Advances in camera technologies have made it easier than ever to collect high-quality underwater imagery, allowing researchers to apply image analysis techniques in many areas of fisheries science, including population assessments and ecological and behavioural studies. This greater ability to capture imagery has resulted in high-volume datasets being collected on a regular basis. As these data are often too large for any team of researchers to analyse fully, researchers have employed-turned their attention to machine learning (ML) as a way of automating tasks (e.g. detection and classification of fishes) that would otherwise require many hours to complete. Machine learning is a branch of artificial intelligence (AI) and computer science which uses data and algorithms to improve performance of a given task, for example, detecting fish in an image.
However, before investing the time needed to develop and train full ML algorithms to complete a task, it’s a good idea to test how well an algorithm can perform a task on data that represents real-world variable conditions. In the latest
ICES Journal of Marine Science Editor's Choice paper, "Evaluating automated benthic fish detection under variable conditions"
, the authors performed a set of object detection experiments on stereo images of orange roughy (
Hoplostethus atlanticus) taken by a net-attached acoustic optical system (AOS) during a biomass survey. Orange roughy is a long-living deep-sea fish that forms large aggregations and is fished commercially in Australia and New Zealand. The authors wanted to know how well a popular object detection algorithm worked under variable fish densities, different benthic substrates, and at different altitudes above the seafloor. The study found that high fish densities (>20 fish per image) caused the object detector to fail as it was unable to distinguish between individual fish, whereas substrate type and altitude did not affect model performance. The authors can now use these findings to move forward with the development of a fully optimised ML algorithm. The paper concludes by recommending that this type of real-world dataset exploration is carried out prior to committing the resources to train final object detectors.