Damage Detection of Unwashed Eggs through Video and Deep Learning

Area of Science:

• Agricultural Engineering
• Computer Vision
• Food Science

Background:

• Broken eggs pose health risks and logistical challenges in production and transportation.
• Accurate, real-time detection of egg integrity is crucial for quality control.

Purpose of the Study:

• To develop a video-based detection model for real-time identification of broken eggs in dynamic environments.
• To enhance the YOLOv5 algorithm for improved accuracy in detecting eggshell integrity.

Main Methods:

• A novel system was designed for continuous egg rotation and translation to capture the entire surface.
• The YOLOv5 model was improved by integrating Channel Attention (CA), BiFPN, and GSConv.
• ByteTrack was employed for object tracking to associate detections across video frames, using a five-frame sequence for classification.

Main Results:

• The enhanced YOLOv5 model demonstrated improvements in precision (+2.2%), recall (+4.4%), and mAP:0.5 (+4.1%) compared to the original YOLOv5.
• The integrated video-based model achieved a field accuracy of 96.4% for broken egg detection.
• The video-based approach proved more effective than single-image methods for detecting moving eggs.

Conclusions:

• The improved YOLOv5 model with ByteTrack offers a robust solution for real-time, video-based detection of broken eggs.
• This non-destructive testing method is suitable for dynamic industrial settings, enhancing quality assurance in egg processing.
• The study provides a valuable reference for developing advanced video-based non-destructive testing technologies.