Image classification is the basic computer-vision task in which the whole picture is assigned one or more labels from a predefined set of classes. The input is a grid of pixels, the output a probability distribution over classes produced by the softmax function; a distinction is drawn between single-label classification and the multi-label task, where independent sigmoids are used instead of softmax. The core difficulty is that a visually identical object produces completely different pixel values depending on lighting, angle, scale, partial occlusion or background, so a representation invariant to those changes must be learned. Historically, hand-designed descriptors such as SIFT and HOG were used with an SVM-style classifier; the decisive break came in 2012 with AlexNet on the ImageNet dataset, after which convolutional networks took over the field completely and, in the form of VGG, ResNet and EfficientNet, gradually reached superhuman accuracy. Today they are being displaced by the Vision Transformer. In practice the task is almost never solved from scratch, but through transfer learning from a pre-trained model together with data augmentation.
It is the simplest question you can ask of a picture: “what is this?” The answer is a single word – cat, invoice, defective part. You ask nothing more, so neither where it is nor how many of them are there. It is hard because the same cat backlit, seen from the side, and half hidden behind an armchair looks to a computer like three completely unrelated clouds of coloured dots. A human does not see it that way, because the brain worked out long ago that “catness” does not live in the particular dots, but in the shapes and the relationships between them.