A novel real-time crayfish weight grading method based on improved Swin Transformer

Area of Science:

• Computer Vision
• Machine Learning
• Aquatic Automation

Background:

• Traditional crayfish weight classification relies on manual inspection, which is labor-intensive and prone to inaccuracies.
• Existing automated methods often lack precision in handling variations in crayfish size and shape.

Purpose of the Study:

• To develop an automated crayfish weight classification system using an improved Swin-Transformer model.
• To enhance the accuracy of weight prediction by integrating image segmentation with regression analysis of crayfish part areas.
• To validate the system's performance in real-time weight grading for industrial applications.

Main Methods:

• An improved Swin-Transformer model was utilized for precise crayfish image segmentation.
• A multiple regression model was developed, correlating projected area with actual weight, incorporating relationships of individual crayfish parts.
• The system was validated on a test set of 40 samples and evaluated through grading experiments.

Main Results:

• The Swin-Transformer model achieved 90.36% Mean Intersection over Union (MIOU) and 99.0% segmentation accuracy, outperforming other leading models.
• The regression model demonstrated a high correlation coefficient (0.983) and an average prediction accuracy of 98.34%.
• The automated grading system achieved over 86.5% accuracy, confirming its practical feasibility.

Conclusions:

• The proposed method offers a novel and accurate approach to automated crayfish weight classification and grading.
• Integrating image segmentation with part-based area-weight correlation significantly improves prediction accuracy over traditional methods.
• This system holds substantial potential for industrial automation in the aquatic processing sector, enhancing quality control and production efficiency.