Image Classification

A factory photographs every part coming off the line, and a model sorts each photo into “pass” or “defect”. A plant-identification app tells you a leaf is from an oak. A photo library groups your pictures into beaches, food and pets. Each is image classification: one image in, one label (or a few labels) out, chosen from categories decided in advance.

Classifiers are usually trained with supervised learning on many example images that people have already labeled, the same idea as statistical classification applied to pixels. The modern era of the task started with ImageNet, a benchmark of over a million labeled photos in 1,000 categories. In 2012, a deep convolutional neural network known as AlexNet cut the error rate far below earlier methods, which helped set off the deep learning boom. Today vision transformers and general multimodal models can also classify images, sometimes with zero-shot learning, where you describe the categories in words instead of training on examples of each.

Image classification is the simplest computer vision task to explain, and its limits are easy to see. It says what is in the picture but not where, or how many. Finding and boxing each item is object detection. Recognizing a specific person from their face is facial recognition, which carries its own privacy and accuracy questions.

Why it matters (editorial synthesis): classification is often the first vision project a business tries, because the output plugs straight into a decision such as accept, reject or route. Accuracy depends heavily on how closely your images match the training data: a model trained on clean studio shots can fail on dim warehouse photos. Check results per category, not just overall, since a model can score well on average while missing the rare class you care about most.

No Chain of Thought guest has discussed image classification in depth yet, so this entry relies on the papers and documentation above.