins, and eye are for identifying salmon from video.
Magnus Wiik
Supervisor: Annette Stahl
Co-supervisors: Christian Schellewald, Rudolf Mester, and Espen Berntzen Høgstedt
Objective analysis of melanin spots as welfare signals
Develops computer-vision pipelines to extract and measure body and head-region melanin spots before and after stress exposure, providing a foundation for studying whether spot changes can indicate welfare state.
Sindre Larsen
Supervisor: Annette Stahl
Co-supervisors: Christian Schellewald, Rudolf Mester, and Espen Berntzen Høgstedt
Image-based quantification of scale loss using machine learning
Detects fish and segments skin and scale-loss regions to quantify a relevant welfare indicator. The project also developed ScaleGuard, an interface that makes the analysis pipeline accessible to non-technical users.
Thomas Fjorden
Supervisor: Annette Stahl
Co-supervisors: Christian Schellewald, Rudolf Mester, and Espen Berntzen Høgstedt
Core team
Computer vision meets fish biology
The core team connects computer vision, machine learning, engineering cybernetics, behavioural biology, and aquaculture research.
Project manager · SINTEF Ocean
Christian Schellewald ↗
Underwater computer vision, AI for aquaculture, visual welfare indicators, and long-standing biometric salmon re-identification research—including the earlier INDISAL project ↗—as well as project leadership.
Project lead’s homepage →
Professor · NTNU
Rudolf Mester ↗
Computer vision, machine learning, robust perception, and salmon re-identification.
Professor · NTNU
Annette Stahl ↗
Computer vision, machine learning, robotics, motion analysis, and engineering cybernetics.
Professor · NMBU
Øyvind Øverli ↗
Fish behaviour, stress biology, welfare indicators, and biological interpretation and confirmation.
PhD candidate · NTNU
Espen Uri Høgstedt ↗
Salmon tracking, re-identification, computer vision, and welfare monitoring in challenging environments.
Senior Research Scientist · SINTEF Ocean · WP3 lead
Aya Saad ↗
AI reasoning, knowledge representation, machine learning, data analysis, and decision support for aquaculture.
Team photos are reproduced from the linked institutional profile pages.
SINTEF OceanNTNUNMBUResearch Council of Norway
Recognition & impact
Part of the UN Ocean Decade
cAIge is endorsed as an action of the United Nations Decade of Ocean Science for Sustainable Development 2021–2030. The project contributes enabling technology for improved fish welfare and more sustainable aquaculture.
Open research data
Datasets from cAIge
Public datasets support reproducible research in salmon tracking, re-identification, segmentation, and underwater computer vision.
The cAIge Zenodo community provides a common home for openly published project datasets and supporting research material.
BoostCompTrack
Training images, annotations, detector models, evaluation material, and underwater video for a multi-purpose salmon welfare tracking framework.
Open dataset ↗Salmon re-identification dataset
Re-identification data, segmentation annotations, trained-model outputs, evaluation results, and verified cross-camera matches.
Open dataset ↗Research outputs
cAIge related publications
Peer-reviewed publications directly connected to the project and its research topics.
Patch ensembles for robust salmon re-identification with weak trajectory labels ↗
E. U. Høgstedt, C. Schellewald, A. Stahl, and R. Mester · Accepted for IEEE ICIP, 2026 · arXiv:2605.18038
Automated computer vision based individual salmon (Salmo salar) breathing rate estimation (SaBRE) for improved state observability ↗
E. B. Høgstedt, C. Schellewald, R. Mester, and A. Stahl · Aquaculture, 2025
A multi-purpose tracking framework for salmon welfare monitoring in challenging environments ↗
E. U. Høgstedt, C. Schellewald, A. Stahl, and R. Mester · ICCV Workshops, 2025
Image-based quantification of scale loss in fish using machine learning and computer vision ↗
T. Fjorden, E. B. Høgstedt, C. Schellewald, R. Mester, M. Remen, A. V. Nytrø, and A. Stahl · ICMV 2024 proceedings, published 2025
Mouth opening frequency of salmon from underwater video exploiting computer vision ↗
C. Schellewald, A. Saad, and A. Stahl · IFAC-PapersOnLine, 2024
Modeling and predicting welfare of farmed Atlantic salmon: Integrative domain representation and reasoning strategies ↗
O. Nissen and A. Saad · Procedia Computer Science, 2024