Custom Computer Vision for Business
We develop computer vision systems that see, understand, and act. Automated quality control, document processing, real-time video analytics, and object detection for any industry.
Two inspection projects, with results
Systems already running on real lines, for clients under NDA. Public figures below; the demo is in each case.
Machines that see and understand the visual world
Computer vision enables machines to interpret images and video with superhuman precision. From detecting microscopic defects on a production line to counting people in a retail space in real time, computer vision transforms visual data into automated decisions and actionable insights.
Human visual inspection is slow, inconsistent, and prone to fatigue-related errors. An operator can inspect 300-500 parts per hour with an 80% defect detection rate. A computer vision system processes thousands of parts per minute with accuracy above 99%. The difference in product quality, defect cost reduction, and production speed is transformative.
But computer vision goes far beyond industrial inspection. In retail, it analyzes in-store customer behavior to optimize layouts and staffing. In logistics, it reads codes, labels, and documents automatically. In security, it detects anomalies and relevant events in video feeds. In healthcare, it assists in medical image analysis with diagnostic accuracy comparable to specialists.
At Soamee we develop custom computer vision solutions with the most advanced architectures: from YOLO models for real-time detection to deep convolutional networks for high-precision classification. We work with static images, streaming video, and multi-camera feeds, deploying on cloud, edge devices, or on-premise GPUs based on each project's latency and privacy requirements.
Detection accuracy
Real-time latency
Undetected defects
Continuous inspection
Computer vision for your industry
Each solution is trained and optimized for your specific use case, with custom datasets and performance metrics tailored to your requirements.
Quality control
Automated visual inspection on production lines: defect detection, assembly verification, dimensional measurement, and quality classification. Systems that operate at real production speeds with accuracy above 99%. Drastic reduction in false positives and defective products reaching customers.
OCR & document processing
Text and structured information extraction from any document type: invoices, forms, labels, license plates, checks, medical prescriptions. Advanced OCR combined with layout and semantic understanding to process complex documents with tables, signatures, and stamps.
Video analytics
Real-time analysis of video feeds: people counting, behavior detection, heat maps, dwell time, queue detection, and anomalous events. Ideal for retail (layout optimization), transport (flow management), and security (incident detection).
Object detection & classification
Models trained to detect, locate, and classify specific objects in images or video. From products on shelves to industrial components, vehicles, wildlife, or any object relevant to your use case. Training with your data for maximum accuracy in your domain.
Medical imaging
Diagnostic assistance through automated analysis of medical images: X-rays, CT scans, MRIs, histological and dermoscopic images. Early anomaly detection, anatomical structure segmentation, and finding quantification. Always as a support tool for specialists.
Retail analytics
Computer vision for commerce: planogram analysis (facing verification), out-of-stock detection on shelves, customer flow analysis, product recognition in self-checkout, and display monitoring. Visual data that translates into business decisions.
Quality control on the production line
Before training any model we decide how the part is lit and photographed. A defect the camera does not capture cannot be recovered by any algorithm.
Screws and turned parts
- Defects
- Damaged thread, deformed head, out-of-tolerance length, rust.
- Capture
- Backlight on a diffuse light table. Telecentric optics to measure in millimetres.
- What is measured
- Silhouette profile, length, thread crest and metal colour.
93.2% OK/NOK accuracy on 1,536 test photos (demo on a public dataset)
Vehicle bodywork
- Defects
- Dents, scratches, broken lights and glass.
- Capture
- Phone photos or an inspection booth. More resolution and controlled light for small dents.
- What is measured
- Damage located with its size, photo by photo, for claims, rentals or used cars.
92% detection precision on the project; around 90% in its tests
Lacquered surfaces and wood
- Defects
- Dust specks, drips, knocks, missing paint, shade differences.
- Capture
- Raking light for relief. Calibrated lighting and a colour target for shade.
- What is measured
- Relief, coating uniformity and colour difference against a reference.
We start with an on-site feasibility phase.
How we take it to a line
- 01
Feasibility
Photos of real parts on site with several camera and lighting setups.
- 02
Data capture
A capture station on the line collects and labels images of each defect.
- 03
Model
Measurement, classification or detection depending on the defect, evaluated on unseen images.
- 04
Pilot
Inspection at one station with operator alerts and a quality dashboard.
- 05
Line
Synchronised triggering, a PLC signal for rejection and history by batch and shift.
Need computer vision for your business?
Free consultation →Computer vision stack
From visual need to model in production
A rigorous process that ensures accuracy and performance under real operating conditions.
Definition & dataset
We define the objects/defects to detect, collect representative images from the real environment, and create a labeled dataset with precise annotations. Data augmentation to maximize diversity.
Model training
We select the optimal architecture (YOLO, ResNet, EfficientNet, ViT) and train with your dataset. We iterate on hyperparameters until reaching target metrics in precision and recall.
Optimization & deployment
We optimize the model for the target hardware (GPU, edge device, cloud). Quantization, pruning, and conversion to efficient formats (TensorRT, ONNX). Deployment with REST API and monitoring.
Production validation
Exhaustive testing under real conditions: lighting variations, angles, wear. Continuous fine-tuning with new production data. Alerts on performance degradation.
You may also be interested in
Frequently asked questions about computer vision
How many images do I need to train a model?
Can it work in real time on a production line?
Can it be deployed on-premise without internet connection?
What accuracy can be expected?
How much does a computer vision project cost?
Give intelligent vision to your business
We help you identify where computer vision can generate the greatest impact on your operations and develop a custom solution that works under real conditions.
Tell us your challenge. We'll propose a solution.
No commitment. We reply within 24 hours, and within 48 hours you have a proposal with scope, timeline and budget. No fine print.