Research project Analaysis and use of synthetic data to optimalise the automatci recognition, classificationa and recognition to measure catches in mixed demersal fisheries

General introduction

AI (artificial intelligence) developments in the marine environment, such as automatically distinguishing between different fish species, sometimes have practical bottlenecks. In the SynFish project, the researchers are focusing on generating synthetic datasets of fish species, to (also) approximate real data. It is difficult to collect enough real training data for the AI training phase, in case of rare or protected species. This creates "class imbalance" in AI training. Synthetic data offers a solution to this and help train recognition software. The software thus learns to better detect and count (even these) fish, even in complex situations. The project forms the basis for further applications of synthetic data in machine vision within fisheries.

Research approach

The project is structured around four main components. The first two cover overall coordination, administration, legal framework, and communication. A third section focuses on generating synthetic datasets for selected fish species, in collaboration with the external partner Vintecc. These datasets are then used in a fourth phase to train and evaluate fish detection software. An iterative approach is applied between these phases to continuously refine the software. The focus lies on improving accuracy and practical usability in more complex scenarios.

Relevance/Valorization

The SynFish project contributes to the improvement and optimization of electronic monitoring systems on board commercial fishing vessels by expanding and enriching annotated datasets for training machine vision and AI systems. By increasing the accuracy of fish detection, classification, and length measurement, more valuable data can be obtained for scientific analyses and stock assessments. As such, the project outcomes have the potential to significantly enhance the scientific and policy relevance of automatically collected fisheries data.

Financing

EFMZVA
FIVA