Computer vision · Crack detection
Every crack,on the map.
AI crack detection is the computer vision technique that finds cracks and other surface defects in photos and outlines them pixel by pixel, through segmentation. We build models adapted to your surfaces, from concrete and masonry to asphalt or render, that work through thousands of inspection images and prepare an ordered list of areas for the engineer to check.
Events
- 10:21:44Image 14/120 · north wall
- 10:21:45Crack traced along its length
- 10:21:45Branch marked for review
- 10:22:10The engineer validates in the report
Visual inspection cannot keep up with the number of images
A single inspection with a drone or a camera can produce thousands of photos. Someone has to look at them one by one, find the cracks, note their position and compare them with the previous inspection. The work is slow, tiring and depends heavily on the attention of the person doing it.
- Fine cracks are easily lost among stains, joints, shadows or formwork marks
- Defect positions are noted by hand, and linking them to the drawing or the structural element is difficult
- Comparison with earlier inspections is done from memory or from reports that are hard to put side by side
- The images collected stay in folders, with no index of the defects found in them
Who it is for
- Managers of buildings, industrial halls and multi-storey car parks who carry out periodic structural inspections
- Construction and technical survey firms that document the condition of concrete or masonry elements
- Operators of roads, bridges, tunnels or industrial infrastructure with large surfaces to inspect
- Teams that already fly drones and have more images than time to analyse them
Step by step
How the system finds a crack, step by step
The five stages on the animation follow a single inspection photo: from the raw image of the wall to a crack that is outlined, measured and validated by the engineer.
Events
Events
- 10:21:44Image 14/120 · north wall
Events
- 10:21:44Image 14/120 · north wall
- 10:21:45Crack traced along its length
Events
- 10:21:44Image 14/120 · north wall
- 10:21:45Crack traced along its length
- 10:21:45Branch marked for review
- 10:22:10The engineer validates in the report
Events
- 10:21:44Image 14/120 · north wall
- 10:21:45Crack traced along its length
- 10:21:45Branch marked for review
- 10:22:10The engineer validates in the report
Step 1 of 5
The inspection photo
Everything starts with a clear image of the surface, taken as square-on to the wall as possible and close enough for the crack to be several pixels wide. Even lighting matters a great deal: a side shadow or a reflection can hide a fine crack or draw a false one. That is why we agree the distance, angle and equipment for the photos from the start.
Technical options for crack detection
A trained segmentation model
A neural network learns from labelled images of your surfaces to tell cracks apart from joints, stains and formwork marks. It suits varied surfaces, drone imagery and several defect types, but it needs a set of examples outlined by specialists.
Classical image processing
On uniform surfaces and photos taken under controlled conditions, cracks can also be found with image rules: lifted local contrast, extraction of thin dark structures and filtering by length and width. It starts without any labelling but struggles with complex textures, so we mainly use it as a starting point or to pre-label images.
Single photos or an orthomosaic
Photos can be analysed one by one, with defects linked to the inspected element, or after they have been stitched into an orthomosaic or 3D model so that cracks appear across the whole façade. The second option needs a flight planned for photogrammetry, but makes comparing inspections much simpler.
What you receive at hand-over
- The segmentation model trained on your surfaces, exported for your local server or chosen cloud environment
- The defect classes and priority rules agreed with the engineers, configured and documented
- The processing pipeline: image intake, positioning on the plan, orthomosaic or map, and export to a report, spreadsheet or GIS
- A validation interface in which the engineer confirms, corrects or rejects each flag
- A measurement report on images from your own inspections, covering missed cracks, false flags and where the model gets it wrong
- Documentation, training for your team and, typically, a retraining procedure for when new surfaces come in
No strings attached: tell us what the cameras should see and we'll tell you honestly whether it can be done.
Let's talkWhat the crack detection model can do
- 01
Outlines cracks on the image
Through segmentation, the model marks the pixels that belong to the crack, not just a box around it, so its shape and path are visible.
- 02
Tells defect types apart
It can separate the classes we agree together, for example cracks, spalling, exposed reinforcement or water ingress, where there are examples of each.
- 03
Estimates approximate dimensions
When the images have a known scale, for example the shooting distance or a size reference, the length and width of the crack can be estimated as guidance for the engineer.
- 04
Orders images by priority
Images with likely defects come first in the review list, with the confidence score shown, while those without flags remain available for spot checks.
- 05
Links defects to a location
Using the photos' GPS coordinates, the orthophoto or the label of the inspected element, each flag can be placed on the drawing or on a map.
- 06
Compares inspections over time
Defects flagged in successive inspections of the same element can be put side by side, to see what is new or what appears to have grown.
How it works
From inspection photo to validated defect
The steps marked in yellow are approved by a person.
- Input
- AI
- Action
- Human in the loop
- Result
- 01
Input
Inspection photos
Images taken with a drone, a camera or a vehicle-mounted camera.
- 02
AI
Defect segmentation
The model outlines cracks and other defects from the defined classes.
- 03
Action
Location and measurement
Flags get a position on the drawing and estimated dimensions, where the scale is known.
- 04
Human in the loop
Engineer validation
A specialist confirms, corrects or rejects every flagged defect.
- 05
Result
Inspection report
Validated defects go into a report, a spreadsheet or a GIS system.
