AI Vision Counting System for Food Production: High-Speed Detection of Overlapping Pickled Mustard Packages

Introduction

Accurate product counting is essential in automated food manufacturing. In high-speed packaging lines, counting errors caused by overlapping products can lead to incorrect package quantities, production interruptions, and increased labor costs.

A food manufacturer producing individually packaged pickled mustard (Zha Cai) faced this exact challenge. Traditional machine vision algorithms struggled to distinguish overlapping packages, resulting in inaccurate counts during continuous production.

To solve the problem, an AI vision counting system based on deep learning object detection and multi-object tracking was deployed, enabling reliable real-time counting even when packages overlap on a fast-moving conveyor belt.

On the pickled mustard packaging line, individual packages move continuously along the conveyor. The machine vision system is responsible for counting each package in real time.

Once the preset quantity is reached, the vision system immediately sends a signal to the upstream conveyor controller, stopping material feeding and allowing the automatic packaging process to proceed.

However, conventional image-processing algorithms were unable to accurately count packages when they overlapped or touched each other during transportation.

To improve counting accuracy and support automated quantitative packaging, the customer upgraded to a deep learning machine vision inspection system capable of object detection and intelligent tracking.

1. Real-Time Processing for High-Speed Production

The production line operates at high speed, leaving only milliseconds for image acquisition, AI inference, object tracking, and counting.

The inspection system must provide:

• Ultra-low latency
• Stable real-time performance
• Continuous high-speed operation
• Accurate counting without slowing production

2. Frequent Package Overlap

Flexible food packages often overlap or stick together while moving on the conveyor.

Traditional rule-based vision algorithms such as threshold segmentation and template matching cannot reliably separate overlapping packages, leading to:

• Under-counting
• Missed detections
• Inconsistent packaging quantities

Deep learning object detection provides far better robustness in these complex scenarios.

The inspection system combines deep learning object detection with multi-object tracking technology to accurately identify and count every package throughout the production process.

Key features include:

• AI-based detection of overlapping packages
• Real-time object tracking across image frames
• Automatic counting with configurable quantity thresholds
• Instant communication with conveyor control systems
• Stable performance under continuous production

Even when multiple packages overlap, the system accurately identifies each individual item and maintains reliable counting performance.

Extensive production testing demonstrated excellent accuracy and speed.

Reliable Detection of Overlapping Packages

The system accurately separates and counts overlapping packages with an overlap rate of up to 50%, maintaining stable counting performance under challenging production conditions.

Ultra-Fast AI Inference

The complete inspection process—including image acquisition, object detection, tracking, and counting—takes only 17 milliseconds per frame, supporting production capacities of more than 400 packages per minute.

High Detection Accuracy

The AI model was trained using only 100 training images and evaluated on 335 test images.

Test results:

• Training Images: 100
• Test Images: 335
• Correct Detections: 334
• Detection Accuracy: 99.7%

These results demonstrate that deep learning can achieve excellent performance even with relatively small training datasets.

The inspection system is built using standardized industrial machine vision components for reliable deployment and easy scalability.

Industrial Camera

MV-CA013-A0GC

Industrial Lens

MVL-HF0828M-6MPE

LED Light Source

MV-LLDS-507-38-W

Industrial Vision Controller

MV-VC4000

The standardized hardware architecture simplifies installation and enables rapid deployment across multiple production lines.

Accurate Counting of Overlapping Packages

Deep learning separates overlapping flexible packages that traditional vision systems cannot distinguish, dramatically improving counting accuracy.

High-Speed Real-Time Inspection

The lightweight AI model performs millisecond-level inference while supporting production speeds exceeding 400 packages per minute.

Excellent Accuracy with Limited Training Data

Only a small number of labeled images are required to achieve nearly 99.7% detection accuracy, reducing project implementation time and annotation costs.

Easy Expansion to Other Food Products

The same AI vision counting solution can be adapted to many food manufacturing applications, including:

• Betel nut counting
• Snack package counting
• Instant noodle accessory inspection
• Candy counting
• Frozen food packaging
• Small sachet counting
• Granular food products

Deep learning machine vision is increasingly replacing conventional photoelectric sensors and traditional image-processing algorithms in food production.

Typical applications include:

• Automatic product counting
• Packaging quantity verification
• Missing item detection
• Conveyor inspection
• Multi-object tracking
• Intelligent packaging systems
• Food quality inspection
• Smart factory automation

Compared with traditional counting technologies, AI vision systems maintain high accuracy even when products overlap, move at variable speeds, or appear in irregular positions.

High-speed food production requires counting systems that are both accurate and capable of real-time operation.

This AI vision inspection solution combines deep learning object detection with intelligent object tracking to accurately count overlapping pickled mustard packages while maintaining production speeds above 400 packages per minute.

With 99.7% detection accuracy, millisecond-level processing, and standardized industrial hardware, the solution helps food manufacturers improve packaging quality, reduce manual intervention, and accelerate the transition toward fully automated smart factories.

Why do traditional machine vision systems fail when food packages overlap?

Conventional image-processing algorithms rely on fixed rules such as threshold segmentation or template matching. When flexible packages overlap or touch each other, these methods cannot accurately separate individual objects, resulting in counting errors.

How does AI improve package counting accuracy?

Deep learning object detection recognizes each package based on learned visual features rather than predefined rules, allowing the system to identify overlapping products and maintain highly accurate counting in real time.

Can this AI vision system be used for other food products?

Yes. The solution can be adapted for counting snacks, betel nuts, instant noodle accessories, candy, frozen foods, seasoning sachets, and many other packaged or loose food products.

Is the system suitable for high-speed production lines?

Absolutely. The complete detection and tracking process requires only 17 ms per frame, enabling stable inspection at production speeds exceeding 400 packages per minute.