Physiological state recognition model of small silkworm based on improved YOLOv5

Area of Science:

• Agricultural Science
• Computer Vision
• Animal Science

Background:

• Silkworm breeding is vital for rural economies but faces challenges in early larval stages due to high mortality.
• Accurate identification of silkworm physiological states is crucial for timely farmer intervention and improved breeding efficiency.

Purpose of the Study:

• To develop and validate an enhanced YOLOv5 model for accurate and rapid identification of small silkworm age and dormancy status.
• To improve silkworm breeding efficiency and support rural revitalization through advanced AI technology.

Main Methods:

• Trained a neural network using manually labeled data to identify silkworm age based on physical characteristics and body length.
• Implemented an improved YOLOv5 architecture with C3-SE attention mechanism and Focal-EIoU loss function.
• Evaluated dormancy by analyzing the movement of the silkworm's center point coordinates against a set threshold.

Main Results:

• The enhanced YOLOv5 model achieved an average accuracy of 92.2%, a 2.29% improvement over the original model.
• Demonstrated a 0.3% increase in accuracy, 3.4% improvement in recall rate, and 7.7% enhancement in frames per second.
• The model effectively identified silkworm age and dormancy, crucial for monitoring vulnerable early developmental stages.

Conclusions:

• The improved YOLOv5 model offers an accurate and efficient solution for monitoring small silkworm physiological states.
• This technology has the potential to significantly enhance silkworm breeding practices and contribute to agricultural development.
• The integration of attention mechanisms and improved loss functions advances the application of computer vision in sericulture.