{"id":1674,"date":"2026-09-02T17:22:37","date_gmt":"2026-09-02T21:22:37","guid":{"rendered":"https:\/\/ripondey.ca\/?p=1674"},"modified":"2026-09-02T17:23:02","modified_gmt":"2026-09-02T21:23:02","slug":"building-a-quantum-camera-counting-photons-one-at-a-time","status":"publish","type":"post","link":"https:\/\/ripondey.ca\/?p=1674","title":{"rendered":"Building a Quantum Camera: Counting Photons One at a Time"},"content":{"rendered":"\n

<\/p>\n\n\n\n

By Ripon Kumar Dey, PhD \u2014 Principal Scientist, Deepera Inc.<\/em><\/p>\n\n\n\n

“Quantum camera” gets used loosely enough in popular coverage that it’s worth being precise about what the term actually means in the context of the platform we’re building at Deepera. It doesn’t mean the camera exploits quantum entanglement or does anything exotic in the science-fiction sense. It means the camera operates at the fundamental quantum limit of light detection: counting individual photons, one at a time, rather than integrating a continuous flow of light the way a conventional camera sensor does. That distinction sounds subtle. In practice, it’s the difference between a camera that stops producing a usable image once the scene gets dark enough, and one that keeps working long after conventional imaging has run out of light to work with.<\/p>\n\n\n\n

This post walks through why that distinction matters, how you actually build a sensor capable of registering single photons, why doing that alone isn’t enough to make a useful camera, and where the harder engineering problems \u2014 the ones that occupy most of my time \u2014 actually live. As with my earlier post on AR waveguide optics, I want to be upfront about scope: some of the specific performance figures behind our platform are proprietary and I won’t be disclosing them here. What I can talk about, and what I think is more interesting anyway, is the physics, the device architecture, and the fabrication reality of building something like this.<\/p>\n\n\n\n

1. Why imaging runs out of light<\/strong><\/h2>\n\n\n\n

A conventional camera sensor \u2014 the kind in your phone \u2014 works by letting photons strike a photodiode and accumulate charge over an exposure window, then reading out the total accumulated charge as a brightness value per pixel. This works extremely well as long as enough photons arrive during that exposure to produce a charge signal well above the sensor’s electronic noise floor. The problem is that photon arrival is fundamentally a random, statistical process \u2014 it follows Poisson statistics, meaning the uncertainty in how many photons you’ve collected scales with the square root of the number collected. At high light levels this uncertainty is a rounding error. At very low light levels, it dominates: if you’re only expecting a handful of photons per pixel across an entire exposure, the shot noise from that small a sample is often larger than the signal itself, and a conventional sensor’s own read noise compounds the problem further, since it’s added on top of a signal that’s already barely above zero.<\/p>\n\n\n\n

This is why conventional cameras get progressively worse in the dark rather than simply dimmer \u2014 they run into a noise floor they can’t design their way past without either gathering more light (bigger aperture, longer exposure) or fundamentally changing how they detect it. A single-photon sensor takes the second path: instead of trying to accurately measure a small, noisy charge accumulation, it’s built to register the arrival of one photon at a time as a discrete, unambiguous digital event. Get the detection efficient and low-noise enough, and you can extract a usable image from scene illumination that would be indistinguishable from black on a conventional sensor.<\/p>\n\n\n\n

2. The single-photon avalanche diode<\/strong><\/h2>\n\n\n\n

The device that makes this possible is called a single-photon avalanche diode, or SPAD. Structurally, it’s a p-n junction \u2014 the same basic building block behind LEDs and conventional photodiodes \u2014 but operated in a very different regime. Bias it in reverse beyond its breakdown voltage (a condition called Geiger mode, borrowing the name from Geiger counters, which detect radiation through a conceptually similar avalanche mechanism), and the electric field across the depletion region becomes strong enough that a single absorbed photon, which liberates just one electron-hole pair, is enough to trigger a self-sustaining chain reaction: that first carrier accelerates through the field, collides with the crystal lattice, liberates additional carriers, which accelerate and collide further, cascading into a macroscopic, easily detectable current pulse \u2014 routinely a multiplication factor of a hundred thousand to a million, from a single photon.<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

Figure 1. A photon absorbed in the depletion region triggers a self-sustaining avalanche under strong reverse bias, producing a detectable pulse from a single photon.<\/em><\/p>\n\n\n\n

