Case study · AI in production
AI Fish Sorting and Grading System for Aquaculture
AI sorting machines that image juvenile fish and classify them by sex and other traits in real time, at around 10,000 fish an hour. WeyBee's computer vision and software engineers work on the product.
AIThe client
In fish farming, separating fish by sex and removing unviable fish early improves growth efficiency and yield, but manual sorting is slow and stressful for the fish. This product images juvenile fish non-invasively as they pass through the machine, and neural networks trained on millions of labelled images and videos sort them in real time.
The challenge
- Classify each fish accurately at production speed, around 10,000 fish an hour.
- Keep accuracy consistent across a full working day and changing conditions.
- Detect more traits over time, such as deformities and early maturation.
- Update software on machines in the field without disrupting operations.
What we did
We provide computer vision and software engineers who work as part of the client's R&D and product teams. They work on the imaging and classification software, the models' integration into the sorting machines, and the tools that support new traits and over-the-air updates. The client's scientists lead the research; WeyBee's engineers help turn it into production software.
What it does
Real-time classification
Fish imaged and classified as they pass, around 10,000 an hour.
Sex sorting
Fish separated by sex so each group can be raised for its best growth.
Trait detection
Detection of unviable fish and deformities, with more traits in development.
Over-the-air updates
New models and software pushed to machines in the field.
Simple operation
Runs with two operators and minimal training, alongside existing vaccination lines.
Results
- Fish are sorted automatically at production speed instead of by hand.
- Consistent, near-complete classification accuracy through the working day.
- Farms can harvest each group at its best time and remove unviable fish early.
Computer vision that sorts
fish in real time
The client's machines image juvenile fish and classify them by sex and other traits as they pass, at around 10,000 fish an hour. Our computer vision and software engineers work inside the client's R&D and product teams.
Deep-learning models classify each fish from images in real time, with no handling or tagging.
Beyond sex, the system detects deformities and other traits that matter to producers.
Improved models and software reach machines in the field without a site visit.
Our engineers contribute to the computer vision and the software that runs it on the machines.