Computer vision app development
Computer vision app development company for camera features that work in the real world
We are a computer vision app development company that builds camera-based features into iOS and Android apps: document and receipt scanning, barcode and label reading, object detection, defect spotting, visual search and photo checks. This page goes one level below our AI app development service and covers the choices that decide whether a vision feature works outside the demo: where the model runs, which toolkit, how privacy is handled and how accuracy is measured.
We judge every feature on your images and your users’ phones, not on a benchmark.
- On-device first where it fits
- LiteRT, ML Kit and Core ML
- Accuracy measured on your data
- Privacy designed in from day one
Computer vision features we build into apps
Most real-world vision work falls into one of these patterns. The first three are often solved with off-the-shelf on-device APIs; the rest usually need a custom or fine-tuned model.
Text and document capture
OCR for receipts, invoices, IDs and forms, with edge detection, perspective correction and field extraction into structured data.
Barcode and code scanning
Fast scanning for retail, inventory, ticketing and logistics, including damaged codes and poor lighting.
Image labelling and visual search
Tagging photos by content, or finding similar products from a picture, using on-device labels or cloud embeddings.
Object detection and counting
Finding and counting items in a frame: stock on shelves, parts on a bench, vehicles in a yard.
Inspection and defect spotting
Flagging damage, wear or missing parts for field service, insurance and manufacturing apps, with a human review step.
Photo quality and compliance checks
Blocking blurry, dark or wrongly framed photos at capture time, so the backend never receives unusable images.
On-device vs cloud computer vision
Where the model runs is the first architecture decision, and it shapes privacy, cost and user experience more than model choice does. Many apps use both: on-device for live camera work, cloud for heavier analysis of a single captured image.
| Factor | On-device | Cloud |
|---|---|---|
| Latency | Works on live camera frames; no network round trip | Seconds per image; not suited to live preview |
| Offline use | Works without a connection | Needs a connection |
| Privacy | Images can stay on the phone | Images leave the device; needs disclosure and a processing agreement |
| Running cost | No per-request fee | Per-image API or hosting cost that grows with usage |
| Model size and power | Limited by app size, memory and battery | Large models and multimodal reasoning possible |
| Updating the model | Ship a new model in an app release or download it | Change on the server without an app release |
| Device spread | Speed varies with phone hardware; test on low-end devices | Same quality on every phone |
For generative or multimodal analysis in the cloud, see our generative AI and LLM integration service.
Toolkits: LiteRT, ML Kit, Core ML and Apple’s Vision stack
ML Kit
ML Kit is Google’s mobile SDK for Android and iOS, and its APIs run on-device: text recognition, face detection, barcode scanning and image labelling among them. When a feature matches one of its APIs, it is usually the fastest and cheapest route, with no model training at all.
LiteRT (formerly TensorFlow Lite)
For custom models we use LiteRT. Google renamed TensorFlow Lite to LiteRT on 4 September 2024. The .tflite model format is unchanged, LiteRT also runs models from PyTorch, JAX and Keras, and apps on the old TFLite packages must move to the LiteRT packages. If your app still uses TFLite, that switch belongs in the next maintenance cycle. Our TensorFlow Lite and ML Kit page covers the technology choice in more detail.
Core ML on iOS
Apple’s Core ML runs models on-device across the CPU, GPU and Neural Engine. We convert custom models to Core ML for iOS when we need the best speed and battery use on iPhone, and use Apple’s Vision framework for common tasks such as text and barcode detection. See Core ML for more.
Apple Foundation Models with images
At WWDC26 Apple added image prompts to its Foundation Models framework, which runs the on-device Apple Intelligence model. That opens up describe-this-photo and question-about-an-image features on supported devices without a cloud call. Because it does not run on every iPhone, we treat it as an enhancement with a fallback, not the only path.
Have a camera feature in mind?
Send us a description and 20 or 30 sample images. We reply within one business day with whether an off-the-shelf API, a custom model or a cloud model fits, and the next step to a fixed quote.
How we evaluate accuracy before launch
A vision feature that scores well on a public dataset can still fail on your users’ blurry, badly lit photos. We measure on your conditions.
Define the error that matters
For a defect check, a missed defect may cost far more than a false alarm. For a receipt scanner, a wrong total is worse than a failed read. We agree which error to minimise and set thresholds for precision and recall accordingly.
