image classification
/ IM-ij klas-uh-fih-KAY-shun /
Image classification asks a machine one blunt question about a picture: what is this? You hand it a photo, and it returns a single label from a fixed list — "cat," "golden retriever," "stop sign." It does not say where the cat is, or how many there are, or what the cat is doing. It just picks the best name for the whole image. This is the simplest and oldest task in computer vision, and most other vision tasks are built on top of it.
Under the hood, the machine never sees a "cat" — it sees a grid of brightness numbers, three of them per pixel for red, green, and blue. Training means showing it hundreds of thousands of labeled photos so it gradually learns which patterns of edges, textures, and shapes tend to go with which label. The output is usually not a flat yes-or-no but a list of confidences: 92% cat, 5% fox, 3% dog. We take the top one as the answer, but that 92% is a statistical bet, not a statement of understanding.
What matters here is honesty about the failure modes. A classifier is only as good as the categories and examples it was trained on. Show it a coffee mug when its list only has animals, and it will confidently call it the nearest animal. It can be fooled by lighting, angle, or a sticker, and it often latches onto background shortcuts — a model that learned "cow" from photos in green fields may stumble on a cow on a beach. High accuracy on a benchmark is not the same as reliable judgment in the wild.
You upload a photo of your tabby. The classifier returns: cat 0.94, lynx 0.03, fox 0.02, other 0.01. The app shows "cat." Crop the same photo to just the carpet, and it might now guess "doormat 0.40" — it has no cat to see and still must pick something.
One label for the whole image, given as confidences. A classifier always picks something, even when nothing fits.
A classifier outputs probabilities over a closed list of categories that must sum to one, so it can never answer "none of these." Anything truly outside its training list gets squeezed into the nearest known label — confidently and wrongly. Recognizing such out-of-list inputs is a separate, harder problem.