What data crack detection needs
- Real photos of your surfaces, taken in the conditions in which inspections will be done: the same distance, the same light, the same type of equipment
- Labelled examples of cracks and defects, but also of features that look like them: joints, stains, shadows, cables, formwork marks
- The criteria engineers use to classify defects and the thresholds that matter to them
- Image metadata (GPS coordinates, altitude, orientation) or a naming system that links each photo to the inspected element
- Drawings, orthophotos or 3D models of the structure, if defects need to be positioned on them
- Processing resources suited to the volume: a local server with a GPU or cloud processing
Possible integrations
- Images from DJI drones and other platforms with EXIF metadata
- Photogrammetry software, for example Pix4D or Agisoft Metashape
- GIS systems, such as QGIS or ArcGIS
- BIM models in IFC format
- Cloud storage, for example Amazon S3 or Azure Blob Storage
- Local servers with GPUs
- Reports exported to PDF and Excel
- REST API and webhooks
- PostgreSQL databases with PostGIS
- Images from DJI drones and other platforms with EXIF metadata
- Photogrammetry software, for example Pix4D or Agisoft Metashape
- GIS systems, such as QGIS or ArcGIS
- BIM models in IFC format
- Cloud storage, for example Amazon S3 or Azure Blob Storage
- Local servers with GPUs
- Reports exported to PDF and Excel
- REST API and webhooks
- PostgreSQL databases with PostGIS
Each integration's compatibility is checked during the analysis phase, for the systems and versions you use.
Control and transparency
What stays with people
- Assessing every flagged defect and deciding whether it is structurally significant
- Conclusions about the safety of the structure and remedial measures, which belong to the engineer or technical expert
- Spot-checking images without flags, to find out what the model misses
- Validating the labels in the training set by people who know the types of defect
Limits and operating conditions
- The model only sees the surface: it cannot tell how deep a crack is, what caused it or whether it affects the strength of the structure
- Very fine cracks may be below what the image can resolve; resolution and shooting distance decide what can be detected
- Joints, damp patches, shadows or vegetation can be mistaken for cracks, especially on surfaces the model has not seen
- The system does not replace a technical survey or the inspections required by regulations; it is a tool that helps the specialist see faster
How we evaluate crack detection quality
We agree the indicators together before launch and track them afterwards. The values depend on the process, the data and the volume.
- Missed cracks
- Cracks marked by engineers on a sample of images but not detected by the model, counted on an independently labelled set
- False flags
- The share of areas marked by the model that turn out, on review, to be joints, stains or shadows
- Outline overlap
- How closely the outline drawn by the model matches the one drawn by a person, measured on the test set with indicators such as IoU
- Stability across surfaces
- Results compared separately by type of material, lighting conditions and camera equipment
How we build a crack detection model, in 4 steps
- 01
We agree the defects and criteria
We discuss with the engineers which types of defect matter, how they are classified and what the final report needs to contain.
- 02
We collect and label images
We start from images from your inspections and outline defects together with the specialists, including the misleading examples.
- 03
We train and test
We adapt a segmentation model, test it on images it has not seen and show openly where it makes mistakes.
- 04
We integrate it into the inspection workflow
We link the results to drawings, GIS or reports and monitor quality over the following inspections, retraining when new surfaces appear.
Use cases
The scenarios below are conceptual examples, designed to illustrate realistic applications. They are not implementations delivered by WEB-AZ.
- Facilities management
Annual inspection of a multi-storey car park
The photos taken on each level are analysed, and cracks in slabs and columns appear on the car park plan, so the engineer's checklist is ready before the site visit.
Conceptual example - Infrastructure
The faces and piers of a bridge
Drone images are worked through automatically, and areas with likely defects are ordered by confidence; the specialist validates each flag and compares it with the previous inspection.
Conceptual example - Industry
The floor of a production hall
A vehicle-mounted camera travels along the aisles, and cracks and floor damage are marked on the hall plan for the maintenance team.
Conceptual example
Useful guides
Before you decide, read
Frequently asked questions
What is AI crack detection?
It is a computer vision technique in which a model analyses photos of surfaces, for example concrete, masonry or asphalt, and outlines the areas where cracks or other defects appear. The result is a list of flags with a position and a confidence score, which an engineer reviews and interprets.
Can AI replace a technical survey?
No. The model only sees what appears on the surface in the image and cannot assess the cause, depth or structural impact of a crack. Conclusions and remedial measures remain with the engineers and technical experts qualified to make them, and the system helps them get through the images faster.
What kind of images are needed?
Clear photos, taken close enough for the cracks of interest to be visible. Resolution, distance, angle and lighting all matter. Before the project, we analyse a few images from your inspections and tell you what can realistically be detected from them.
Can the length and width of a crack be measured?
They can be estimated when the image has a known scale, for example the shooting distance or a size reference in the frame. The values are indicative and are confirmed on site with the usual measuring tools when they matter.
Do you work with drone imagery?
Yes, we can process drone images, including their GPS coordinates, to place defects on a map or an orthophoto. Flights are carried out by you or by an authorised operator, who is responsible for complying with the rules on drone operations.
What happens to the cracks the model does not see?
That is why we explicitly measure missed cracks on a sample checked by people and keep spot checks of images without flags. Missed cases become new examples for retraining the model.
Can’t find the answer? Write to us and we’ll reply directly.
Do you have inspection images waiting to be analysed?
Send us a few real photos of your surfaces and the types of defect you are interested in. We will tell you what can be detected from them and what further data would be needed.
Related solutions
- Industrial visual inspectionCameras on the production line flag parts with visible defects, and the operator confirms every rejection.
- Custom computer vision modelsWe develop computer vision models for your task end to end: data, labelling, training, evaluation, optimisation, deployment and retraining.
- Object detectionModels that recognise the objects that matter to your process in images and pass them on as structured data.