Image classification answers one question about a picture: which single label describes it?
The model outputs a score for every class it was trained on, and the highest one wins.
It says nothing about where anything is, which is the job of object detection instead.
Training needs one label per image, so it is the cheapest kind of vision data to collect.
Almost every other vision task starts from a classifier and reuses what it learned to see.
Image classification is the task the whole field was built on, and it is deliberately narrow.
One image goes in and one label comes out, chosen from a fixed list decided before training.
The network ends in a layer with one output per class, and softmax turns those into probabilities.
Training minimises cross-entropy between the predicted probabilities and the single correct label.
Accuracy is the headline metric, though top-five accuracy is still quoted for large label sets.
The famous benchmark is ImageNet, a thousand classes and over a million labelled photographs.
Progress on it drove almost every architecture worth knowing, from AlexNet to ResNet to vision transformers.
The task assumes exactly one thing matters in each picture, which real images often break.
Multi-label classification relaxes that and lets several classes be true at once.
Nothing in the output says where the object is, or how many of them there are.
That information is what object detection and image segmentation add.
Classification remains the cheapest task to collect data for, since one label per image is quick to produce.
It is also the standard pretraining task, because the features learned transfer to almost everything else.
Transfer learning starts from a trained classifier and replaces only the final layer.
So even when you never ship a classifier, you very likely ship what one learned.
One question on this card. Every option explains itself once you pick, right or wrong.
22 concepts in this module, then a graded test with every answer explained.