Image Classification
Getting an AI to look at an image and say what it is — the foundational task of computer vision.
Reviewed July 10, 2026Stable
When not to use it
- When you need to know where, not just what. Classification gives one label per image. If position, count, or multiple objects matter, this is the wrong task.
- On images unlike your training data. Different camera, lighting, angle, or population and accuracy falls off a cliff — quietly.
- For anything safety-critical without a confidence threshold and a human path. A confident wrong label is the dangerous output.
Reach for something else instead
- Object detection when there's more than one thing, or location matters.
- Segmentation when you need exact boundaries — medical imaging, manufacturing defects.
- Pretrained vision-language models for open-ended questions about an image; they need no training set and answer in words.
Further reading
- Krizhevsky, Sutskever & Hinton (2012), ImageNet Classification with Deep Convolutional Neural Networks — AlexNet, the result that started the deep learning era.
- He et al. (2016), Deep Residual Learning for Image Recognition — ResNet, and why depth stopped hurting.
- Recht et al. (2019), Do ImageNet Classifiers Generalize to ImageNet? — accuracy drops on a fresh test set drawn the same way. Read it before trusting a benchmark.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Training on clean stock images and deploying to a phone camera in a warehouse.
- Ignoring class imbalance, so the model learns to always guess the common class.
- Trusting the confidence score as a probability. It's usually poorly calibrated and overconfident.
At a glance
FieldComputer Vision
Core idealabel what's in an image
Launchedmodern deep learning (2012, ImageNet)
Key methodCNNs, now also ViT
DifficultyBeginner → Intermediate
Where this sits
9 concepts come first. Understanding it opens up 7 more.
6Levelsteps in
9Needs firstconcepts
7Opens up2% of 310
1Areastays here
Computed from the prerequisite graph, not assigned. How this works