How to Accelerate Image Labeling with NLP & Image-Prompted Custom Coding
Classical image-processing algorithms can generate a first pass of defect labels before any deep-learning model is trained. In ReliVision you describe the defect in plain language, attach a reference image, and the platform generates a custom image-processing node that pre-labels it for you. This guide walks through that workflow end to end, using porosity detection in weld X-ray images as the worked example.
- 4 steps
- Under 5 minutes setup
- 1 node — no separate model to train
- No code
The background
Accelerating image labeling with classical computer vision
Traditional algorithms can quickly generate initial labels of suspected defects before training a deep learning (DL)-based inspection model.
Overall detection workflow
Define
Generate custom node(s) to detect likely porosity candidates.
Refine
Flag low-confidence detections and improve instantly.
Example: ReliVision uses NLP and image-prompted custom coding to generate classical image-processing nodes that label porosity in weld X-ray images, reducing manual labeling effort and creating a better starting dataset for supervised deep learning.
What a custom node is
A custom node is generated from a written description and a reference image, then dropped into the pipeline like any built-in node — there is no separate model to train.
Open the Add Custom Node menu
On the Pipeline Editor page, scroll down in the left menu. Click the Add Custom Node button to create a custom node.
What to do
- Open ReliUI and go to the Pipeline Editor page.
- Scroll to the bottom of the node list in the left-hand menu.
- Click Add Custom Node. The node creation dialog opens.
%20-%20Copy.jpg)
Enter a node name
Name the node for what it detects. In the worked example the node is called Porosity Detection.
What to do
- Enter a descriptive node name — for this example, Porosity Detection.
- Confirm to move on to the description step.
.jpg)
Describe the target defect and add a reference image
Provide a detailed description and upload a reference image to help the model better understand the expected output.
Write the description
Describe the defect the way you would explain it to a new inspector: what it looks like, where it appears, and how it differs from the surrounding material. For porosity in weld X-rays, that means small dark rounded voids inside the weld bead, distinct from the lighter parent metal and from the straight edges of the bead itself.
- State the shape and contrast of the defect — for example, small dark circular voids against a lighter weld bead.
- State where it occurs, so the node is not looking at the whole frame: inside the weld bead only.
- Name what it should not flag — scratches, bead edges, or surface marks that read as dark but are not porosity.
- Give an approximate size range if you have one, in pixels or millimetres.
“I have an image as attached. Create a node to find the white porosity on the image.”
Attach a reference image
Upload an image that shows the expected output rather than a perfect specimen. A frame containing the defect, at the resolution and lighting your line actually produces, tells the generator more than a textbook example does.
.jpg)
Verify before you rely on it
A generated node is a starting point, not a validated inspection rule. Run it across a handful of images you already know the answer for — some with porosity, some without — and check both what it catches and what it wrongly flags before you let it pre-label a dataset.
Add the node to your pipeline
Newly created nodes appear in the left menu. Drag and drop them directly onto the workflow area, then link the node to other relevant blocks.
What to do
- Find the new node under Custom Nodes in the left-hand menu.
- Drag it onto the workflow canvas.
- Connect its input to the block that supplies the image, and its output to whatever consumes the detections — usually the annotation or export block.
- Run the pipeline across your unlabelled set. The node writes its pre-labels as you go.
.jpg)
Porosity detection pipeline
Built end-to-end with ReliVision custom image-processing nodes. In practice the pipeline localizes targets accurately under real inspection conditions.
from a blank pipeline to a working porosity detector.
added to the existing pipeline — no separate model to train.
Refinement
Refining low-confidence results
If a result falls short, the pipeline is refined with a sample image.
Flag the result
Mark the missed or low-confidence detection.
Feed it back
The example image refines the custom node.
Detection improves
The updated node redeploys instantly across the pipeline.
.jpg)
Before and after refinement on the same weld X-ray image.
Next steps
With the node running, the pre-labelled set becomes the starting point for expert review rather than a blank slate. Reviewers correct and confirm rather than draw every annotation from scratch, and the corrected set becomes the training data for the deep-learning model that will do the production inspection.
Take the pattern to the next defect type
Once one node works, repeat the same description-and-reference-image loop for the next defect class, then use the pre-labelled set as training data for your deep-learning inspection model.
Benefits of NLP and image-prompted data labeling
Faster label creation
Labels likely porosity candidates before manual review.
Safer than manual-only labeling
Provides consistent first-pass labels that can be corrected by experts.
Better DL training data
Produces structured annotations for DL training.
Practical for variable data
Classical algorithms are used for labeling support, while DL remains the approach for robust inspection.
See it on your own images
Request a demo to try image-prompted custom node generation on your own defect images, or download ReliVision and build your first pipeline.
Download ReliVision
