Ghost images refer to two very different things depending on context. In optics and physics, ghost imaging is a real and actively developed technique that reconstructs a picture of an object using light that never directly interacted with it. In everyday photography, medical scanning, radar, and even human vision, “ghost image” describes an unwanted duplicate or false copy of something in the scene. Both uses share a common thread: an image appearing where you would not expect one. The physics version is a genuinely strange phenomenon that has evolved from a quantum curiosity into a practical tool, while the artifact version is a persistent headache across dozens of technologies.
How Ghost Imaging Works
Ghost imaging splits light into two paths. One beam illuminates the object but is collected by a simple single-pixel detector that records only total brightness, with no spatial detail at all. The second beam never touches the object but is captured by a camera that can resolve spatial patterns. Neither detector alone produces a useful picture. The image emerges only when the intensity fluctuations recorded by both detectors are mathematically correlated. Because the spatial information comes from the beam that avoided the object, the result feels ghostly: an image built from light the object never saw.
The earliest demonstrations in the mid-1990s relied on pairs of quantum-entangled photons produced by a process called spontaneous parametric down-conversion. One photon from each pair hit the object, the other went to the spatially resolving camera, and the quantum link between them made the correlation possible. For years, many physicists assumed entanglement was essential to the trick. That assumption turned out to be wrong.
You Do Not Need Quantum Entanglement
In 2004, researchers showed that splitting ordinary thermal light from an incoherent source into two beams could produce the same ghost-imaging result, because the classical intensity correlations between the split beams were formally analogous to the quantum case.1PubMed. Ghost imaging with thermal light: comparing entanglement and classical correlation This was a significant finding. It meant ghost imaging did not require expensive single-photon sources or cryogenic detectors. A hot lamp and a beam splitter could, in principle, do the job.
The field quickly expanded into three broad illumination categories: thermal-state light (classical), biphoton-state light (quantum), and classical phase-sensitive light.2Advances in Optics and Photonics. Ghost imaging: from quantum to classical to computational The quantum version still holds advantages in certain low-light scenarios, but the classical version made ghost imaging far more accessible and opened the door to a third approach that dispensed with beam splitting entirely.
Computational Ghost Imaging
Computational ghost imaging, or CGI, replaces the reference beam with math. Instead of splitting physical light, a device called a spatial light modulator projects a sequence of known random patterns onto the object. A single-pixel detector records how much total light bounces back from each pattern. Because the pattern shapes are already stored in a computer, the system can correlate the known patterns with the measured intensities and reconstruct the image without a camera at all.
The bottleneck has been speed. Each pattern projection is one measurement, and a full image at decent resolution can require thousands of them. One team addressed this by sweeping a pair of galvanic mirrors across a high-resolution spatial light modulator, effectively multiplying the rate at which patterns could be projected. Their proof-of-concept setup achieved ghost imaging at 42 frames per second for an 80 × 80 pixel image, roughly five times faster than the previous state of the art, with room for further improvement through better hardware.3Scientific Reports. High Speed Computational Ghost Imaging via Spatial Sweeping
Speed can also be gained by being smarter about which patterns to project. An adaptive compressive sampling approach borrows from the mathematics of compressive sensing: rather than projecting every possible pattern, the system learns on the fly which patterns carry the most information and skips the rest. This lets it reconstruct an image from far fewer measurements than the number of pixels in the final picture, and the result is available immediately with no heavy computation after the fact.4PubMed Central. Compressive adaptive computational ghost imaging
Seeing Through Fog and Turbulence
One of the most appealing aspects of ghost imaging is its resilience in harsh conditions. Conventional cameras struggle when the air between them and the object is full of scattering particles, like fog, smoke, or atmospheric turbulence. Ghost imaging has an inherent advantage here because the correlations that form the image are statistical: random scattering degrades individual measurements but does not easily destroy the correlation pattern across many measurements.
