When I first began experimenting with computer vision in industrial facilities in 2018, the deep learning AI models that exist today were still in their infancy.
At the time, training an AI model for a specific industrial use case was not nearly as straightforward as it is today. In many cases, you had to build the model block by block, referring back to original research papers just to understand how everything connected.
Every project became a messy collection of custom code, data pipelines, scripts, model files, and post-processing logic. You just hoped the project would never need to be handed off to another developer because of how complex and custom everything had become.
From 2018 to 2021, I consulted on at least a dozen different vision projects for industrial customers. Each project had unique requirements, custom models, and specific post-processing logic.
But even though the use cases were different, many of the underlying needs were the same.
Every project needed some way to save video files, extract images for annotation, train AI models, deploy and run models on edge devices, display results in web-based dashboards, monitor system health, and support long-term operation.
Those shared needs became the inspiration behind the Canopy Vision platform.
Why We Built a Platform
By building a single platform that covered these common needs, we could approach each new project without having to rebuild the same foundation from scratch every time.
That mattered for two reasons.
First, it allowed us to move faster. Instead of spending project time building basic infrastructure, we could focus on solving the customer's specific problem.
Second, it gave us a standard platform that multiple people on our team already understood. That made projects easier to support, easier to hand off, and easier to improve over time.
Over the years, we added capabilities that were specifically designed for industrial AI vision deployments. These included a gateway instance that proxies requests to and from multiple edge devices, a network separation layer for firewalls, a model training environment, centralized software updates, Single Sign-On for enterprise customers, robust health monitoring, standardized data validation tools, and the ability to deploy custom dashboards on top of the core solution.
The result is a flexible platform that has helped reduce the AI vision pilot timeline from about one year to roughly three months.
That matters because a shorter pilot timeline reduces the cost of piloting new technology and makes it easier for facilities to prove whether AI vision is feasible before committing to a permanent deployment.
Below are a few of the specific Canopy Vision platform features that help make the initial pilot phase more successful.
Video Review and Image Extraction
During the pilot phase of most AI vision projects, cameras are used to record video to a hard drive so a data scientist or developer can review the footage.
They may need to find specific videos where interesting events occurred, such as activity in the process, diverse lighting conditions, unusual operating states, or situations that should be included in the model training process.
From there, they extract images from those videos and use them to train the model.
This process can take a lot of time.
The Canopy Vision platform speeds this up with a built-in video browser that allows users to bookmark and tag videos. Tagged videos can be archived so they are not overwritten.
That means anyone with access to the platform can help flag useful video, even if they are not the data scientist or developer working on the model.
For example, if a process engineer or operator knows something interesting happened during a shift, they can mark that video before it disappears from the hard drive. That helps the project team collect better data faster.
The platform also includes tools to extract selected images from video files. This reduces the amount of time spent moving video files around, writing custom scripts, or manually converting video into training images.
Once the first model is deployed, even if the initial accuracy is low, the platform provides a dashboard that shows when different classes or events were detected. Users can click on the timeline to view video from that exact moment. If the video is useful, they can tag it, archive it, or download it directly from the platform.
This significantly reduces the time spent on video review, data gathering, and data wrangling, which are some of the most time-consuming parts of the AI vision pilot process.
Model Training and Deployment
In most vision projects, model training and model inference are not performed on the same machine.
Training usually happens on a server or workstation with a large GPU. The model is trained over hours of GPU compute time and then exported as a file that can run in real time for inference.
Inference may happen on an edge compute device in the field or on another GPU-enabled server inside the facility.
That separation makes sense technically, but it can create a lot of project overhead.
Images, configurations, and model files often need to move from one machine to another, and sometimes from one network to another. If the training environment is in an IT-managed Azure cloud environment and the edge device is on an OT network, the model may need to pass through multiple steps before it can be deployed.
And if the model has issues running on the edge device, the process may need to be repeated multiple times.
Ask me how I know.
With Canopy Vision, models trained using the training server are output into a standard format designed to work with Canopy Vision at the edge. If there is a network path through the appropriate firewalls between the training server and the edge device, the model can be deployed to the edge device through the platform without needing to SSH into the device.
The Canopy Vision training server can also be self-hosted or deployed offline to meet customer networking, security, or data policy requirements.
This makes the model training and deployment process faster, more repeatable, and easier to support.
Health Monitoring and Alerts
Industrial vision projects usually involve a mix of operations teams, engineers, data scientists, developers, networking professionals, and project managers.
During the active pilot phase, many people are watching the project closely. If an edge device goes offline or a vision pipeline crashes, the data scientist or developer working on the project will usually notice quickly.
But once the project reaches a stable state, attention naturally shifts to the next priority.
The project is successful, the model is working, and everyone assumes it will continue running forever.
Just kidding.
In reality, AI vision systems still need monitoring. Devices can lose power, pipelines can crash, storage can fill up, cameras can shift, and network conditions can change.
If there is no live stream being reviewed and no PLC data being actively monitored by an operator, a system can stop working without anyone noticing for days or even weeks.
Canopy Vision helps solve this with health monitoring and email alerts.
Users can create alerts for downtime and other health issues. These alerts are sent from the Canopy Vision gateway, so they can still be sent even if individual edge devices lose power in the field.
If the Canopy Vision gateway is provisioned in an offline environment, customers can use their own internal SMTP settings so alerts can still be sent internally.
This helps ensure that deployed AI vision systems remain visible, supported, and operational after the initial excitement of the pilot phase is over.
Controlled and Documented Data Validation
AI vision accuracy is not always easy to define.
If you ask, "What accuracy do we need for this project to be successful?" you may get a range of answers.
But the bigger question is often:
"What defines accuracy for this specific use case?"
If the project is tracking railcars in a railyard, accuracy should probably reflect whether the system correctly predicts the location of cars in the yard. It may not matter as much whether every bounding box is perfectly drawn.
If the project is estimating the amount of impurities in a material flow, then root mean square error, or RMSE, may be a better way to measure performance.
This becomes even more complicated when the AI model is being compared against human interpretation.
For example, if a system is visually estimating impurity levels on a scale of low, medium, high, and very high, two different people may disagree about how the same image should be classified.
These factors can make it difficult for a team to agree on whether a vision system is accurate enough to provide value.
Canopy Vision addresses this with structured data validation experiments inside the platform.
These experiments retain the model output data, human observations, video files, and notes. They can be used to document model performance issues, support future improvements, and even assist with control loop tuning.
Some of the experiments that can be run within Canopy Vision include:
- Step-change tests, where the team makes a step change to the process and records how the AI model responds
- Disturbance tests, where the camera or environment is physically disturbed to evaluate model robustness
- Transition tests, where the process is maintained at a transition point between two states to see how the AI model responds
This controlled and standardized approach helps everyone stay aligned. It also creates documentation that can support future decisions, troubleshooting, model updates, and operational changes.
A Faster Path to Industrial AI Vision
These features are only a few of the capabilities we have developed through our experience building Canopy Vision and deploying industrial AI vision systems.
The larger goal is simple: make AI vision easier to deploy, easier to support, and easier to scale inside real industrial environments.
Every new project teaches us something. As we continue working with more customers and more use cases, we will keep building features that make AI vision deployments faster, more reliable, and more practical.
Our vision is to make deploying AI vision feel as approachable as installing a new thermocouple or building a new spreadsheet model.
That does not mean every project is simple.
But it does mean the foundation should not need to be rebuilt from scratch every time.