Computer vision is the field of computing and artificial intelligence concerned with enabling machines to obtain useful information from images, video, and other visual data. Its objectives include detecting objects, recognising patterns, identifying structures, estimating geometry, tracking movement, understanding scenes, reconstructing three-dimensional information, and generating or transforming visual content. Traditional computer vision used manually designed image-processing operations such as edge detection, thresholding, filtering, feature extraction, geometric transformations, and feature descriptors.
Modern computer vision relies heavily on machine learning, particularly deep neural networks, which can learn visual representations directly from data. Inputs may come from conventional cameras, medical imaging devices, satellites, industrial sensors, depth cameras, LiDAR systems, microscopes, or other imaging equipment. Computer vision is therefore broader than image classification: classification is only one of many visual-computing tasks.
Digital images are numerical representations of visual information. A conventional colour image can be represented as an array of pixels, with each pixel containing values for channels such as red, green, and blue. Grayscale images contain a single intensity value per pixel, while other representations use colour spaces such as HSV or Lab. Images can also contain alpha channels, depth information, infrared measurements, multispectral bands, or other sensor-derived values.
Resolution describes the spatial dimensions of the image, while bit depth determines how many discrete values can represent each channel. Computer-vision algorithms process these numerical structures using operations such as convolution, filtering, geometric transformations, feature extraction, and learned representations. The physical scene itself is not identical to the image: lighting, viewpoint, lens properties, occlusion, sensor noise, motion blur, compression, and exposure all influence the captured data.
Classical computer vision often constructed explicit features from images before applying statistical models or decision procedures. Edge detectors identify rapid intensity changes, which can indicate object boundaries. Corners and local features can provide stable points for matching across images. Image segmentation divides an image into regions, while morphological operations manipulate binary or structured image regions. Feature descriptors such as SIFT and related techniques were historically important for image matching and recognition.
Geometric methods can estimate camera motion, homographies, stereo depth, or three-dimensional structure from multiple views. These approaches remain useful because they can provide precise, interpretable transformations and can work effectively in constrained environments. Deep learning has reduced the need for manually engineered features in many tasks, but classical image-processing and geometric methods continue to be integrated into modern systems.
Convolutional neural networks became highly influential in computer vision because convolution can exploit local spatial structure and parameter sharing. A convolutional layer applies learned filters across an image, allowing the network to detect patterns such as edges, textures, shapes, and increasingly complex structures across successive layers. Pooling or strided operations can reduce spatial resolution while increasing receptive fields. More recent architectures include vision transformers, which apply attention mechanisms to visual representations, and hybrid systems combining convolution and attention.
Object detection models locate and classify objects, usually producing bounding boxes or other spatial representations. Semantic segmentation assigns a class to each pixel or region, while instance segmentation distinguishes individual object instances. Image classification assigns one or more labels to an image or visual region. These tasks require different output structures and evaluation metrics even though they all involve visual input.
Computer vision also involves three-dimensional perception and temporal reasoning. Stereo vision estimates depth by comparing images from different viewpoints, while monocular depth estimation attempts to infer depth from a single image using learned or geometric cues. Optical flow estimates apparent pixel motion between frames. Object tracking maintains identities or trajectories across video frames. Simultaneous localisation and mapping, or SLAM, estimates an observer's position while constructing a representation of an environment and is important in robotics and augmented reality.
Three-dimensional reconstruction can combine multiple images, depth sensors, or point clouds to estimate scene geometry. Video understanding can involve action recognition, event detection, temporal segmentation, and spatiotemporal prediction. These problems are harder than static image classification because the system must reason about geometry, motion, time, viewpoint, and changing visual context.
Computer vision has extensive practical applications. Industrial systems use cameras for quality inspection, defect detection, measurement, assembly verification, and robot guidance. Healthcare systems can analyse medical images such as radiographs, CT scans, MRI images, pathology slides, and retinal photographs, although clinical deployment requires rigorous validation and appropriate professional oversight. Autonomous systems use vision for perception of lanes, objects, pedestrians, signs, and environmental structure. Retail systems can use visual recognition for inventory and checkout processes.
Security systems can analyse video or access-control imagery, raising significant privacy and governance considerations. Agriculture can use aerial and ground imagery to monitor crops and identify disease or stress. Scientific research uses computer vision to analyse microscopy, astronomy imagery, environmental observations, and experimental data. The suitability of a vision system depends heavily on sensor characteristics, lighting, environmental conditions, dataset coverage, and the cost of false positives and false negatives.
Vision models face challenges including occlusion, viewpoint variation, illumination changes, background clutter, unusual objects, adversarial manipulation, domain shift, and limited training examples. A model can achieve high benchmark accuracy while failing when cameras, environments, populations, or operating conditions differ from its training distribution. Dataset annotations can also be imperfect, particularly for ambiguous boundaries or subjective categories. Object detectors may miss small objects, while segmentation systems can struggle with fine boundaries.
Recognition of faces or other sensitive visual attributes introduces additional privacy, security, and governance concerns. Computer vision systems should therefore be evaluated under representative operating conditions, including variations in weather, lighting, camera quality, motion, and relevant demographic or environmental factors. Confidence scores should not automatically be treated as guaranteed probabilities, and automated visual decisions may require human review when errors have significant consequences.
Computer vision is related to but distinct from image processing, pattern recognition, graphics, and multimodal AI. Image processing primarily transforms or analyses images, such as by denoising or sharpening, without necessarily attempting high-level semantic interpretation. Computer graphics generally generates images or visual scenes from models, whereas computer vision infers information about scenes from visual observations, although modern generative systems blur this distinction. Pattern recognition is a broader concept covering recognition in visual and non-visual data.
Multimodal AI combines visual information with language, audio, or other modalities and can use computer-vision components to convert images into representations that another model can reason over. Computer vision therefore spans low-level signal processing through high-level scene understanding, with modern systems increasingly integrating learned representations, geometry, language, and external knowledge.
