Abstract
Preserving fish welfare is of major priority for fish farming companies and is essential for the sustainability and future growth of the aquaculture industry. Wounds represent a serious welfare issue for the fish, both as a symptom of existing problems and a precursor to potential future vulnerabilities. Existing static or moving vision-based sensor systems paired with computer vision methods offer promise for monitoring different welfare indicators, but have not yet been well adapted for wound detection. Targeting this challenge, this paper introduces a computer vision framework that leverages object detection and tracking algorithms to automate wound detection and tracking in video recordings. Tested on unseen video images from operational salmon fish farms, the model achieves 90% accuracy and showcases enhanced robustness in highly complex and dynamic scenarios due to the added tracking capabilities. Currently an assistive tool for detecting wounds during manual camera inspections, the framework has the potential for future development into a fully autonomous, data-driven approach to fish welfare monitoring.