Surface defect detection
Finding scratches, stains and deformations on metal, plastic or textile surfaces.
AI · Vision
Computer Vision Applications
Image processing and computer vision solutions for quality control, counting, recognition and measurement.
In brief
- Imaging setup first, then the model.
- Threshold tuning based on the cost of each error type.
- Results are passed to your existing automation and ERP systems.
Computer vision is a way to analyze images from cameras or sensors and perform tasks continuously and consistently where the human eye tires, slows down or makes mistakes. It has a wide range of applications, such as defect detection in manufacturing, counting in warehouses, recognition in security and crop analysis in agriculture.
The key to success is the right imaging setup before the algorithm: lighting, camera angle, resolution and data variety determine model quality.
In detail
Computer Vision Applications: how do we do it?
How does a computer vision project progress?
First the goal (which defect, at what speed, with what accuracy?) and the imaging environment are defined. Then sample images are collected and labeled, and models are trained. The prototype is tested under real line conditions.
Field conditions (changing light, vibration, product variety) differ from the lab, so a pilot period and model retraining are part of the plan.
Integration and monitoring
Model outputs are passed to the PLC, ERP or monitoring dashboard; actions such as rejecting defective products or alerting operators are triggered automatically.
Model performance is monitored in production; the model is updated when a new product variant appears or conditions change.
Scope
What do we offer?
01
Defect and anomaly detection
Automatic detection of surface, assembly and packaging defects.
02
Object recognition and counting
Counting and classifying products, pallets or parts.
03
Measurement
Image-based methods for measuring size and position.
04
OCR / text recognition
Extracting text from labels, serial numbers and documents.
05
Real-time analysis
Low-latency inference that runs at line speed.
06
Integration
Passing results to the PLC, ERP or dashboard.
Problem → solution
Which problems do we solve?
Problem
Quality control depends on the operator and is inconsistent.
Our solution
We set up automated visual inspection that applies the same criteria to every product.
Problem
Counting and measuring are done by hand.
Our solution
We deliver speed and accuracy with image-based counting and measurement.
Problem
Defective products reach customers.
Our solution
With automated inspection at the end of the line we reduce the risk of missed defects.
Process
How does the process work?
- 01
Discovery
Goal, environment and sample data.
Deliverable: Feasibility note - 02
Data and labeling
Image collection and labeling.
Deliverable: Labeled dataset - 03
Model development
Training and validation.
Deliverable: Validated model - 04
Pilot and integration
Field testing and system connection.
Deliverable: Pilot report
Use cases
How is it used in real life?
Packaging and label inspection
Checking for missing labels, misprints and caps.
Warehouse and pallet counting
Automatic product counting from images.
Things to watch out for
Common mistakes and the right way
Common mistake
Not controlling the lighting
The right way
The right way: make lighting part of the design.
Common mistake
Training with little, uniform data
The right way
The right way: collect data under varied conditions.
Common mistake
Skipping the pilot
The right way
The right way: validate under real line conditions.
Portfolio
Our projects related to this service
AI ApplicationsAutomotive supplier · 2026
AI ApplicationsAutomated Quality Control with Computer Vision on the Production Line
A computer vision system installed at the quality control station of a metal parts manufacturer catches surface defects at line speed.
−%38 Defective parts reaching shipmentView project AI ApplicationsRetail and distribution · 2025
AI ApplicationsStock Optimization with a Demand Forecasting Model
For a multi-warehouse distributor we developed a model that produces weekly demand forecasts from sales history and suggests orders.
−%27 Number of stockoutsView project VR / AR ApplicationsIndustry and training · 2026
VR / AR ApplicationsVR Simulation for Industrial Safety Training
A virtual reality training simulation that lets trainees safely experience hazardous work environments and measures training performance.
+%40 Knowledge retention (example)View projectFAQ
Frequently asked questions
How many images are needed for computer vision?
It depends on the problem; simple cases may need hundreds of examples, complex ones thousands. Data augmentation and correct labeling reduce the need.
How important are lighting and camera conditions?
Very important. Controlled lighting and the right camera choice can affect model success even more than the algorithm.
Does the system work in real time on a live line?
Yes; by optimizing the model for the target speed and hardware, inference can run at line speed.
How is the false alarm rate managed?
With threshold tuning and routing suspicious cases to a human, the balance between false alarms and missed defects is adjusted to the business goal.
What's the difference between image processing and deep learning?
Classical image processing works with rule-based methods such as edges, color and shape. Deep learning recognizes complex patterns by learning from examples. In simple, controlled environments, classical methods may be enough; in variable conditions, deep learning performs better.
What kinds of defects and objects can you detect?
We can detect visually distinguishable conditions such as surface scratches, missing parts, incorrect assembly, label errors, counts, measurements and fill levels. Feasibility is confirmed with a preliminary trial on sample images.
What camera and lighting equipment is needed?
The choice of camera, lens and lighting depends on the size of the defect to be detected and the line speed. The right equipment affects the result as much as the algorithm. We prepare a list of recommended hardware and provide installation support.
How fast does the system make decisions?
It's designed to make decisions in milliseconds depending on line speed. We get the required speed in the analysis phase and choose model size and processing hardware accordingly.
Can the results be transferred to the ERP or production system?
Yes. Detection results can be stored in a database with image evidence and passed to ERP, MES or reporting dashboards. Quality analysis can be performed on the number of defective products and their causes.
How is the model adapted to new product types?
Sample images are collected for the new product, and the model is retrained or fine-tuned with this data. We recommend setting up a data collection and training flow that makes adding new products easy.
Where is the image data stored?
Data is stored on an on-premises server or in a secure cloud environment, as you prefer. Retention period, access rights and KVKK requirements are set in writing at the start of the project.
Explore
Related services
Artificial Intelligence
AI Applications
We build custom AI systems for your organization that turn data into decisions.
Explore AutomationAutomation Systems
We turn repetitive work into smart workflows for greater efficiency and speed.
Explore R&DR&D Services
Original R&D projects built with scientific method; from academic knowledge to technology products.
Explore