Glossary
What is a convolutional neural network?
In short
A convolutional neural network (CNN) is a neural network for images and signals that scans a picture with small windows and builds whole objects from small details.
Machine learning · convolutional neural network
MMOOZZGG.com editorsChecked: 1 min read
For a child
Imagine you look at a big painting through a small window cut in a sheet of paper. You move the window and notice: here is an ear, there are whiskers, next to them a tail. Then you put the findings together and see that it is a cat. Such a program looks at a photo the same way, piece by piece, and then builds the whole picture from the pieces.
In plain words
For a computer, a photo is a table of numbers, one number for each point and color. A convolutional network lays small template filters over it and checks where the image has a vertical line, a corner or a spot. The next layers look for combinations of these findings: an eye, a wheel, a letter. The same filter works anywhere in the frame, so a cat is recognized both in the corner and in the center. This is how face unlock on a phone, license plate reading and defect search in photos work.
For business
Most industrial computer vision runs on these networks: defect control on a conveyor, counting goods on shelves, reading scans and passports, analysis of medical images. They are usually more compact than transformer models and run on low-cost graphics cards and even on cameras with a built-in processor, which matters for a workshop without a good connection. Training from scratch is usually not needed: you take a pretrained model and fine-tune it on your own images. Costs depend on the number of defect types, the volume of labeling and the shooting conditions. The main risk is a change of light, camera or product: accuracy drops without anyone noticing, so teams monitor it by MLOps rules. The work is first checked on a real line and only then scaled.
For a specialist
A convolutional neural network (CNN) is an architecture in which the main layers compute a convolution of the input with small trainable kernels. Its properties: local receptive fields, weight sharing between positions and equivariance to shifts; pooling or strided convolution lowers the resolution and widens the context. Depth gives a hierarchy of features from edges to parts and objects. AlexNet (Krizhevsky, Sutskever, Hinton, NIPS 2012), with 60 million parameters and 500,000 neurons, showed the breakthrough in image recognition. Residual connections in ResNet (He et al., 2015) made it possible to train networks up to 152 layers deep; an ensemble of such networks reached 3.57% error on the ImageNet test and first place in ILSVRC 2015. Variants: one-dimensional convolutions for sound and time series, U-Net for segmentation, object detectors. Vision Transformer (Dosovitskiy et al., 2020) showed that a transformer on image patches, with large-scale pretraining, matches a CNN.
Example
A warehouse puts a camera above a belt with boxes. A network fine-tuned on images of damaged packaging marks dented and torn boxes, and a worker sets them aside before they ship to the customer.
Related terms
Questions and answers
Are convolutional networks outdated after transformers?
Sources
- Krizhevsky, Sutskever, Hinton, ImageNet Classification with Deep Convolutional Neural Networks, NIPS 2012papers.nips.cc
- He, Zhang, Ren, Sun, Deep Residual Learning for Image Recognition, 2015arxiv.org
- Dosovitskiy et al., An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale, 2020arxiv.org
Where companies use it
Not sure whether your company needs this?
AI solutions map