Computer-Simulated Virtual Image Datasets to Train Machine Learning Models for Non-Invasive Fish Detection in...

Area of Science:

• Aquaculture technology
• Computer vision
• Machine learning

Background:

• Recirculating aquaculture systems (RASs) benefit from AI and ML for management.
• High fish density and turbidity in RASs challenge underwater image acquisition for ML models.
• Manual image annotation is subjective, time-consuming, and labor-intensive.

Purpose of the Study:

• To simulate fish schooling behavior for RAS conditions.
• To investigate the use of computer-simulated virtual images for training fish detection models.
• To develop a process for expediting model training and automating virtual image annotation.

Main Methods:

• Developed a process flow for simulating fish schooling behavior and automating virtual image annotation.
• Trained and compared fish detection models using solely virtual images, solely real images, and a combination of both.
• Evaluated model performance using mean average precision (mAP) and F1 score.

Main Results:

• A model trained only on virtual images performed poorly (mAP = 62.8%, F1 = 0.61).
• A mixed model (M6) with a 90:10 virtual-to-real image ratio (630 virtual, 70 real) achieved high performance (mAP = 91.8%, F1 = 0.87).
• The M6 model's training time was seven times shorter than the model trained on real images.

Conclusions:

• Virtual image simulation is a promising approach for rapidly training reliable fish detection models for RAS.
• Combining a small proportion of real images with virtual images significantly enhances model performance.
• This method offers a more efficient alternative to traditional data acquisition and annotation for ML in aquaculture.