Experiments with computational ghost imaging through artificial fog confirmed that image quality held up well over short distances and degraded only gradually as the fog-filled path grew longer. Switching from a standard Gaussian-shaped light source to one with a Lorentzian profile helped maintain quality at longer ranges.5Optics & Laser Technology. Demonstration of computational ghost imaging through fog Atmospheric turbulence poses a similar challenge, and adaptive optics systems have been combined with ghost imaging to recover usable images even in conditions where conventional ghost imaging failed entirely.6Optics Express. Adaptive optical ghost imaging through atmospheric turbulence
This robustness makes ghost imaging attractive for remote sensing and surveillance where the environment between the sensor and the target is unpredictable.
X-Ray Ghost Imaging and Radiation Dose
Perhaps the most medically exciting application is ghost imaging with X-rays. A 2016 demonstration showed that the ghost-imaging principle works with hard X-rays, raising the prospect of imaging biological tissue or industrial materials with dramatically less radiation exposure than conventional X-ray methods.7PubMed. Experimental X-Ray Ghost Imaging
That promise has started to materialize. A more recent experiment using a specially designed X-ray beam splitter estimated that the radiation dose received by the sample in their ghost-imaging setup was about 0.33% of the dose used in conventional projection imaging.8Journal of Synchrotron Radiation. X-ray ghost imaging with a specially developed beam splitter If that kind of reduction translates to clinical or industrial settings, it could make a genuine difference in medical diagnostics and in studying delicate materials that degrade quickly under X-ray exposure.
Looking Around Corners and Detecting Threats
Ghost imaging also enables tricks that sound like science fiction. Because the single-pixel detector does not need a direct line of sight to the object, researchers have demonstrated imaging around corners. The setup collects light that bounces off a nearby wall and correlates it with the known illumination patterns to reconstruct what is hidden from view.9Optik. Imaging around corners with single-pixel detector by computational ghost imaging The images are grainy, but they clearly resolve the shape of the concealed object.
On the detection side, quantum ghost spectroscopy uses pairs of photons at different wavelengths to identify substances in wavelength ranges where detectors are expensive or unreliable. A recent study demonstrated that correlated photon pairs allowed spectral features in the near-infrared to be measured using ordinary visible-light detectors.10The European Physical Journal Plus. Near infrared quantum ghost spectroscopy for threats detection The idea is that you illuminate the substance with one wavelength, detect the correlated photon at a different wavelength where your detector works best, and still extract spectral information about the substance. Potential applications include screening for explosives or hazardous chemicals without needing specialized infrared sensors.
Improving Resolution With Machine Learning
Ghost images tend to be noisy and low-resolution compared to what a good camera can produce directly. Researchers have increasingly turned to machine learning to close that gap. One approach combined a quantum ghost imaging setup with neural networks trained to enhance resolution. The system achieved super-resolved images with fidelity close to 90% while using only the number of measurements needed for a lower-resolution picture.11PubMed Central. Super-resolved quantum ghost imaging In other words, the neural network learned to fill in the gaps, producing a sharper image than the raw data alone would support. This is a growing area, and it reflects a broader trend in imaging science where deep learning compensates for physical limitations in the optical system.
Ghost Images as Unwanted Artifacts
Outside the laboratory, “ghost image” more commonly describes something you do not want: a duplicate, echo, or phantom copy of a real object appearing where it should not. The causes vary enormously depending on the technology involved.
