Computer Vision
Computer vision is the field of AI that interprets images and video — recognising what is in them, reading text, locating objects, verifying faces, detecting manipulation — from raw pixels rather than from structured input; deep learning made it work (2012 onward) and it now runs in every phone. For a fintech it is document capture and reading, identity verification, card scanning, fraud evidence review, and the defence against generated and manipulated images; most of it is rented from vendors.
- Also known as
- Vision AI, Image recognition, Visual AI, CV
- Used at
- identity verification (document authenticity, face match, liveness, injection detection), document capture and OCR, card and cheque scanning, KYC and KYB evidence review, insurance claims (photo damage assessment), retail and unattended payments (walk-out stores), and fraud (screenshots, receipts, generated images)
- Binding
- biometric rules — face recognition is special-category data under GDPR and its uses are restricted; the EU AI Act prohibits some biometric uses and classes remote biometric identification as high-risk; identity-verification standards (eIDAS, the UK's digital identity framework) set requirements for document and face checks
- Typical models
- convolutional networks for classification, detection and segmentation on-device; vision transformers and CLIP-style encoders for general recognition and image–text tasks; OCR and document-understanding models for text; specialised models for face embedding, liveness and forgery detection; multimodal language models for reading documents and answering about images
- Typical performance
- document classification and OCR above 98% on clean images and far lower on poor captures, which is why capture quality is engineered; face match at false-accept rates below 0.01% with false-reject rates of a few percent; liveness and injection detection in an arms race with generated media, with vendor accuracy on synthetic faces rising from around 60% to above 95% through 2023–2025
- Parties
- the identity-verification and document-AI vendors, the labs whose general models the vendors build on, the device makers whose neural engines run models on phones, the regulators of biometrics, the fraudsters generating what the models must catch, and the fintech teams choosing and monitoring vendors
How computer vision works in practice
A customer opens an account. The app's on-device model finds the ID document in the camera frame, checks focus, glare and framing, and captures a good image; a server-side pipeline classifies the document type, reads the fields with OCR, checks security features and the machine-readable zone, compares the extracted data with the application, and scores the document's authenticity. A selfie is captured with liveness detection — passive signals that a real face is in front of a real camera, an active challenge if the risk warrants — and injection detection that checks the camera feed is genuine; a face-embedding model compares the selfie with the document photo. The whole runs in seconds; 90% of applicants pass automatically; the rest go to a reviewer with the evidence displayed. Every step is a vision model, most rented from a vendor, and the fintech's job is choosing, integrating, monitoring and testing them against its own fraud cases.
The arms race defines the current state. Generated faces, face swaps, printed and screen replays, and injected synthetic video each defeated the previous generation's checks; each was answered by new detection — texture and depth analysis, challenge-response, camera-integrity checks, models trained on generated media — and each answer was tested by the next attack. A fintech's identity stack is therefore a vendor relationship with a review cadence, a red-team programme that tests the flow with current attack tools, and monitoring of pass rates and fraud outcomes by segment. The in-house vision work is where the company's images are unique: its own document types, its merchants' storefronts for KYB, its dispute evidence. For a buyer, an identity or document vendor is diligenced on its detection of current synthetic media and its biometric-law compliance; a fintech is diligenced on its vendor choices and the fraud its flow let through.
For a fintech, computer vision is the technology in its onboarding and its fraud evidence, mostly rented and always under attack; for an acquirer, the question is how recently the stack was tested against what the generators now produce.
Example
A neobank's onboarding passes 92% of applicants automatically with a vendor's document and face checks. A fraud investigation finds 1,200 accounts opened with synthetic faces injected into the capture flow over four months — the vendor's liveness check detected replays and prints but not injected video. Response: the vendor's injection-detection module is enabled (it had been optional), a second vendor's liveness runs in parallel on high-risk segments, a quarterly red-team test with current deepfake tools is instituted, and pass rates are monitored by device type and country. Synthetic-face fraud falls by 97%; automatic pass rate falls to 90% and recovers to 91% after tuning; the two-vendor cost adds €0.20 per verification on the 15% of applications routed to it. The bank's next investor diligence reads the red-team reports.
Common mistakes
- Face match without liveness and injection detection. - Capture flows that accept poor images and then blame the model. - Biometric uses without the special-category-data and AI Act analysis. - Vendors chosen once and never re-tested against current attacks. - Building in-house vision for tasks vendors solve, and renting it for the images only the company has.