Abstract
Salmon farming is becoming increasingly important in the global food chain, which necessitates a deep understanding of how internal salmon states relevant for production yield and welfare is evinced externally on salmon, in order to facilitate real-time detection of inadequate conditions. In this thesis, one particular external trait of salmon, ventilation frequency, is elucidated in great detail to evaluate how it can inform on the condition of salmon in tanks and net pens. To allow for easy and automatic extraction of this trait, a complete pipeline capable of estimating ventilation frequency of individual salmon from a video recording is developed. By the use of State Of The Art deep learning methods, the algorithm is capable of detecting and tracking salmon, estimating mouth poses in order to calculate ventilation frequency, and determining the unique salmon individual a fish belongs to. Upon completion of the salmon ventilation frequency extraction method, it was applied to data from a salmon stress experiment, unveiling that salmon ventilation frequency increase in response to reduced dissolved oxygen content or disturbances, and that individual salmon show consistent respiration patterns across hours and weeks. By observing the capability of the constructed salmon ventilation frequency pipeline to discern salmon welfare trends in a stress experiment, it can be concluded that the method is capable of automated evaluation of a salmon welfare indicator.