Convolutional Neural Network
A convolutional neural network is a neural network built for grid-shaped data — images above all — that slides small learned filters across the input to detect local patterns (edges, textures, then shapes, then objects) and stacks these layers so that detail builds into recognition. It started the deep-learning era (AlexNet, 2012), dominated image recognition for a decade, and still runs wherever speed on small hardware matters, including the document and card scanners in fintech apps.
- Also known as
- CNN, ConvNet, Convolutional network
- Used at
- image classification and detection, optical character recognition, document and ID scanning, face detection and liveness on mobile devices, medical imaging, and, adapted, time-series and audio; in fintech, the card-scanning, cheque-imaging and document-capture models on phones
- Binding
- none for the method; face recognition uses fall under biometric rules
- Typical mechanics
- a convolution layer applies the same small filter (3×3 or 5×5 pixels) across the whole image, producing a map of where the pattern occurs; pooling reduces resolution; dozens of such layers build from edges to objects; a final layer classifies. Sharing filters across positions makes the network small and translation-invariant — a passport number is recognised wherever it sits on the page
- Typical history
- LeNet for postal digits (1998), AlexNet's ImageNet win (2012), VGG, ResNet's skip connections (2015) enabling hundreds of layers, EfficientNet; vision transformers (2020) matched them at scale, and hybrid architectures followed
- Parties
- the researchers (LeCun's group, the ImageNet competitors), the mobile-hardware makers whose neural engines run them, and the vision-AI vendors
How a convolutional neural network works in practice
The first layer's filters learn to respond to edges at various angles; the second combines edges into corners and curves; further layers respond to eyes, digits, the MRZ strip of a passport; the final layers say "this is a UK driving licence, front side, at this angle". Because the same filters run everywhere, the network needs far fewer parameters than a fully connected one and learns from fewer images, which is why CNNs succeeded when data and compute were a fraction of today's. Training is backpropagation over labelled images — millions for a general classifier, thousands for a fine-tuned one that starts from a pre-trained network and adapts (transfer learning).
In fintech the CNN is in the phone. When a customer scans a card, a CNN finds the card in the frame, another reads the digits; when they photograph an ID, a CNN classifies the document, locates the fields, and passes crops to an OCR model; a liveness check may run a CNN on the video frames to detect a screen or a printout. These models must run in milliseconds on a mid-range phone without sending frames to a server, and a compact CNN does that where a transformer would not. On the server, larger models — CNN or transformer — do the forensic work.
For AI M&A the CNN era's companies were the computer-vision startups of 2014–2019, many acquired for their teams and their models; the document-capture and identity vendors built on them are the fintech-adjacent legacy.
Example
A neobank's onboarding requires an ID photo. Its first release sends every frame to a server for analysis; capture takes 12 seconds on a median device, drop-off at that step is 19%, and the server bill is material. The team replaces the server round-trip with an on-device CNN (4 MB, quantised) that detects the document, checks glare, blur and framing in real time, and fires the capture only when a good frame is in view; the server receives one image instead of forty. Capture time falls to 3 seconds, drop-off to 7%, server cost by 90%, and the share of images the forensic pipeline rejects for quality falls from 22% to 4% — the largest single gain in the onboarding funnel that year.
Common mistakes
- Sending frames to a server for work a small on-device CNN does in milliseconds. - Training from scratch when a pre-trained network fine-tuned on a few thousand images would do. - Reading vision transformers' benchmark wins as CNNs' obsolescence on mobile. - Ignoring capture quality; the best classifier fails on a blurred image. - Face-recognition uses without the biometric-law analysis.