image classification

Image classification is the canonical computer-vision task of assigning a whole image to exactly one category, or in the multi-label variant to a set of categories, drawn from a fixed predefined list of classes. The input is a single image, that is a grid of pixels usually with three colour channels, and the output is a label such as 'cat', 'fire truck', or 'melanoma'. Crucially, classification answers the question 'what is in this image?' at the level of the entire frame; it does not say where the object is, which is detection, nor which pixels belong to it, which is segmentation.

Formally we learn a function f that maps an image x to a vector of scores over C classes, and a decision rule, almost always argmax, picks the predicted class. Training uses a labelled dataset of (image, class) pairs and adjusts the model parameters so the predicted class matches the ground-truth label as often as possible. Modern systems implement f as a deep neural network, historically a convolutional network such as ResNet and now frequently a Vision Transformer, ending in a linear classification head that produces one score per class.

Image classification matters far beyond its own applications because it is the workhorse pretraining task for visual representation learning: a network trained to classify ImageNet learns general-purpose features that transfer to detection, segmentation, retrieval, and more. It is also the simplest setting in which to study almost every deep-learning idea, including optimisation, regularisation, calibration, robustness, and fairness, which is why so many terms in this field are framed around it.