Training images from the wrong place
Stock photos or a clean test bench look nothing like a dim warehouse, a dusty line or a customer’s phone camera. Accuracy measured on them tells you very little.
Computer vision development
Computer Vision Development for Images and Video
Models that inspect, recognise, count and read from camera feeds, photos and scans — defects on a production line, stock on a shelf, damage in a claim photo, findings on a medical image for a clinician to review. We handle the whole route: collecting and labelling data, training and testing the model, and running it on a device on site or in the cloud.
- Tested on images from your own conditions
- Edge or cloud, chosen per site
- Privacy designed in when people are in frame
Why vision projects stall
Vision models learn only the conditions they are shown.
Most computer vision failures trace back to the data: images that do not match the lighting, angles and cameras of the real site, too few examples of the rare cases that matter, or labels nobody agreed on.
Where vision reliably helps
Repetitive visual checks with a clear definition of good and bad, done at a volume people find tiring:
- Surface and assembly inspection on production lines
- Counting and measuring stock, parcels or produce
- Reading labels, meters, plates and serial numbers
- Assessing damage from claim or inspection photos
- Flagging findings on medical images for clinician review
Too few examples of what matters
Defects, damage and unusual findings are rare by definition. Without a plan to collect or simulate them, the model learns the normal case well and misses the one you care about.
Labels nobody agreed on
Where one inspector sees a scratch, another sees a mark. Without a written labelling guide and checks on agreement, the model inherits the inconsistency.
People in frame, no privacy plan
Cameras that capture staff, customers or patients bring UK GDPR obligations. Retrofitting blurring, retention and access controls after launch is slower and riskier than designing them in.
What we build
Vision systems, from camera to decision.
The model is one component. Around it sit the data pipeline, the device it runs on, the screen where a person checks its work, and the loop that makes it better over time.
Visual inspection
Defect, contamination and assembly checks on lines or in the field, with borderline items sent to an operator rather than silently passed or failed.
Manufacturing software
Detection and counting
Find, locate and count items in images or video — stock, parcels, vehicles, produce — and track them across frames where the job needs it.
Discuss this
Label, meter and plate reading
Read labels, serial numbers, meters, barcodes and number plates from photos or live video, checked against your reference data before anything is recorded.
Document processing
Assistive medical imaging
Models that highlight regions of interest on X-rays, scans or clinical photos for a qualified professional to review — support for a decision, never the decision.
Dental imaging case study
Data collection and labelling
Capture plans, labelling guides, annotation tooling and quality checks on the labels themselves — often the largest single part of a vision project.
Discuss this
Edge and cloud deployment
Models compressed to run on cameras and small devices on site, or served in the cloud for photo uploads, with monitoring and a safe way to update them.
Cloud services
How we work
How a vision project moves from pilot to site.
The usual seven stages, with extra weight on data: what the cameras actually see, how it is labelled, and whether results hold up in the real environment.
- 01
Discovery
We review sample footage and photos with your team on video calls, map the camera setup and lighting, and write down exactly what counts as a defect, item or finding.
A visual definition of the task
- 02
Strategy
Pretrained, fine-tuned or custom model; edge device or cloud; how many images to collect and label. Each choice is costed against the accuracy the task needs.
A data plan, deployment route and quote
- 03
UX & architecture
We design the operator or reviewer screen: the image with the finding highlighted, a one-tap confirm or reject, and a clear view of why an item was flagged.
A review screen people will use
- 04
Development
Two-week sprints alternating between data and model: label a batch, train, measure on a held-back test set, find the failure cases, collect more of them.
A model measured on your own images
- 05
Testing
Accuracy by category and condition — lighting, angle, camera — plus speed on the target hardware, and checks for uneven performance across groups where people are in frame.
Results by condition and category
- 06
Launch
Shadow mode first: the model runs alongside your current checks without acting, so you can compare results before it is trusted with anything.
Evidence from real conditions before go-live
- 07
Optimisation & support
Confirmed and rejected findings become new labels, changes in camera or product conditions are monitored, and models are retrained on a schedule you agree.
A model that keeps up with your site
Technology
The vision stack, from sensor to server.
Frameworks and hardware chosen for the task and the site: what must run on a device, what can go to the cloud, and what the budget allows.
Frameworks
PyTorchTensorFlowOpenCVONNX Runtime
Model types
Object detectionSegmentationImage classificationVision-language models
Data and labelling
Annotation toolingLabelling guides and QADataset versioningAugmented and synthetic data
Edge deployment
NVIDIA Jetson devicesModel quantisationOn-device inferenceOffline operation
Cloud and MLOps
AWS SageMakerContainerised inference APIsTraining pipelinesModel monitoring
Privacy
Face and plate blurringProcessing on the deviceRetention limitsAccess logging
Where cameras capture people, UK GDPR applies, and using faces to identify individuals involves special category biometric data with stricter conditions. We design for data minimisation and support your DPIA; the lawful basis and sign-off rest with your DPO or legal adviser.
Use cases
Where UK operations put cameras to work.
Each one pairs a camera or photo source with a narrow question, and a person who acts on the answer.
