A Dual-Segmentation Framework for the Automatic Detection and Size Estimation of Shrimp

Area of Science:

• Aquaculture technology
• Computer vision in agriculture
• Deep learning for biological measurements

Background:

• Accurate shrimp size estimation is vital for health and growth assessment in aquaculture.
• Traditional manual methods are labor-intensive, while existing computer vision techniques lack precision.
• Shrimp's physical characteristics, like curved posture and background blending, complicate size measurement.

Purpose of the Study:

• To develop an accurate and efficient deep learning framework for shrimp size estimation.
• To overcome the limitations of manual and conventional computer vision methods in aquaculture.
• To enable precise size measurements for large-scale shrimp farming monitoring.

Main Methods:

• Proposed a dual-segmentation deep learning framework integrating instance and semantic segmentation.
• Utilized the RTMDet-m model for initial shrimp instance segmentation.
• Developed a custom semantic segmentation model for precise centerline prediction.

Main Results:

• The RTMDet-m model achieved 96% average precision (AP50) with high frames per second (FPS).
• The custom segmentation model achieved the highest FPS and an F1-score of 88.3%.
• The dual-segmentation framework demonstrated superior accuracy with a mean absolute error of 1.02 cm and root mean square error of 1.27 cm.

Conclusions:

• The proposed dual-segmentation deep learning framework significantly improves shrimp size estimation accuracy.
• This method offers a more efficient and precise alternative to traditional and existing computer vision techniques.
• The framework has the potential to enhance shrimp farming management through accurate, large-scale monitoring.