That avalanche has to be deliberately stopped once it’s served its purpose \u2014 left alone, the cascade would continue indefinitely and the diode couldn’t detect a second photon until it wound down on its own. A quenching circuit, either passive (a simple resistor that starves the avalanche of current) or active (a fast electronic circuit that senses the pulse and forcibly resets the bias), cuts the avalanche short and resets the diode to be ready for the next photon, typically within nanoseconds. The output is inherently digital: rather than a graded brightness value, each pixel produces a stream of timestamped “click” events, one per detected photon, and an image is built up by accumulating or statistically processing those clicks rather than reading out an analog voltage.<\/p>\n\n\n\n

The metrics that actually matter<\/em><\/strong><\/p>\n\n\n\n

A handful of device parameters determine whether a SPAD is any good, and it’s worth naming them even without disclosing our specific numbers. Photon detection efficiency (PDE) is the probability that a photon actually hitting the device gets registered at all, as opposed to passing through or recombining without triggering an avalanche \u2014 it depends on the semiconductor material’s absorption at the wavelength of interest and the depletion region’s geometry. Dark count rate (DCR) is the rate at which the device registers a “click” with no photon actually present, driven by thermally generated carriers or crystal defects that trigger a spontaneous avalanche on their own \u2014 every dark count is indistinguishable from a real detection event and directly limits how faint a signal you can pull out of the noise. Timing jitter is the uncertainty in exactly when a detected photon’s timestamp is recorded relative to when it actually arrived, which matters enormously for any application that measures time-of-flight rather than just counting photons. And afterpulsing is a related nuisance effect where carriers trapped during one avalanche release later and trigger a spurious second avalanche, correlated with \u2014 and therefore harder to average away than \u2014 genuine dark counts.<\/p>\n\n\n\n

Every one of these parameters is set by a combination of material choice, junction doping profile, and fabrication quality, and they trade off against each other in ways that are never fully independent \u2014 pushing detection efficiency higher by widening the depletion region, for instance, tends to increase timing jitter, because photons absorbed at different depths in a wider region take measurably different amounts of time to drift to the multiplication zone. Chasing any one metric in isolation without understanding these couplings is a reliable way to end up with a device that looks great on one datasheet line and is unusable in practice.<\/p>\n\n\n\n

3. The fill-factor problem, and why we build metalenses<\/strong><\/h2>\n\n\n\n

Here’s a problem that doesn’t show up until you try to build an actual imaging array rather than a single detector: the avalanche region of a SPAD, along with the guard-ring structure needed around it to prevent premature edge breakdown, takes up real chip area \u2014 area that isn’t photosensitive. Pack SPADs into a dense array the way you would a conventional sensor’s pixels, and the fraction of each pixel’s area that’s actually able to detect a photon (the fill factor) can end up disappointingly small, because so much of each pixel’s footprint is consumed by circuitry rather than active detector. A sensor that’s individually excellent at detecting single photons can still perform poorly as an imaging array if most of the incoming light is landing on dead silicon between the active regions rather than on the detectors themselves.<\/p>\n\n\n\n

This is where my own background in nanophotonic device fabrication intersects directly with the sensor architecture, and it’s the piece of Deepera’s platform I’ve spent the most hands-on time developing. Rather than accepting the native fill factor, we place a metalens directly over each pixel \u2014 a flat optical element built from an array of nanoscale pillars, each individually shaped to impart a specific, precisely controlled phase delay to the light passing through it. Engineer that phase profile correctly across the array and the metalens behaves like a conventional refractive microlens, focusing the full pixel’s worth of incident light down onto the small active detector area, but it does so in a layer only a few hundred nanometers thick, fully compatible with wafer-level semiconductor fabrication rather than requiring separately molded refractive optics bonded on afterward.<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

Figure 3. A metalens concentrates the full pixel’s incident light onto the SPAD’s small active area, recovering fill factor that would otherwise be lost.<\/em><\/p>\n\n\n\n

