Project Overview
Traditional stress monitoring techniques are often invasive and time-consuming, as they require the collection of biological samples (e.g. blood, mucus or scales) for subsequent laboratory analysis. This may necessitate the sacrifice of otherwise healthy-seeming individuals. Using computer vision, we can monitor stress non-invasively through the detection of behaviours commonly associated with stress, such as erratic swimming, surface-seeking and the formation of dense groups (also known as 'balling' behaviour).
This project, developed with a local Australian aquaculture farm, involved the creation and training of a bespoke computer vision model to detect balling behaviour in barramundi (Lates calcarifer). As part of a full-stack solution, this model is designed to integrate seamlessly with the farm's existing infrastructure and CCTV systems. Additionally, the capacity for continual learning (CL) will be implemented in a subsequent update to ensure the model continues to learn and grow with new data, and enable farmers to improve the model's performance over time.
Project Information and Development Process
As this project was developed as part of my PhD research, this project was at no cost to the partnering farm. Similarly, the development time of this model is not neccessarily reflective of the time it would take to develop a similar model in a non-academic context, as I sought to not only develop a functional model for industrial application, but used this dataset to contribute to the scientific understanding of the applications of attention modules in computer vision models for aquaculture. In particular, I conducted a series of experiments to test the contributions of three different attention modules to the performance of computer vision models in aquaculture. A research paper describing these experiments and contributions is currently in preparation, and will be made available here upon publication.
Human-Centered AI Design
This project was developed with a human-centered AI (HCAI) design approach, which involved the active involvement of farmers and other end-users throughout the development process to ensure the model is designed with their needs and preferences in mind. This included meetings and feedback sessions with the partnering farm, both online and in-person, to gather their input and insights on the model's design and functionality, as well as the creation of user-friendly documentation and training materials to support the farm in using and maintaining the model.