An artificial vision sensor inspired by the human pupil adapts automatically to light, maintaining clear perception from darkness to glare and sharply improving recognition accuracy in machine learning systems.
(Nanowerk Spotlight) Machine vision has long faced a basic trade-off. In dim light sensors generate random noise that buries detail, while in bright light they saturate and lose information. The human eye handles both extremes with remarkable ease. It converts light into brief electrical impulses called spikes and adjusts sensitivity through a reflex in the iris known as the pupillary light reflex. The pupil narrows when light is strong, limiting the influx of photons, and widens in the dark to admit more. This simple mechanism protects the retina and keeps neural activity within an efficient range.
Researchers have sought to reproduce this adaptability in artificial sensors. Neuromorphic vision, an area that merges sensing and computation, aims to mimic the way biological neurons encode information through discrete spikes. Many devices can already detect changes in brightness or motion and extend their working range through electronic control circuits that adjust bias currents.
These methods improve performance but require external energy and control, and they change signal amplitude rather than spike timing. A system that could adapt to light intensity on its own while producing spike-based outputs would bring artificial vision a step closer to biology.
A paper published in Advanced Materials (“A Pupillary Light Reflex Inspired Self‐Adaptive Spiking Visual Neuron”) presents such a system. The work describes a device that imitates the eye’s reflex by combining an adaptive optical film, a light sensitive transistor, and a memory component that generates spikes. The three layers work together to regulate light, convert it into current, and encode the result as electrical pulses.
At the top of the stack lies a film that darkens automatically as light intensity rises and clears again when it falls. This photochromic material, made from a compound called DMTMM in the spirooxazine family, acts as a self regulating pupil. The molecules in the film change structure when exposed to light, switching from a transparent form to a darker one and reversing when the light weakens.
Measurements show that the film transmits more than eighty percent of light at low intensity and about four percent at high intensity. It therefore limits the light reaching the transistor without any need for external control signals or added power.
Once the light passes through this film it reaches the transistor, which converts it into electrical current. The transistor’s channel is made from a material called IGCdO, a close relative of the oxide semiconductor IGZO. IGCdO has a narrower bandgap that lets it absorb red light and can be deposited at room temperature using standard sputtering methods. Its structure is amorphous, avoiding the grain boundaries that often create irregularities in crystal materials.
Self-adaptive spiking visual neuron inspired by the pupillary light reflex. a) The mechanism of pupillary light reflex and its significant role in visual encoding b) The biological coding frequency limit and the extension of visual adaptation range by pupillary light reflex. c) The structure and optical image of self-adaptive spiking visual neuron, and the photochromic principle of DMTMM film. Region I and II are the IGCdO-based transistor and TaOX-based memristor, respectively. d) Cross-section images of I and II regions of the device in c). e) The pupillary light reflex function represents a promising strategy for enhancing the visual performance of bionic vision systems. (Image: Reprinted with permission from Wiley-VCH Verlag) (click on image to enlarge)
Tests show that it achieves a current on to off ratio of about 2.45 times ten to the eighth, a field effect mobility near 21.34 square centimeters per volt second, and a subthreshold slope of roughly 71.53 millivolts per decade. The gate insulator, made from hafnium aluminum oxide, provides high capacitance and an extremely low leakage current density below one times ten to the minus eight amperes per square centimeter. The transistor remains stable during long stress tests and responds reliably to red light at 655 nanometers.
The electrical current from the transistor is then transformed into spikes by a device called a memristor. A memristor is a component whose resistance depends on the history of voltage and current. The version used here is built from layers of silver, tantalum oxide, and aluminum. When voltage across it rises to a threshold, silver ions move through the oxide and create a conductive path. When voltage falls below a certain level, the path breaks down.
This repeating cycle of formation and rupture produces distinct voltage pulses. Inside the memristor, a small capacitance gradually charges under constant current until it reaches the threshold, then discharges in a sharp spike before recharging. The frequency of these spikes increases with stronger input current. Experiments across currents from 0.1 to 100 nanoamperes confirm stable, repeatable spiking that scales predictably with input strength.
The researchers combined the photochromic layer, the IGCdO transistor, and the TaOx memristor into a single vertical structure that functions as a complete neuron. Arrays containing ten by ten cells were fabricated on two inch wafers using standard semiconductor processes. The results show that the optical film expands the usable light range dramatically.
Without the film, the device saturates at about 51 milliwatts per square centimeter. With the film, it continues operating up to 1.64 watts per square centimeter. The adaptive version reaches a spike frequency of about 800 hertz at this upper limit, whereas the nonadaptive version reaches the same rate at roughly 20 milliwatts per square centimeter. This demonstrates that the optical front end both widens the dynamic range and protects the transistor beneath from overexposure.
The study defines its dynamic perception range as twenty times the logarithm base ten of the ratio between the maximum measurable current and the minimum detectable off current. By this measure the new neuron achieves about 160 decibels. This corresponds to a difference in brightness of roughly ten million to one, similar to the operational range of human vision. The system reaches this range while keeping the output in spike form and without any external control circuits.
To test how this adaptation affects perception, the spike output was connected to a spiking neural network, a computational model that processes information through discrete events instead of continuous values. The researchers created a dataset of images with a wide range of brightness and trained two identical networks. One used data from the adaptive neuron with the photochromic film, and the other used data from a fixed sensitivity neuron.
After one hundred training cycles, the adaptive network achieved about 86 percent recognition accuracy, while the fixed network reached only about 20 percent. This result shows that hardware level adaptation helps stabilize signals and improves the ability of spike-based algorithms to interpret visual scenes under variable lighting.
Several design features contribute to the stability of this system. The oxide semiconductor can be fabricated at room temperature, which supports compatibility with other chip processes. Its amorphous structure provides uniform conduction, and the high permittivity dielectric minimizes leakage while reducing operating voltage.
The memristor’s silver ion mechanism gives repeatable threshold switching over many cycles. The concentration of DMTMM in the photochromic film can also be tuned to balance transparency and adaptation strength, allowing designers to set the tradeoff between sensitivity and protection.
This approach arrives at a time when artificial vision is moving toward sensors that combine optical capture and computation within the same unit. Robots, vehicles, and mobile systems need eyes that adjust automatically to glare or shadow without relying on complex exposure control or large power budgets.
The self adaptive spiking neuron described in this work demonstrates that a thin optical film coupled with spiking electronics can achieve this function. It delivers a wide operational range, maintains event-based coding, and improves recognition accuracy under varied lighting. While it does not recreate every feature of the biological eye, it reproduces one of its most essential reflexes, a pupil that adapts by itself, and connects it directly to the generation of neural spikes. This combination points toward vision hardware that can remain reliable and efficient when the environment shifts from darkness to bright sunlight in an instant.
For authors and communications departmentsclick to open
Lay summary
Prefilled posts
Nanowerk Newsletter
Get our Nanotechnology Spotlight updates to your inbox!
Thank you!
You have successfully joined our subscriber list.
Become a Spotlight guest author! Join our large and growing group of guest contributors. Have you just published a scientific paper or have other exciting developments to share with the nanotechnology community? Here is how to publish on nanowerk.com.
