Abstract

This project details developing a robotic sorting system for the Food Industry Quality Control with Robotic Conveyor Belt Inspectors that uses computer vision and deep learning to enhance efficiency and reduce human error. The system, implemented as a conveyor belt inspector, identifies contamination and pollution in glass jars before they are filled with food. By utilizing YOLOv6 and OpenCV for object detection and image processing, the system improved detection accuracy by 20%. Challenges, such as distinguishing scratches from contamination, were addressed with anomaly detection techniques. Additionally, a CI/CD pipeline and MLOps practices were introduced to streamline model development and deployment, resulting in a 50% reduction in human error and a 25% increase in production.

Introduction

Ensuring food safety and quality is critical in the food industry. Manual inspection of containers often leads to inefficiencies and inconsistencies, impacting production and customer satisfaction. To address these issues, we developed an autonomous robotic system that inspects glass jars on a conveyor belt, detecting contamination and pollution using advanced computer vision and deep learning techniques. This solution reduces human error and boosts production efficiency while maintaining high-quality standards.

Methodology

1. System Design The robotic system consists of:

• Conveyor belt with camera-mounted inspection stations
• Embedded processing devices (Jetson NX) for real-time analysis
• Integration with anomaly detection for specialized cases

2. Object Detection and Image Processing YOLOv6 was chosen for its high speed and accuracy in object detection, coupled with OpenCV for image preprocessing. Key steps included:

• Training a model on images of glass jars with and without contamination
• Addressing challenges such as scratches misclassified as contamination
• Implementing image enhancement to improve detection accuracy

3. Anomaly Detection for Scratches A separate anomaly detection module was developed to distinguish scratches from contamination. This module uses:

• Unsupervised learning to identify deviations from normal patterns
• Integration with the YOLOv6 pipeline for enhanced precision

4. CI/CD Implementation and MLOps To ensure efficient deployment and monitoring, a CI/CD pipeline was designed:

• Automated model training, testing, and deployment
• Continuous integration of updates using cloud-based services
• Monitoring system performance and retraining as needed

5. Real-Time Performance The system was deployed on Jetson NX devices, enabling:

• Camera-based real-time detection
• On-device processing for reduced latency and improved throughput

Experiments

The robotic sorting system achieved the following:
• 20% increase in detection accuracy compared to previous methods
• 50% reduction in human error during quality inspections
• 25% increase in overall production efficiency
• Effective differentiation between scratches and contamination using anomaly detection

Results

1. Scratch Misclassification Scratches at the bottom of jars were frequently detected as contamination. This was resolved by:
   • Developing a custom anomaly detection algorithm
   • Enhancing model training with additional labeled data
2. System Integration Integrating real-time detection and cloud-based monitoring required careful orchestration of software and hardware. This was achieved through robust CI/CD pipelines and modular design.

Conclusion

The conveyor belt inspector robot represents a significant advancement in food industry automation. By leveraging YOLOv6, OpenCV, and MLOps practices, the system improved quality control, reduced human error, and enhanced production efficiency. Future work will focus on expanding the system’s capabilities to include other container types and integrating predictive maintenance features for further optimization.