Designing the nanopillar geometry is an electromagnetic simulation problem \u2014 get the pillar dimensions and spacing right and you get a clean, efficient focal spot; get them wrong and you get scattered light, unwanted diffraction orders, or a focal spot that doesn’t land where the detector actually is. But the harder half of the problem, in my experience, is fabricating that pillar array with the placement accuracy and sidewall quality the simulation assumes. This is electron-beam lithography and etch territory \u2014 the same nanofabrication discipline I wrote about in the context of AR waveguide gratings \u2014 applied to a different optical function but facing very similar tolerance-sensitivity challenges: pillar height variation, sidewall angle, and edge roughness all degrade focusing efficiency in ways that are straightforward to model and genuinely difficult to hold at production yield.<\/p>\n\n\n\n

4. Material and wavelength choices<\/strong><\/h2>\n\n\n\n

One decision that ripples through the entire system is which semiconductor material to build the SPAD in, and at which wavelength the whole platform is optimized to operate. Silicon SPADs are the most mature and manufacturable option, benefiting from decades of CMOS process infrastructure, but silicon’s absorption drops off well before it reaches the short-wave infrared (SWIR), which limits it to visible and near-infrared applications. Compound semiconductors \u2014 indium gallium arsenide (InGaAs) paired with an indium phosphide (InP) substrate being the most common combination for SWIR work \u2014 extend usable detection out into wavelength bands, such as the 1310\u20131550 nm telecom band, that silicon simply cannot reach.<\/p>\n\n\n\n

That extended reach matters for reasons beyond raw material capability. Wavelengths in the SWIR band are substantially safer for the human eye than visible or near-infrared light at equivalent power, because the eye’s lens and cornea absorb SWIR before it can focus onto and damage the retina \u2014 a genuinely important consideration for any active-illumination system, such as a LiDAR, that has to operate around people. SWIR also benefits from better atmospheric transmission through common obscurants like fog, haze, and smoke than visible light does, which is directly relevant to the wildfire and long-range surveillance applications discussed later. None of this comes for free: compound semiconductor fabrication is less mature, more expensive per wafer, and often requires bonding or integrating a III-V detector layer onto a separate silicon readout circuit rather than fabricating everything monolithically the way a pure silicon sensor can \u2014 which becomes its own substantial fabrication problem, covered next.<\/p>\n\n\n\n

5. From single detector to wafer-scale array<\/strong><\/h2>\n\n\n\n

Everything described so far explains how a single SPAD pixel works. Building a useful camera means replicating that pixel by the thousands or millions across a sensor array, and that scaling step introduces engineering problems that don’t exist at the single-device level at all.<\/p>\n\n\n\n

When the detector material and the readout electronics live in different semiconductor systems \u2014 a III-V detector layer paired with a silicon CMOS readout circuit, for instance, since fabricating high-speed timing electronics directly in InGaAs is impractical \u2014 the two have to be physically joined without damaging either, typically through a process called hybrid bonding: the detector wafer and the readout wafer are each processed independently, then brought together and bonded at the wafer level, with dense arrays of metal interconnects (often through-silicon vias, or TSVs) providing an electrical path between each detector pixel and its corresponding readout circuit below it. Getting this bond uniform and void-free across an entire wafer, with interconnect pitches fine enough to match a dense pixel array, is a serious process-development challenge in its own right \u2014 a misaligned or incomplete bond doesn’t just lose a pixel, it can compromise the hermeticity or electrical integrity of neighboring pixels as well.<\/p>\n\n\n\n

Layered on top of the detector-to-readout bond is the metalens integration discussed earlier \u2014 a third physical layer that has to be aligned to the pixel array beneath it with a placement tolerance that’s a small fraction of the pixel pitch itself, since a metalens that’s well designed but misregistered against its target pixel simply focuses light onto the wrong, non-photosensitive region of silicon. Multiply a single-pixel alignment tolerance across a full wafer containing many thousands of identical pixels, each needing that same tight registration, and you get a sense of why wafer-scale integration is treated as its own discipline rather than an afterthought to device design \u2014 it’s frequently the actual limiting factor on array yield, more so than any single detector’s intrinsic performance.<\/p>\n\n\n\n

6. Characterization: measuring a device that only speaks in clicks<\/strong><\/h2>\n\n\n\n

A conventional image sensor’s performance is straightforward to characterize \u2014 expose it to a known, calibrated light source and measure the resulting signal. A SPAD’s output is a stream of discrete timestamped events, and extracting each of the performance metrics introduced earlier requires its own dedicated measurement methodology.<\/p>\n\n\n\n

