Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.edgeimpulse.com/llms.txt

Use this file to discover all available pages before exploring further.

Visual regression lets you train a model that outputs a continuous numeric value from an image, rather than a discrete class label. This is useful whenever you need to measure or estimate something from a visual input: a fill level in a container, the angle of a dial or knob, the degree of wear on a component, or any other quantity that lies on a scale. By the end of this tutorial, you will have trained and tested a visual regression model using Edge Impulse Studio, using images of vial tubes filled to different levels as the example dataset.

​
What You’ll Need

  • A free Edge Impulse account.
  • Images of the object or scene you want to measure, each labeled with its numeric value. For this tutorial you can use the Vial Tubes dataset or collect your own.

​
Getting Started

The quickest way to follow along is to clone the public project, which already contains the dataset, impulse design, and a trained model:
  1. Open the Vial Tubes public project in Edge Impulse Studio.
  2. Click the Clone this project button in the top-right corner.
  3. Follow the prompts to copy the project into your account.
If you’d rather start from scratch with your own data, create a new project and continue with the steps below.

​
Building the Project Pipeline

​
Collect and label data

For visual regression, every image needs a numeric label — the value the model will learn to predict. In this dataset, images are labeled with the fill level of the vial (e.g., 0, 0.5, 1, 1.5, 2), where -1 indicates an empty or inverted vial.
Consistent framing and lighting across your images will significantly improve model accuracy. Capture images across the full range of values you want to predict, and distribute them evenly across the Training and Test sets. You can upload images using any of the standard Edge Impulse methods: For AI-assisted labeling of images with continuous values, see AI labeling for regression.

​
Create the impulse

Go to Create Impulse in the menu. Add an Image processing block and a Regression learning block.
The Image processing block resizes your images and converts them to the format the model expects. A resolution of 96×96 pixels offers a good balance between accuracy and on-device performance — you can increase it if your target hardware allows. Once the blocks are configured, click Save Impulse.

​
Generate features

Open the Image section from the menu and click Save parameters, then navigate to the Generate features tab and click Generate features. Edge Impulse will extract visual features from every image in the Training set. When complete, the Feature Explorer shows a 2D projection of those features — images that look similar appear close together. For regression, you should see a smooth gradient across the feature space that correlates with the labeled values.

​
Training the model

Navigate to the Regression section and click Start training. Edge Impulse trains a neural network that maps image features to a predicted numeric value. After training you will see metrics including Mean Absolute Error (MAE) — the average difference between the predicted and true values across the test set. Lower MAE indicates better accuracy.

​
Testing the Model

​
Model testing

Go to Model Testing and click Classify all. Edge Impulse runs every image in the Test set through the trained model and displays the predicted value alongside the true label.
Review the results to check whether predictions are within an acceptable tolerance for your application. If accuracy is lower than expected, consider:
  • Collecting more images, especially near values where the model struggles.
  • Increasing image resolution.
  • Rebalancing the Training/Test split.
  • Using the EON Tuner to automatically find better impulse configurations.

​
Live classification

Connect a device or upload an image directly in the Live classification tab to test the model against new, unseen images in real time.

​
Deploying the Model

Once you are satisfied with accuracy, go to Deployment to export the model for your target hardware. Edge Impulse supports a wide range of embedded targets — see the full list of supported hardware. The regression block produces a single floating-point output per inference, which your application can read and act on directly (for example, triggering an alert when the predicted fill level drops below a threshold).

​
Next Steps