One-stage CNN detector-based benthonic organisms detection with limited training dataset.

Area of Science:

• Marine biology
• Computer vision
• Machine learning

Background:

• Benthonic organism detection faces challenges due to unique shapes and limited training data.
• Accurate identification is crucial for ecological monitoring and resource management.

Purpose of the Study:

• To propose an effective one-stage CNN detector-based benthonic organisms detection (OSCD-BOD) scheme.
• To address challenges of unique shape dimensions and limited datasets in benthonic organism identification.

Main Methods:

• Developed an OSCD-BOD scheme integrating generalized intersection over union (GIoU) for localization accuracy.
• Employed K-means-based dimension clustering to derive multiple benthonic organisms anchor boxes (BOAB) for improved recall.
• Utilized geometric and color transformations (GCT) for data augmentation to prevent overfitting and enhance generalization.

Main Results:

• The OSCD-BOD scheme significantly enhanced localization accuracy and recall ability.
• Data augmentation prevented overfitting and improved detection generalization in complex underwater conditions.
• Achieved superior mean average precision compared to Faster R-CNN, SSD, YOLOv2, YOLOv3, and CenterNet.

Conclusions:

• The proposed OSCD-BOD scheme effectively addresses limitations in benthonic organism detection.
• The integration of GIoU, BOAB, and GCT provides a robust solution for underwater ecological surveys.
• OSCD-BOD demonstrates significant performance improvements over existing object detection methods.