Photon detection efficiency is measured by illuminating the device with a calibrated, attenuated source at a known photon flux \u2014 attenuated enough that the true arrival rate is well characterized \u2014 and comparing the detected click rate against the known incident rate, correcting for any residual dark counts and afterpulsing so they don’t get miscounted as detected photons. Dark count rate is measured the more direct way: cover the device completely, eliminate every stray light path into the measurement chamber, and count clicks over a long integration window in complete darkness, since any residual light leakage will inflate the apparent dark count rate and understate true device performance. Timing jitter uses a technique called time-correlated single-photon counting (TCSPC): a pulsed laser with a precisely known, narrow pulse width repeatedly illuminates the detector at low flux, and a histogram of the measured arrival-time delay between the laser trigger and each detected click is built up over millions of events \u2014 the width of that histogram, rather than any single measurement, is the jitter, since it’s fundamentally a statistical property of the detection process, not a fixed offset.<\/p>\n\n\n\n

All three measurements are sensitive to their own experimental artifacts in ways that are easy to get subtly wrong \u2014 stray light contaminating a dark count measurement, an insufficiently characterized attenuator skewing a PDE measurement, timing-electronics jitter in the measurement setup itself getting conflated with jitter intrinsic to the device under test. A meaningful fraction of the value in a mature characterization lab isn’t the measurement technique itself, which is well established in the literature, but the discipline of ruling out every one of these artifacts before trusting a number enough to feed it back into the next design iteration.<\/p>\n\n\n\n

7. A fabrication lesson: when a design that simulates perfectly doesn’t etch that way<\/strong><\/h2>\n\n\n\n

It’s worth grounding all of this in a concrete example, in the same spirit as the checkerboard-defect story from my earlier post on AR waveguide gratings, because the shape of the problem generalizes even when the specific numbers here are kept general rather than exact. Early in metalens development for a SPAD concentrator array, we had a nanopillar design that performed excellently in electromagnetic simulation \u2014 high, uniform focusing efficiency across the array, a clean focal spot well matched to the detector’s active area. The as-fabricated devices, however, showed a systematic falloff in focusing efficiency toward the edges of each imaging die that the simulation gave no reason to expect, since the simulated design was, by construction, spatially uniform.<\/p>\n\n\n\n

The investigation followed the same design-of-experiments discipline as the earlier waveguide example: rather than guessing at a cause, we cross-sectioned pillars from multiple locations across the die and measured actual sidewall angle and pillar height as a function of position, then compared that measured geometry against the simulation’s assumed geometry at each location. The pattern that emerged was a modest but systematic etch-rate gradient across the die \u2014 a well-known phenomenon in plasma etch processes, generally called etch loading or microloading, where local pattern density affects the local concentration of reactive species in the plasma, so pillars in a sparser region of the pattern etch slightly differently than pillars in a denser region. Because our nanopillar phase profile intentionally varies pillar dimensions across the metalens to create the focusing effect in the first place, different regions of the same die had different local pattern densities by design \u2014 which meant the etch-loading effect wasn’t uniform across the part, and neither was the resulting deviation from the intended geometry.<\/p>\n\n\n\n

The fix combined a revised etch recipe with better plasma uniformity control and a pre-compensated design file that intentionally adjusted target pillar dimensions in anticipation of the known loading effect at each position \u2014 closing the loop between what the etch tool actually does and what the mask design assumes it will do, rather than treating the two as independent. As with the waveguide grating story, the broader point isn’t the specific mechanism, which is particular to plasma etch physics and won’t recur identically in every process; it’s that a design which is provably correct in simulation still has to survive contact with a real fabrication process, and the gap between the two is very often where the real engineering work lives.<\/p>\n\n\n\n

8. The data-rate problem nobody mentions at first<\/strong><\/h2>\n\n\n\n

There’s a practical engineering consequence of event-based detection that’s easy to overlook until you’re actually building the system: a SPAD array doesn’t produce a tidy frame of pixel values the way a conventional sensor does, it produces an asynchronous stream of individually timestamped events, potentially from every pixel, arriving whenever a photon happens to arrive rather than on a fixed frame clock. At a large array size and a realistic photon flux, the aggregate data rate coming off the sensor can be substantial \u2014 before any image has even been formed, you’re moving and processing a raw event stream that dwarfs what an equivalent conventional sensor would produce for a comparable field of view.<\/p>\n\n\n\n