Build a test set from real capture
Images taken on the phones, lighting, angles and backgrounds your users will have, labelled by people who know the domain. Kept separate from any training data.
Measure on target devices
Accuracy, latency per frame, memory and battery drain on a low-end, mid-range and flagship phone. A model that is accurate but drops the camera preview to a stutter is not shippable.
Design for uncertainty
Confidence thresholds, “try again” guidance at capture time, and a manual entry or human review path when the model is unsure.
Monitor after release
With user consent, log anonymised failure signals such as low-confidence results and manual corrections, so the model can be improved where it actually struggles.
Privacy and store rules for camera and image data
Camera features handle some of the most sensitive data an app touches: faces, homes, documents and children. Processing on-device is the simplest privacy story, because images never have to leave the phone. When they do, we keep uploads to the minimum needed, strip metadata such as location where it is not required, and set retention limits.
Store rules apply too. Apple’s guideline 5.1.2(i) requires apps to disclose clearly when personal data is shared with third parties, including third-party AI, and to get explicit permission first. Your App Store privacy labels and Google Play data safety form must match what the feature actually does. In the EU, AI Act transparency obligations apply from 2 August 2026. Where children may use the app, the amended US COPPA Rule treats biometrics as personal information. This is general information, not legal advice; your lawyer should confirm what applies to you.
Privacy defaults we apply
- On-device processing unless there is a reason not to
- Camera permission requested in context, with a clear purpose string
- No face recognition or identification without explicit legal review
- Images deleted after processing unless the user saves them
- Third-party AI sharing disclosed and consented in the app
- Provider settings that exclude your images from training
Custom models: data, training and keeping them current
When no off-the-shelf API fits, the cost driver is data, not code. A custom detector needs labelled examples of every class in realistic conditions, and the labelling effort usually exceeds the model training effort. We often start by fine-tuning an existing model on your images rather than training from scratch, then convert it for LiteRT on Android and Core ML on iOS.
If feasibility is still uncertain, for example whether a defect is visible at phone-camera resolution at all, we run a short proof of concept first with agreed pass criteria. For apps that also overlay results on the live scene, our augmented reality app development team handles the AR layer.
Computer vision projects we have delivered
[[Case study: feature, on-device or cloud approach, toolkit, evaluation results on the client’s test set, and business outcome. Confirm with the client before publishing.]]
[[Client testimonial with name, role and company, used with written permission.]]
Computer vision app development company FAQs
Should computer vision run on the phone or in the cloud?
Run it on the phone when you need live camera results, offline use or the strongest privacy, and when an on-device model meets your accuracy target. Use the cloud for heavier analysis of single images, large or multimodal models, or when you need to update the model without an app release. Many apps combine both.
Do we need to train our own model?
Often not. ML Kit’s on-device APIs already handle text recognition, barcode scanning, face detection and image labelling, and Apple’s frameworks cover similar ground on iOS. Custom models are needed for domain-specific objects, defects or classes. In that case, labelled images from real conditions are the main cost, more than the training itself.
What happened to TensorFlow Lite?
Google renamed TensorFlow Lite to LiteRT on 4 September 2024. The .tflite model format did not change, so existing models still work, and LiteRT can also run models built in PyTorch, JAX and Keras. Apps using the old TFLite packages need to switch to the LiteRT packages, which we usually handle during routine maintenance.
How do you measure whether a vision feature is accurate enough?
We agree which error matters most, then build a test set from real images captured on your users’ devices and conditions, kept separate from training data. We measure precision, recall, latency and battery use on low-end and flagship phones, and only ship when the agreed thresholds are met.
Can a computer vision app work offline?
Yes, if the model runs on-device. ML Kit APIs, LiteRT models and Core ML models all run without a connection. The trade-offs are app size, the speed of older phones and the need to ship an app update or a model download to improve the model. We test offline behaviour on real devices.
What privacy rules apply to camera and image features?
It depends on your market and users. Apple requires clear disclosure and explicit permission before sharing personal data with third-party AI. EU AI Act transparency obligations apply from 2 August 2026, and the amended COPPA Rule treats biometrics as personal information. On-device processing reduces exposure. Your lawyer should confirm what applies.
Build a camera feature you can trust
Book a 30-minute call. Bring sample images and describe the decision the feature should support. We follow up within one business day with a recommended approach and evaluation plan.