How can object detection in industrial machine vision systems be reliably implemented under challenging lighting conditions?

Conventional rule-based image processing quickly reaches its limits when dealing with reflections, varying surface properties, and changing lighting conditions. The use of deep learning models such as YOLOv8 enables robust object detection through training on realistically augmented datasets. With ibaVision, such models can be integrated into existing camera systems and continuously retrained. This ensures stable detection performance even under varying production conditions.

How can AI-based machine vision systems be integrated into existing production environments?

Efficient integration requires seamless connectivity to existing data and system landscapes without adding hardware complexity. ibaVision enables the integration of AI models via standardized interfaces, for example based on Python. Detected objects and results can be directly transferred to ibaPDA and linked with process data. This creates a consistent data foundation for further analysis.

How can end-to-end traceability of visual quality characteristics in production be ensured?

Complete traceability requires storing both the original images and the derived analysis results. With ibaCapture, image data and object detection results can be archived in time synchronization. In combination with process data from ibaPDA, this creates a complete context for every detected event. This enables detailed traceability and subsequent root cause analysis.

How can extracted image data be leveraged for advanced process analysis?

Integrating machine vision results into process data systems enables direct correlation between visual events and machine states. ibaPDA captures extracted metrics, such as object counts or states, in real time. This data can then be analyzed in detail using ibaAnalyzer. As a result, correlations can be identified and optimization potential can be derived.

How can AI models for industrial applications be efficiently trained and scaled?

A structured training process begins with the annotation of real-world image data and is enhanced through targeted data augmentation to reflect production variability. Models such as YOLOv8 can be iteratively improved by continuously integrating new data into the training process. With ibaVision, trained models can be deployed into existing systems without the need for specialized hardware. This enables scalable deployment across multiple plants with minimal infrastructure effort.

AI-Powered Coil Strap Detection with ibaVision

Success Story

In the production of aluminum coils, proper strapping is essential to ensure safe handling and transport. Missing or improperly fastened straps pose a significant safety risk, especially in high-throughput finishing lines where manual checks are prone to human error.

Solution Benefits Conclusion Used Functions More Success Stories

The objective of the project was to automatically detect both the number of coils and the number of straps using existing camera infrastructure, while ensuring seamless integration into the plant’s process data environment. The solution was not intended as a safety-certified system, but rather as a tool to improve process reliability and provide operators with valuable insights in real time. However, coil strap detection under real-world conditions presented multiple difficulties: coils differ in size, surfaces can be shiny or matte, and reflections and varying lighting conditions made traditional image processing unreliable. The fixed camera perspective further limited options for optimized image capture.

Implementation with ibaVision

A deep learning approach with YOLOv8, a fast and lightweight object detection framework, offered the flexibility needed to handle visual variations and improve detection accuracy over time by simply adding more training images. Unlike static filter-based methods, AI models can be retrained and optimized continuously with minimal effort. An initial dataset of coil and strap images was manually annotated with bounding boxes. To account for real-world variability, the dataset was augmented by altering brightness, contrast, and sharpness and by simulating blur and noise. Three separate YOLOv8 models were trained in successive stages. The trained model was then integrated into ibaVision via a Python-based interface. The solution runs on the customer’s existing virtual machine without requiring GPU hardware. For every detected coil, ibaVision logs both the original camera image and the processed detection result into ibaCapture, while the number of coils and straps is recorded in ibaPDA for further process analysis.

System diagram showing synchronized data flow between ibaPDA, ibaCapture, and ibaVision YOLOv8.

Benefits

Reliable detection under challenging real-world conditions, even with reflections and lighting variations
Seamless integration into the iba environment, enabling contextualized process analysis
Easy retraining and redeployment, making the solution adaptable for future lines or additional projects

Conclusion

By combining YOLOv8’s AI-based detection with ibaVision’s powerful visualization and logging capabilities, the project successfully demonstrated how modern machine vision can increase process transparency and reliability without requiring complex infrastructure or advanced AI expertise on the user side.

With the combination of ibaVision, ibaCapture, and ibaPDA, we created a practical AI-based solution that fits seamlessly into the customer’s existing environment. The detection model, running within ibaVision, allows us to reliably count coils and straps, while ibaPDA logs the results and ibaCapture stores the corresponding images.

Artuur Krekels
Application & Consulting Engineer, iba Benelux

Used Functions

The following functions of the iba system are used in this application.

Explore More Success Stories

Read more inspiring success stories of our satisfied customers.

Frequently Asked Questions (FAQ)

How can object detection in industrial machine vision systems be reliably implemented under challenging lighting conditions?: Conventional rule-based image processing quickly reaches its limits when dealing with reflections, varying surface properties, and changing lighting conditions. The use of deep learning models such as YOLOv8 enables robust object detection through training on realistically augmented datasets. With ibaVision, such models can be integrated into existing camera systems and continuously retrained. This ensures stable detection performance even under varying production conditions.
How can AI-based machine vision systems be integrated into existing production environments?: Efficient integration requires seamless connectivity to existing data and system landscapes without adding hardware complexity. ibaVision enables the integration of AI models via standardized interfaces, for example based on Python. Detected objects and results can be directly transferred to ibaPDA and linked with process data. This creates a consistent data foundation for further analysis.
How can end-to-end traceability of visual quality characteristics in production be ensured?: Complete traceability requires storing both the original images and the derived analysis results. With ibaCapture, image data and object detection results can be archived in time synchronization. In combination with process data from ibaPDA, this creates a complete context for every detected event. This enables detailed traceability and subsequent root cause analysis.
How can extracted image data be leveraged for advanced process analysis?: Integrating machine vision results into process data systems enables direct correlation between visual events and machine states. ibaPDA captures extracted metrics, such as object counts or states, in real time. This data can then be analyzed in detail using ibaAnalyzer. As a result, correlations can be identified and optimization potential can be derived.
How can AI models for industrial applications be efficiently trained and scaled?: A structured training process begins with the annotation of real-world image data and is enhanced through targeted data augmentation to reflect production variability. Models such as YOLOv8 can be iteratively improved by continuously integrating new data into the training process. With ibaVision, trained models can be deployed into existing systems without the need for specialized hardware. This enables scalable deployment across multiple plants with minimal infrastructure effort.