Computer vision is the subfield of Artificial intelligence concerned with enabling machines to extract meaningful information from images and video, including recognizing objects, understanding scenes, tracking motion, and, in generative applications, producing new visual content. It draws on Machine learning, signal processing, and geometry, and has been one of the most consistently productive testbeds for Deep learning research.
History
Early computer vision, from the 1960s through the 2000s, relied on hand-engineered features, edge detectors, and geometric models designed by researchers, combined with classical machine-learning classifiers. Progress was slow and performance on real-world images remained limited. The field's defining shift came with the ImageNet dataset, released in 2009, and the annual ILSVRC competition built around it, which gave researchers a large, standardized benchmark for object recognition. In 2012, AlexNet, a Convolutional neural network trained on GPUs, won ILSVRC by a wide margin, an event often cited as the start of the modern deep-learning era, discussed further in the entry on ImageNet 2012 (AlexNet moment). Convolutional networks dominated the field for the following decade.
Core tasks
Classic computer vision tasks include image classification, assigning a label to an image; object detection, locating and labeling multiple objects within an image; semantic and instance segmentation, labeling every pixel by category or object instance; and pose estimation. Video tasks add tracking objects across frames and recognizing actions over time. Each task historically required specialized architectures, though later transformer-based vision models increasingly unified several tasks under shared backbones.
From CNNs to transformers
Convolutional networks remained the dominant architecture through the mid-2010s, with progressively deeper and more efficient designs. Starting around 2020, researchers adapted the Transformer (architecture) architecture, originally developed for Natural language processing, to images by treating an image as a sequence of patches, an approach known as the Vision Transformer. Transformer-based vision models proved competitive with or superior to CNNs at scale and became easier to combine with text, contributing to the rise of Vision-language models such as CLIP, which learns a shared representation for images and captions.
Generative and multimodal vision
Since the early 2020s, computer vision increasingly overlaps with generative modeling: Generative adversarial networks and later Diffusion models made it possible not just to interpret images but to produce them, powering Text-to-image generation and text-to-video systems. Vision-language models embedded in general-purpose systems extended computer vision from narrow recognition tasks toward open-ended visual question answering and reasoning about images as part of multimodal AI.
Applications and concerns
Computer vision underlies facial recognition, medical image analysis, autonomous vehicle perception, industrial quality inspection, and content moderation, among many other uses. Its deployment has drawn scrutiny over Algorithmic bias, particularly documented accuracy disparities in facial recognition across demographic groups, and over surveillance uses that raise civil-liberties concerns, feeding into broader debates about acceptable uses of the technology.