End-of-line defect checks
Cameras at the end of a line flag surface and assembly faults, operators confirm each flag, and results feed a defects dashboard by shift.
Parcel and pallet counting
Count, measure and read labels on goods as they pass a door or conveyor, and reconcile them against the expected delivery.
Damage photo triage
Assess claim photos for the type and extent of damage and route them to the right handler, who makes the settlement decision.
Image review support
Highlight areas of interest on X-rays or clinical photos so clinicians can review faster, with every finding confirmed by the clinician.
Shelf and layout checks
Photos taken in store compared with the expected layout to spot gaps, wrong facings and missing price labels that need attention.
Meter and asset readings
Read meters, serial plates and asset condition from photos engineers take on site, with low-confidence reads sent back for a retake.
Relevant work
Products we have designed and built.
Dental.AI analyses dental X-rays and intraoral images to highlight possible conditions for dentists to review — assistive computer vision, with the clinician making the call.
All case studiesWhy Techsleight
What working with us is actually like.
No inflated numbers — just how we run projects, and what you can hold us to.
Product engineering, not ticket-taking
We ask what the software is for before we estimate it — and we will tell you when something should not be built, or should be bought instead.
AI where it earns its place
LLM features, retrieval and automation built with evaluation, guardrails and cost controls, and plain software where that is the better answer.
Full-stack under one roof
Design, frontend, backend, mobile, cloud and QA in one team, so nothing falls between suppliers.
UK-focused delivery
UK business hours, estimates in pounds, and a contract with a UK company. Our engineers are based in the UK and India.
Flexible engagement
A fixed-scope project, dedicated developers or a monthly retainer — and you can move between them as the work changes.
You own everything
Code, IP, cloud accounts and documentation are yours from day one. We sign an NDA before discovery if you need one.
Support after launch
We stay on for fixes, upgrades and new features, or hand over cleanly to your in-house team with the documentation to match.
FAQs
Questions we get asked.
Straight answers on scope, cost, timelines and how we work. If yours is not here, ask us directly.
How many images do we need to train a computer vision model?
Often fewer than people expect when starting from a pretrained model, but it depends on how varied your conditions are and how rare the cases you care about. We start with a small labelled set, measure, and then collect more of the specific cases the model gets wrong, rather than labelling everything up front.
Should the model run on the edge or in the cloud?
On the edge when you need fast responses on a line, when connectivity is poor, or when video should not leave the site for privacy reasons. In the cloud when images arrive as uploads, volumes are uneven, or models need heavy compute. Many systems use both: a small model on site, with uncertain cases sent up to a larger model or a person.
Can we use our existing CCTV or cameras?
Sometimes. We review sample footage for resolution, angle, lighting and frame rate early on. If the existing cameras cannot show what the model needs to see, we will say so and suggest what would, before any model is built.
Who labels the training images?
Usually a mix: your experts label a small reference set and agree the rules with us, then the bulk is labelled against that written guide, with agreement checks and expert review of disputed images. For specialist fields such as medical imaging, qualified reviewers need to be involved.
How do you handle images of people under UK GDPR?
By collecting only what the task needs: blurring faces and plates where identity is irrelevant, processing on the device where possible, limiting retention and logging access. Identifying people from their faces involves special category biometric data, and the EU AI Act may also apply if you sell into the EU. We support your DPIA with technical detail; your DPO decides.
How accurate will the model be?
We will not quote a number before testing on your images. In discovery and the first sprints we measure accuracy per category and condition on a held-back test set, show you the failure cases, and agree the threshold at which a finding is acted on or sent to a person.
Can computer vision support medical diagnosis?
As an assistant to a qualified clinician, yes: highlighting regions of interest or prioritising images for review, with the clinician making the diagnosis. Software with a medical purpose may be regulated as a medical device in the UK, so your regulatory adviser should assess that early; we provide the technical detail they need.
What does a computer vision project cost?
Data collection and labelling are often the largest cost, followed by training and deployment. A discovery sprint from £2,000 reviews your images and camera setup and produces a fixed quote for a pilot. Cloud compute, edge hardware and any annotation service fees are separate costs.
What happens when conditions on site change?
New lighting, a new product or a moved camera can reduce accuracy. We monitor confidence and review rates to spot drift, feed operator corrections back as labels, and retrain on a schedule you agree, testing each new model before it replaces the old one.
How do you work with us during a vision project?
Remotely, with engineers in the UK and India working to UK business hours: a shared channel, weekly written updates and a demo every fortnight on your own images. Footage and photos are shared securely, and your experts join short video calls to review labels and failure cases.
Start a project
Have a visual check to automate?
Tell us what needs to be seen, where the images come from and how many there are. You will get questions back, and an honest view on whether computer vision is the right tool.
What happens next
- 1A senior engineer reads your brief within one working day, and replies with questions or a first view.
- 2A 30-minute call to understand the goal, constraints and what good looks like — no sales script.
- 3A written proposal with scope, milestones, team and a GBP estimate you can take to your board.
Techsleight Labs is a trading name of Krapton IT Consultancy.
Explore