In photography, ghost images often arise from internal reflections and scattering inside the camera lens. Light bouncing between lens elements creates faint secondary images of bright sources, a problem closely related to lens flare. The physics involves reflections, scattering, diffraction, and dispersion within the optical system, and the severity depends on lens coatings, element count, and the angle of incoming light.12arXiv. Toward Flare-Free Images: A Survey Modern lens coatings have reduced these artifacts substantially, but they are almost impossible to eliminate completely, especially when shooting toward a strong light source. Deep learning methods are now being developed to remove lens flare and its associated ghost artifacts after the fact, including self-supervised networks that can disentangle different flare types in nighttime photography without requiring manually labeled training data.13Proceedings of the AAAI Conference on Artificial Intelligence. Disentangle Nighttime Lens Flares: Self-supervised Generation-based Lens Flare Removal
Ghost Targets in Radar
Radar systems face their own version of the ghost-image problem. When a radar signal bounces off one object, then off a wall or another surface before returning to the receiver, the system can interpret the delayed, indirect echo as a second target at a different location. These ghost targets are a serious concern for automotive radar, where self-driving systems need to distinguish real obstacles from phantom ones. In complex environments with many reflective surfaces, indirect paths from one object can corrupt the angle estimation of another, generating false detections.14arXiv. Detection of Ghost Targets for Automotive Radar in the Presence of Multipath
The problem is even more acute for through-wall radar, which is used in search-and-rescue and security applications to detect people behind walls. Interior walls inside a room create specular reflections that produce multipath ghosts looking just like real human targets, increasing false alarm rates.15IET Radar, Sonar & Navigation. Multipath ghost elimination for through‐wall radar imaging Specialized algorithms exist to identify and suppress these phantoms, but it remains an active area of research, especially as radar proliferates in consumer vehicles and urban infrastructure.
Ghost Images in MRI and Human Vision
In medical imaging, ghost artifacts in MRI scans are among the most common quality problems. When a patient moves during a scan, even just breathing or having a heartbeat, the motion gets encoded into the data in ways that produce repeated copies of anatomical structures displaced along the phase-encoding direction of the image. These ghosts can overlap with the area of clinical interest and obscure real pathology. Over three decades of research have produced a large toolkit of methods to reduce motion artifacts, including breath-hold protocols, cardiac gating, and various post-processing corrections, but no single solution works for every situation.16PubMed Central. Motion artifacts in MRI: A complex problem with many partial solutions
Ghost images can even happen inside your own eye. Irregularities in the cornea’s surface can split incoming light into separate focal paths, creating conditions where a person sees a faint second copy of an object alongside the real one. This is a form of monocular diplopia, meaning the doubling persists even when one eye is closed. For these ghost images to be noticeable, the corneal irregularity has to be large enough to produce a secondary image that is both displaced from the primary image and sufficiently focused and bright to compete with it.17PubMed. Corneal aberrations, monocular diplopia, and ghost images: analysis using corneal topographical data Causes include keratoconus, scarring, and post-surgical irregularities. People experiencing this sometimes struggle to get a diagnosis because standard eye exams may not reveal the underlying corneal shape problem without topographic mapping.
Why Ghost Imaging Is Not Just an Academic Curiosity
The deliberate form of ghost imaging occupies an interesting niche. It will probably never replace conventional cameras for everyday photography, but its unusual properties fill gaps that conventional imaging cannot. A single-pixel detector is far cheaper, smaller, and easier to build at exotic wavelengths than a full imaging sensor. That makes ghost imaging particularly valuable in the terahertz band, the infrared, and with X-rays, where high-resolution detector arrays are prohibitively expensive or simply do not exist yet. The ability to image through scattering media and around corners adds further appeal for defense, industrial inspection, and medical applications.
The convergence with machine learning is accelerating the field. Neural networks compensate for the inherently noisy, low-resolution raw data that ghost imaging produces, making the technique viable at lower measurement counts and faster speeds. Meanwhile, compressive sensing techniques continue to shrink the gap between the number of measurements needed and the resolution achieved. At the same time, a parallel engineering effort is underway across automotive radar, medical MRI, and computational photography to identify and suppress the unwanted ghost images that remain stubbornly common in those domains. Whether you are trying to create ghost images or eliminate them, the underlying challenge is the same: understanding how light, radio waves, or other signals correlate across space and time, and using that understanding to separate real information from noise.