text.skipToContent text.skipToNavigation

Running AI Models with ONNX Directly in uniVision 3

With the “Image ONNX” module, you can run AI models directly on the uniVision 3 hardware such as the B60 smart camera or MVC machine vision controller. Leverage seamless ONNX inference for industrial image processing tasks – without additional hardware or customization of your development environment.

AI model training

Export to ONNX format

u3o package

Integration into uniVision 3

Interface connection

Step 1: Your Training, Your Framework

Use your preferred machine learning framework such as PyTorch, TensorFlow or other ONNX-compatible tools without changing your existing development environment. You are not tied to proprietary software or cloud platforms and retain maximum flexibility in the development of your AI models.
Train your models on-premises, in the cloud or on your existing GPU clusters, keeping full control of your data, model architectures and hyperparameters at all times. Your data scientists can continue to work with the tools they know and use efficiently.

Step 2: Export to ONNX Format

ONNX (Open Neural Network Exchange) is an open, vendor-independent standard for exchanging AI models. Conversion from your framework takes place with a few lines of code via integrated export functions.
wenglor notebooks show best practices and support with export, validation and further processing. The technologies classification (multi-class and multi-label) and object detection are supported – including optional quantization for optimized performance.

Step 3: u3o Package

The u3o format is the deployment-ready package for devices with the uniVision 3 software. It combines the ONNX model with all relevant metadata such as input/output specifications, preprocessing and class labels.
Creation takes place automatically via wenglor Python scripts in GitHub, including validation for the target hardware. Optionally, the AI model can be quantized to further increase performance.

Step 4: Integration into uniVision 3

The “Image ONNX” module enables the direct import of u3o packages into uniVision 3. Depending on the application, flexible hardware options are available: the B60 series smart cameras for compact AI inference directly on the system or the MVC series machine vision controllers for computing-intensive models. Combined with additional modules, powerful and flexible machine vision applications can be realized.

Step 5: Connection

Integration is seamless via existing interfaces to PLCs, robots and IT systems. The existing communication structure of uniVision 3 is used – without any additional middleware.

Supported Technologies and Model Types for the ONNX Module

The ONNX format is supported by all common machine learning frameworks, so your existing setup can continue to be used unchanged.

Supported model format	ONNX
Training framework	PyTorch, TensorFlow, scikit-learn, Apache Spark, Keras, Caffe2, Microsoft Cognitive Toolkit (CNTK), Theano*
Supported model types	Multi-class classification Multi-label classification Object detection
Optimization (optional)	INT8 quantization is supported and recommended for running on the B60 smart camera. When running on the MVC machine vision controller, both quantized FP32 networks and quantized INT8 networks are supported.

*In addition to the listed frameworks, further integrations are possible.

The ONNX format is supported by all common machine learning frameworks, so your existing setup can continue to be used unchanged.

Supported model format

ONNX

Training framework

PyTorch, TensorFlow, scikit-learn, Apache Spark, Keras, Caffe2, Microsoft Cognitive Toolkit (CNTK), Theano*

Supported model types

Multi-class classification
Multi-label classification
Object detection

Optimization (optional)

INT8 quantization is supported and recommended for running on the B60 smart camera.
When running on the MVC machine vision controller, both quantized FP32 networks and quantized INT8 networks are supported.

*In addition to the listed frameworks, further integrations are possible.

Why Use GitHub for Machine Learning Projects?

Ready-made Jupyter notebooks for every step of the machine learning workflow
Clear metadata structure thanks to detailed JSON specification
Regular updates for new framework versions
Easily customizable and expandable Python code examples

What Are the Advantages of ONNX in uniVision 3?

Open standard ONNX without vendor lock-in

Use the framework of your choice (e.g. PyTorch, Tensorflow, etc.).
Your existing processes remain unchanged.

AI and rule-based image processing in one workflow

Benefit from combining neural networks with rule-based tools.
Leverage seamlessly integrated tools such as masking, ROI definition and post-processing.

Hardware and software optimized for ONNX runtime

The B60 smart cameras have a neural processing unit for accelerated AI calculation.
All MVC machine vision controllers ensure optimal process times with a powerful CPU.

Numerous industrial interfaces for seamless integration

All results can be processed directly via the existing uniVision 3 communication infrastructure.

Two Ways to Run AI Models on wenglor Hardware

The uniVision AI license package enables the execution of AI models either directly from the AI Lab or trained externally via the ONNX format. uniVision AI ensures seamless integration of your AI models, while end-to-end model training can take place in the AI Lab or with common machine learning frameworks in its own toolchain.

“Image ONNX” Module – Integration of Externally Trained AI Models in ONNX Format

Integrate your AI models trained with open source frameworks.

Model training	In your own toolchain (e.g. PyTorch, Tensorflow) locally or externally
Interface	AI model import in ONNX format (via GitHub)
Transparency	Heatmap support (model dependent)
Integration	Direct execution in uniVision – compatibility and conversion according to documentation

“Image AI” Module – AI Model Training in the AI Lab

Leverage the entire AI workflow from a single source.

Model training	In the cloud in the AI Lab
Data flow	Image transfer from uniVision 3 to the AI Lab via weHub
Transparency	Maximum traceability via heatmaps and evaluations
Access	Intuitive user interface for structured projects without their own machine learning toolchain

The Right Hardware for Optimal Execution of AI Models

With the uniVision 3 software, the wenglor machine vision hardware provides the perfect basis for running AI models efficiently.

On the B60 series smart cameras, the integrated Neural Processing Unit (NPU) ensures efficient and fast running of the AI models. The Intel® OpenVINO™ acceleration and multicore processor of the MVC series machine vision controllers enable efficient, parallel execution of multiple AI models and complex process workflows.

Smart Cameras B60

Find out more

Machine Vision Cameras with Machine Vision Controller

Find out more

With uniVision 3, you get a scalable solution for easily integrating AI into your industrial image processing applications!

Licenses for the Use of ONNX in uniVision 3

For the execution of the AI models created in the AI Lab, the “Image AI” module is available on the relevant machine vision hardware. This is part of the “uniVision AI” license package.

The “uniVision AI” license package includes the activation of the following modules:

“Image AI” module (for AI models from the AI Lab)
“Image ONNX” module (for AI models in ONNX format)

The “uniVision AI” license is available as follows:

For B60 series smart cameras: License DNNL031
For MVC series machine vision controllers: License DNNL032

Product Comparison

3/4

4/4

5/4

6/4