Handling that data volume is very much a co-design problem rather than something that can be solved purely downstream. Some of it gets addressed in the pixel and readout architecture itself \u2014 on-chip logic that bins nearby events, applies coincidence detection to suppress likely dark counts before they ever leave the sensor, or compresses the timestamp stream before it hits the digital interface. Some of it is addressed architecturally, by deciding early which applications genuinely need full per-photon timestamp data (time-of-flight ranging, for instance, generally does) versus which can work from a pre-aggregated photon-count image without needing individual event timing preserved (most passive low-light imaging falls here). Getting this wrong in either direction has real costs: over-preserve data fidelity and you build a system that can’t actually move or process its own output fast enough to be useful in real time; over-aggregate too early and you throw away exactly the statistical information the reconstruction model downstream needs to do its job well. It’s a genuinely unglamorous piece of systems engineering, and in my experience it’s one of the pieces that separates a sensor that performs well on a bench from a camera that performs well as a shipped product.<\/p>\n\n\n\n

9. Imaging in the photon-starved regime<\/strong><\/h2>\n\n\n\n

Even with an efficient, low-noise SPAD array with good fill factor, there’s a further wrinkle specific to very low light imaging: at the illumination levels this platform is designed for, you’re often working with well under one photon per pixel across a reasonable exposure window. A raw image built directly from those counts looks like what it is \u2014 a sparse scatter of individual detection events, visually closer to noise than to a photograph, with large stretches of the frame that received zero photons at all purely by chance rather than because nothing was there.<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

Figure 2. At sub-photon-per-pixel illumination, raw single-photon counts are dominated by Poisson shot noise; reconstruction recovers a usable image from that sparse data.<\/em><\/p>\n\n\n\n

This is precisely the problem the AI half of Deepera’s platform is built to solve, and it’s why we describe the system as a dual-innovation platform rather than just a sensor. Because photon arrivals follow known Poisson statistics, and because real-world scenes have strong spatial and temporal structure \u2014 edges, smooth regions, motion that’s continuous rather than random from frame to frame \u2014 a reconstruction model trained on those statistics can distinguish genuine sparse signal from shot noise in a way that a naive frame-average or simple denoising filter cannot. The model isn’t inventing detail that isn’t there; it’s making a statistically informed estimate of the most likely scene given a sparse, noisy sample of photon arrivals, similar in spirit to how modern computational-photography low-light modes work, but operating on true single-photon event data rather than a low-but-nonzero analog signal.<\/p>\n\n\n\n

I’ll say plainly that this is the one area of the platform where I’m not the domain expert \u2014 my own background is on the hardware and fabrication side \u2014 but from where I sit, watching the reconstruction pipeline turn a frame that’s genuinely unreadable to the naked eye into a recognizable scene is the single most visually convincing demonstration of why the hardware and software have to be co-designed rather than developed independently and stitched together afterward. A reconstruction model tuned against the actual noise statistics, dark count behavior, and timing characteristics of our specific sensor performs meaningfully better than a generic low-light denoiser applied after the fact, which is the whole argument for treating this as one integrated platform rather than a sensor project and a software project running in parallel.<\/p>\n\n\n\n

10. Putting it together: the full signal chain<\/strong><\/h2>\n\n\n\n

Assembled end to end, the system looks like this: incident photons first pass through the metalens concentrator array, which focuses each pixel’s collection area down onto the much smaller SPAD active region; each SPAD registers photon arrivals as discrete avalanche events; dedicated timing and readout electronics timestamp and digitize those events with minimal added jitter or dead time; and the resulting sparse, timestamped photon stream is passed to the reconstruction model, which outputs a final image. Every stage has to be engineered against the stages before and after it \u2014 a beautifully efficient metalens is wasted if the SPAD behind it has a high dark count rate that swamps the extra signal, and the best reconstruction model in the world can’t recover information the sensor never captured in the first place.<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

Figure 4. The full signal chain: metalens concentration, single-photon detection, timing electronics, and AI reconstruction, each stage co-designed with the others.<\/em><\/p>\n\n\n\n

