Computational photography is the use of algorithms, digital signal processing, multiple image captures and computational resources to produce photographs or visual information that cannot be obtained as effectively from a single conventional optical exposure alone. It combines photography with computer vision, image processing, sensor fusion and increasingly machine learning.
A smartphone camera can use computational photography to reduce noise, expand dynamic range, improve detail, simulate shallow depth of field, stabilise video, correct lens distortions, merge information from multiple cameras and improve images captured in low light. The central principle is that the camera system treats the photograph as a computationally reconstructed result rather than merely the direct output of a lens and sensor. This approach is particularly valuable in smartphones because their physical dimensions limit sensor size, lens diameter, optical zoom and other traditional photographic capabilities. Software can compensate for some of these constraints by extracting and combining information from multiple measurements.
Multi-frame image fusion is one of the most important computational-photography techniques. Instead of relying on one exposure, the camera can capture several frames and align them before combining them. Random noise tends to vary between frames, whereas genuine scene structures remain more consistent, so carefully averaging or otherwise fusing multiple aligned measurements can improve signal-to-noise ratio. Alignment is critical because the camera and scene may move between exposures. Algorithms can estimate motion and transform frames so corresponding features overlap.
If a subject moves independently, however, simple averaging can create ghosting or other artefacts. More sophisticated systems detect moving regions and choose or synthesise appropriate information. This approach underlies many smartphone high-dynamic-range and low-light modes, where several exposures provide more information than a single frame could capture.
High-dynamic-range photography uses computational techniques to represent scenes containing a wider brightness range than a single exposure can record without substantial loss. A camera may capture multiple exposures, including shorter exposures that preserve bright highlights and longer exposures that reveal shadow information. Software aligns and combines these frames and applies tone mapping to produce an image suitable for a display or file format.
The final image is not simply the arithmetic combination of exposures because the algorithm must decide how to handle moving objects, noise, colour and local contrast. Computational HDR therefore involves both measurement and interpretation. The result can preserve information that would otherwise be clipped in highlights or buried in noise in shadows, although aggressive processing can produce halos, excessive local contrast or an appearance that differs from natural scene brightness.
Night photography is another major application. In low light, a sensor receives fewer photons, so random photon noise and electronic read noise become more significant relative to the useful signal. A smartphone can compensate by collecting more light over a longer period or by capturing multiple frames and combining them. Multiple frames allow the system to reduce some random noise while preserving consistent details, provided motion can be estimated and managed.
The phone may also apply specialised denoising and tone-mapping algorithms to brighten dark regions. Computational night modes therefore do not literally make the sensor more sensitive than the laws of photon collection permit; instead, they use time, repeated measurements and processing to extract more useful information from the available signal. They remain limited by moving subjects, extremely low light, sensor noise and the physical aperture and sensor characteristics of the camera.
Computational photography also enables synthetic depth effects. Portrait modes can use two or more cameras with different viewpoints to estimate depth through stereo information, or they can use machine-learning models to segment a person from the background. Once a depth or segmentation map is available, software can apply spatially varying blur to approximate the appearance of a camera with a larger aperture and longer focal length. This is sometimes called computational bokeh.
The effect is not identical to true optical depth of field because real optical blur depends on lens geometry, aperture shape, focus distance and three-dimensional scene structure. Algorithms can make mistakes around hair, transparent objects, fine branches or overlapping subjects. Nevertheless, depth estimation combined with image segmentation allows smartphones to create convincing background separation using relatively small physical camera modules.
Computational photography can combine information from multiple physical cameras. A phone may have wide, ultra-wide and telephoto modules with different focal lengths and sensor characteristics. Software can select the most appropriate camera for a requested framing or blend information across modules. Digital zoom can use cropping and image reconstruction, while hybrid zoom may combine optical magnification with computational processing. Cross-camera processing requires calibration because each module has different colour response, perspective, distortion and physical alignment.
Algorithms can correct these differences before combining the data. Some systems can also use one camera to assist focus, depth estimation or detail reconstruction for another. Multi-camera computation therefore turns separate sensors into a coordinated imaging system rather than treating each camera as an entirely independent device.
Machine learning has become an important component of computational photography. Neural networks can be trained to identify objects, estimate depth, classify scenes, reconstruct details, reduce noise or determine how different regions of an image should be processed. A learned model can perform complex transformations that would be difficult to describe through simple hand-written rules. However, machine-learning-based image enhancement can also introduce plausible-looking details that were not directly captured by the sensor.
This distinction is important because computational photography may optimise for perceptual quality rather than strict measurement fidelity. A photograph can look sharper or cleaner while containing algorithmically inferred information. For scientific or forensic applications, preserving raw measurements and documenting processing steps may therefore be more important than maximising subjective visual quality. Consumer smartphone photography generally prioritises attractive, usable images instead.
Computational photography has expanded the practical capabilities of small cameras but does not eliminate optical constraints. Algorithms cannot recover arbitrary information that was never captured, although they can estimate or infer information from patterns and multiple measurements. A larger sensor can collect more photons, a larger lens can admit more light, and a longer physical focal length can provide genuine optical magnification.
Computational methods can partially compensate for these constraints through multi-frame capture, sensor fusion and reconstruction, but they introduce processing time, energy consumption and potential artefacts. The field therefore represents a partnership between optics and computation. Modern smartphone imaging is best understood as a complete computational imaging system in which lenses and sensors acquire physical measurements while algorithms transform those measurements into images optimised for particular purposes, lighting conditions and viewing environments.
