Robust Crayfish Size Estimation Under Complex Poses: A Hybrid YOLO-LSTM Framework With Eye Distance Regression

Area of Science:

• Aquaculture automation
• Computer vision
• Crustacean biometrics

Background:

• Accurate size estimation of crayfish (Procambarus clarkii) is crucial for aquaculture automation.
• Complex poses like curled bodies and occlusions pose significant challenges for current estimation methods.

Purpose of the Study:

• To develop a Robust Crayfish Size Estimation (RCSE) framework for automated aquaculture.
• To introduce a novel pose-invariant biometric feature (eye distance) for improved size estimation.

Main Methods:

• A two-stage system combining YOLO-based object detection and LSTM regression.
• Utilized YOLOv11 for eye detection and segmentation, with CBAM-enhancement for complex scenarios.
• Integrated eye distance with pincer length and head dimensions in an LSTM network for size prediction.

Main Results:

• Achieved 98.2% precision in eye detection and 99.50% mAP50 in body segmentation.
• The LSTM network attained R² = 0.897 for size estimation, a 23.6% improvement over traditional methods.
• Demonstrated robustness to morphological variations and complex poses.

Conclusions:

• The RCSE framework offers a new vision-based paradigm for crustacean biometrics.
• The system is highly applicable to automated grading systems in precision aquaculture.
• Eye distance proves to be a valuable pose-invariant feature for accurate size estimation.