The one performance figure I can share concretely, because it’s been cleared for public discussion, is that this combined approach \u2014 concentrator optics, low-noise single-photon detection, and statistics-aware reconstruction working together \u2014 has demonstrated image quality improvements in the range of 10 to 16 dB in peak signal-to-noise ratio (PSNR) compared to conventional low-light imaging approaches under matched, extremely photon-limited conditions. In practical terms, that’s the difference between an unusable, noise-dominated frame and a clearly interpretable image, achieved at light levels where conventional sensors have essentially nothing left to work with.<\/p>\n\n\n\n

11. Why bother: applications that actually need this<\/strong><\/h2>\n\n\n\n

None of this complexity is worth taking on unless it enables something a conventional camera genuinely can’t do, and the honest answer is that most everyday photography doesn’t need single-photon sensitivity \u2014 well-lit scenes are exactly what conventional sensors already handle well. The applications that justify this approach are specifically the ones defined by an absence of light, or where every remaining photon carries real information that can’t be recovered any other way.<\/p>\n\n\n\n

Defense and security surveillance<\/em><\/strong><\/p>\n\n\n\n

Operating in genuinely unlit conditions \u2014 no moon, no illuminator, no active infrared floodlight to give away a sensor’s position \u2014 is a scenario where the only available light is whatever ambient starlight or airglow is present, several orders of magnitude dimmer than the scenes conventional night-vision equipment is rated for. A passive sensor that needs no active illumination is also a sensor that doesn’t broadcast its own position, which matters as much operationally as the raw sensitivity gain does.<\/p>\n\n\n\n

Industrial and infrastructure monitoring<\/em><\/strong><\/p>\n\n\n\n

Inspecting equipment or environments where adding illumination is impractical or unsafe \u2014 confined spaces, hazardous or explosive atmospheres where an active light source introduces its own risk, or long-range monitoring where placing a light source anywhere near the target isn’t physically possible \u2014 favors a sensor that can extract an image from whatever ambient or incidental light is already present rather than one that depends on supplying its own.<\/p>\n\n\n\n

Wildfire detection<\/em><\/strong><\/p>\n\n\n\n

Identifying very faint thermal or optical signatures at long range, often through smoke or haze that further attenuates an already weak signal, is a case where the value of extra sensitivity compounds: a fire’s earliest optical signature is faint by nature, atmospheric scattering degrades it further with distance, and the operational payoff for detecting it minutes earlier \u2014 before it becomes a large, fast-moving front \u2014 is disproportionately large relative to the sensing improvement that bought that time.<\/p>\n\n\n\n

LiDAR and time-of-flight ranging<\/em><\/strong><\/p>\n\n\n\n

Here the timing-jitter properties of the SPAD matter as much as pure sensitivity, since distance is inferred from photon time-of-flight rather than intensity. A detector that’s both single-photon sensitive and precisely timed extends useful ranging distance well beyond what conventional avalanche photodiodes achieve, particularly against low-reflectivity or distant targets that return only a handful of photons per pulse \u2014 exactly the regime where a conventional receiver’s noise floor swamps the return signal entirely.<\/p>\n\n\n\n

Medical and biological imaging<\/em><\/strong><\/p>\n\n\n\n

Several fluorescence and luminescence-based imaging techniques are intrinsically photon-starved by the nature of the biological signal itself, not by any choice of illumination \u2014 the fluorophores or luminescent markers involved simply don’t emit very many photons to begin with. In these applications, more sensitive detection translates directly into either meaningfully better image quality at a fixed excitation dose, or an equivalent image quality achieved at a lower, safer excitation dose \u2014 a genuinely valuable trade in a clinical or biological context where excitation intensity carries its own risk or cost.<\/p>\n\n\n\n

12. The broader landscape<\/strong><\/h2>\n\n\n\n

It’s worth situating this within the wider trajectory of the field rather than treating it as an isolated engineering exercise. Single-photon detection technology has existed in specialized scientific instrumentation \u2014 photon-counting for astronomy, fluorescence lifetime imaging in biology, quantum key distribution research \u2014 for decades, generally at costs and form factors that kept it confined to laboratory settings. What’s changed, and what makes a platform like this viable as a product rather than a research instrument, is the maturation of semiconductor fabrication techniques capable of bringing SPAD arrays, nanophotonic optics, and CMOS readout electronics together at a scale and cost structure closer to conventional image sensor manufacturing than to bespoke scientific instrumentation.<\/p>\n\n\n\n

