Foundations
Computer VisionCV
Also known as: machine vision
The field of AI that lets computers interpret and understand images and video.
Computer vision is the part of AI that lets machines "see", to make sense of photos and video. It is what recognises faces to unlock your phone, spots a pedestrian for a car's safety system, reads a licence plate, or sorts your photos by who is in them. The computer does not see like we do; it finds patterns in the millions of coloured dots that make up an image.
Computer vision (CV) is the field enabling computers to extract meaning from visual data: images and video. Typical tasks include classification (what is in this picture?), object detection (where are the objects?), segmentation (which pixels belong to what?), face recognition and optical character recognition (reading text in images).
The deep-learning leap
CV was transformed in 2012 when a deep convolutional neural network dramatically won the ImageNet competition. Convolutional networks became the workhorse, and more recently vision transformers have joined them.
Where it meets language
Modern multimodal systems combine vision with language, so you can show an AI a photo and ask questions about it, blurring the old line between computer vision and NLP.
Computer vision develops algorithms to infer semantic and geometric structure from visual signals, spanning recognition (classification, detection, segmentation), reconstruction (depth, pose, SLAM) and generation. The modern era is anchored by deep learning: CNNs (AlexNet 2012 onward) established learned hierarchical features over hand-crafted descriptors, and vision transformers (ViT) subsequently showed attention-based architectures can match or exceed convolutions at scale.
Current directions include self-supervised and contrastive pretraining (learning representations without exhaustive labels), open-vocabulary recognition via image-text alignment (CLIP-style), and integration into multimodal LLMs where a vision encoder feeds visual tokens into a language backbone. Persistent challenges are robustness to distribution shift and adversarial perturbation, 3D and video understanding, data and annotation cost, and, as with all AI, bias and fairness, documented starkly in facial-analysis audits.