Image classification is the task of looking at a whole image and assigning it one label from a fixed list: “cat”, “pneumonia”, “defective part”. It is the simplest core task in computer vision, and for about a decade it was the benchmark that drove the field forward.
It is a kind of classification where the input happens to be a grid of pixels. A 224×224 colour image is about 150,000 numbers, and the model has to map all of them to a single category while ignoring lighting, angle, background and the thousand ways a cat can sit.
How a classifier works
Reading the diagram left to right: the image goes into a feature extractor, which turns raw pixels into a compact description of what’s in the picture. For years that meant a convolutional network; increasingly it’s a vision transformer, which cuts the image into patches and treats them like tokens. A final layer produces one score per class, and a softmax turns the scores into probabilities. The top one is the answer.
Everything spatial is thrown away at the end. The classifier says the image contains a cat, not where it is or how many there are. Those questions belong to object detection and semantic segmentation, which usually start from a backbone first trained on classification.
The ImageNet years
The field’s yardstick was the annual challenge built on ImageNet, a dataset introduced in 2009; the challenge version has 1,000 categories and about 1.4 million labelled images. The organizers’ 2014 retrospective records top-5 classification error (the right answer missing from the model’s five guesses) falling from 28.2% in 2010 to 16.4% in 2012, when a deep CNN, AlexNet, won decisively, and to 6.7% by 2014. One trained human annotator, tested on a sample, scored 5.1%. In 2015 ResNet reached 3.57% with networks up to 152 layers deep.
The catch
High benchmark accuracy doesn’t guarantee the model learned what you meant. A classifier can latch onto shortcuts, such as a background or a watermark that happens to co-occur with one class, and fail when that cue changes. It is also forced to pick from its fixed label list, so an image of something unfamiliar still gets a confident-looking answer. In practice most teams don’t train from scratch: they take a network pre-trained on a large dataset and fine-tune it on their own labels, a form of transfer learning that needs far less data than starting over.