That maturation is still in progress, not finished. Array sizes are growing, but a large-format single-photon sensor is still a harder manufacturing target than an equivalent conventional CMOS sensor, for exactly the wafer-integration and yield reasons discussed above. Cost per pixel is falling, following a trajectory that looks broadly similar to earlier phases of conventional image sensor development, though it would be premature to claim it will follow an identical curve, since the underlying device physics and fabrication requirements are genuinely more demanding. What I’m confident of is the direction: the applications listed above are already commercially real today at the performance and cost points this technology currently reaches, and the addressable set of applications only grows as array size increases and cost per pixel continues to fall \u2014 the same broad pattern that took specialized imaging technology from laboratory bench to mainstream product in prior sensor generations, playing out again on a different underlying device.<\/p>\n\n\n\n

13. Where this is headed<\/strong><\/h2>\n\n\n\n

The quantum and computational imaging space is still early relative to its ultimate ceiling, and most serious market analyses of single-photon and quantum-enhanced imaging technology project substantial growth over the next decade as the underlying sensor and fabrication costs continue to fall \u2014 a trajectory that mirrors what happened with conventional CMOS image sensors a generation ago, when a technology that started in specialized scientific and defense applications eventually became cheap and reliable enough to end up in every phone. I don’t think single-photon imaging follows an identical path \u2014 the physics keeps some applications inherently specialized \u2014 but I do think the direction of travel is the same: falling cost per pixel, rising array sizes, and a widening set of applications that stop being niche as the hardware matures.<\/p>\n\n\n\n

From where I sit, the constraint on getting there isn’t a lack of clever algorithms or clever optical designs on paper \u2014 it’s the same fabrication reality that runs through everything I’ve written about on this blog so far. A metalens design that performs beautifully in simulation, a SPAD architecture that looks excellent in a device physics paper, a reconstruction model that works well on synthetic data \u2014 none of it matters until it can be manufactured repeatably, at a yield and cost that makes the resulting camera something you can actually build a product around. That’s the work. It’s less glamorous than the headline capability, and it’s most of what actually determines whether a technology like this reaches the applications above or stays a laboratory demonstration.<\/p>\n\n\n\n

Further reading<\/strong><\/h2>\n\n\n\n

Cova, S., Ghioni, M., Lacaita, A., Samori, C., and Zappa, F. “Avalanche photodiodes and quenching circuits for single-photon detection.” Applied Optics 35, no. 12 (1996): 1956\u20131976.<\/p>\n\n\n\n

Bruschini, C., Homulle, H., Antolovic, I. M., Burri, S., and Charbon, E. “Single-photon avalanche diode imagers in biophotonics: review and outlook.” Light: Science & Applications 8, no. 87 (2019).<\/p>\n\n\n\n

Khorasaninejad, M., and Capasso, F. “Metalenses: Versatile multifunctional photonic components.” Science 358, no. 6367 (2017).<\/p>\n\n\n\n

Hadfield, R. H. “Single-photon detectors for optical quantum information applications.” Nature Photonics 3 (2009): 696\u2013705.<\/p>\n\n\n\n

Rapp, J., Tachella, J., Altmann, Y., McLaughlin, S., and Goyal, V. K. “Advances in single-photon lidar for autonomous vision-based navigation.” IEEE Signal Processing Magazine 37, no. 4 (2020): 62\u201371.<\/p>\n\n\n\n

Deepera Inc. Technical overview and application brief. (Internal \/ confidential \u2014 figures cited in this post are limited to those cleared for public disclosure.)<\/p>\n","protected":false},"excerpt":{"rendered":"

By Ripon Kumar Dey, PhD \u2014 Principal Scientist, Deepera Inc. “Quantum camera” gets used loosely enough in popular coverage that it’s worth being precise about what the term actually means in the context of the platform we’re building at Deepera. It doesn’t mean the camera exploits quantum entanglement or does anything exotic in the science-fiction […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"ocean_front_end_style_editor":"no","ocean_post_layout":"","ocean_both_sidebars_style":"","ocean_both_sidebars_content_width":0,"ocean_both_sidebars_sidebars_width":0,"ocean_sidebar":"","ocean_second_sidebar":"","ocean_disable_margins":"enable","ocean_add_body_class":"","ocean_shortcode_before_top_bar":"","ocean_shortcode_after_top_bar":"","ocean_shortcode_before_header":"","ocean_shortcode_after_header":"","ocean_has_shortcode":"","ocean_shortcode_after_title":"","ocean_shortcode_before_footer_widgets":"","ocean_shortcode_after_footer_widgets":"","ocean_shortcode_before_footer_bottom":"","ocean_shortcode_after_footer_bottom":"","ocean_display_top_bar":"default","ocean_display_header":"default","ocean_header_style":"","ocean_center_header_left_menu":"","ocean_custom_header_template":"","ocean_custom_logo":0,"ocean_custom_retina_logo":0,"ocean_custom_logo_max_width":0,"ocean_custom_logo_tablet_max_width":0,"ocean_custom_logo_mobile_max_width":0,"ocean_custom_logo_max_height":0,"ocean_custom_logo_tablet_max_height":0,"ocean_custom_logo_mobile_max_height":0,"ocean_header_custom_menu":"","ocean_menu_typo_font_family":"","ocean_menu_typo_font_subset":"","ocean_menu_typo_font_size":0,"ocean_menu_typo_font_size_tablet":0,"ocean_menu_typo_font_size_mobile":0,"ocean_menu_typo_font_size_unit":"px","ocean_menu_typo_font_weight":"","ocean_menu_typo_font_weight_tablet":"","ocean_menu_typo_font_weight_mobile":"","ocean_menu_typo_transform":"","ocean_menu_typo_transform_tablet":"","ocean_menu_typo_transform_mobile":"","ocean_menu_typo_line_height":0,"ocean_menu_typo_line_height_tablet":0,"ocean_menu_typo_line_height_mobile":0,"ocean_menu_typo_line_height_unit":"","ocean_menu_typo_spacing":0,"ocean_menu_typo_spacing_tablet":0,"ocean_menu_typo_spacing_mobile":0,"ocean_menu_typo_spacing_unit":"","ocean_menu_link_color":"","ocean_menu_link_color_hover":"","ocean_menu_link_color_active":"","ocean_menu_link_background":"","ocean_menu_link_hover_background":"","ocean_menu_link_active_background":"","ocean_menu_social_links_bg":"","ocean_menu_social_hover_links_bg":"","ocean_menu_social_links_color":"","ocean_menu_social_hover_links_color":"","ocean_disable_title":"default","ocean_disable_heading":"default","ocean_post_title":"","ocean_post_subheading":"","ocean_post_title_style":"","ocean_post_title_background_color":"","ocean_post_title_background":0,"ocean_post_title_bg_image_position":"","ocean_post_title_bg_image_attachment":"","ocean_post_title_bg_image_repeat":"","ocean_post_title_bg_image_size":"","ocean_post_title_height":0,"ocean_post_title_bg_overlay":0.5,"ocean_post_title_bg_overlay_color":"","ocean_disable_breadcrumbs":"default","ocean_breadcrumbs_color":"","ocean_breadcrumbs_separator_color":"","ocean_breadcrumbs_links_color":"","ocean_breadcrumbs_links_hover_color":"","ocean_display_footer_widgets":"default","ocean_display_footer_bottom":"default","ocean_custom_footer_template":"","ocean_post_oembed":"","ocean_post_self_hosted_media":"","ocean_post_video_embed":"","ocean_link_format":"","ocean_link_format_target":"self","ocean_quote_format":"","ocean_quote_format_link":"post","ocean_gallery_link_images":"on","ocean_gallery_id":[],"footnotes":""},"categories":[24],"tags":[],"class_list":["post-1674","post","type-post","status-publish","format-standard","hentry","category-steam","entry"],"_links":{"self":[{"href":"https:\/\/ripondey.ca\/index.php?rest_route=\/wp\/v2\/posts\/1674","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/ripondey.ca\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/ripondey.ca\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/ripondey.ca\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/ripondey.ca\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=1674"}],"version-history":[{"count":1,"href":"https:\/\/ripondey.ca\/index.php?rest_route=\/wp\/v2\/posts\/1674\/revisions"}],"predecessor-version":[{"id":1679,"href":"https:\/\/ripondey.ca\/index.php?rest_route=\/wp\/v2\/posts\/1674\/revisions\/1679"}],"wp:attachment":[{"href":"https:\/\/ripondey.ca\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=1674"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/ripondey.ca\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=1674"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/ripondey.ca\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=1674"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}