CMOS Image Sensors Explained: Pixel Size, Resolution, and Advanced Sensing

CMOS image sensors convert light into digital information for cameras, industrial inspection systems, vehicles, scientific instruments, and connected devices. Their development focuses on reproducing fine detail, maintaining image quality across bright and dark scenes, and capturing motion accurately within limited sensor area and power. This article explains how CMOS pixels collect and read light, how pixel size affects resolution, sensitivity, noise, and dynamic range, and how technologies such as pixel binning, stacked construction, and advanced isolation structures address these tradeoffs. It also examines sensing methods that extend beyond conventional photography, including time-of-flight depth measurement, infrared and spectral imaging, polarization detection, and event-based vision.

Catalog

Simulated Human Eye

A CMOS image sensor, or CIS, converts an optical scene into electrical data that can be processed into a digital image. Its purpose is similar to that of the human visual system: detect light, preserve spatial detail, distinguish brightness and color, and respond to motion.

However, a camera does not reproduce human vision directly. The eye, retina, pupil, lens, and brain work as one adaptive system, while a camera divides these functions among the lens, image sensor, exposure controls, image signal processor, and display. Human-eye comparisons are therefore useful for explaining design goals, but direct megapixel or dynamic-range comparisons should be treated cautiously.

How a CMOS Image Sensor Converts Light

A CMOS image sensor contains an array of light-sensitive pixels. Each pixel normally includes a photodiode and several transistors used to reset, transfer, amplify, and read the accumulated charge.

The imaging process follows four basic stages:

• The lens focuses light onto the sensor

• Microlenses and color filters direct selected wavelengths into individual pixels

• Photodiodes convert absorbed photons into electrical charge

• Readout circuits measure the charge and convert it into digital data

When a photon has enough energy to cross the semiconductor band gap, it generates an electron-hole pair. The photodiode collects charge carriers during the exposure period. The amount of stored charge is approximately proportional to the number of absorbed photons until the pixel approaches saturation.

Silicon has a band gap of about 1.12 eV at room temperature. It responds across the visible spectrum and part of the near-infrared range while remaining compatible with established semiconductor manufacturing. This compatibility allows the pixel array, timing circuits, analog readout, data conversion, and digital control functions to be integrated within a compact device.

Pixel Structure and Signal Readout

A common modern pixel uses a pinned photodiode and a four-transistor, or 4T, architecture. It normally contains:

• A pinned photodiode that stores photo-generated charge

• A transfer gate that moves charge to a floating diffusion node

• A reset transistor that prepares the node for measurement

• A source-follower transistor that buffers the signal

• A row-select transistor that connects the pixel to the column circuit

The sensor first samples the reset level of the floating diffusion. It then transfers the photodiode charge and samples the signal level. Subtracting the two measurements through correlated double sampling reduces reset noise and part of the low-frequency circuit noise.

Column circuits amplify or digitize the resulting voltage. The digital data then passes to an image signal processor for defect correction, black-level adjustment, demosaicing, color correction, tone mapping, and noise reduction.

Front-Side, Back-Side, and Stacked Sensors

In a front-side-illuminated sensor, incoming light passes through interconnect and transistor layers before reaching the photodiode. These structures reduce the effective light-sensitive area and limit the angles at which light can enter the pixel.

A back-side-illuminated, or BSI, sensor reverses the light-entry direction. The silicon substrate is thinned so light reaches the photodiode without first passing through most of the wiring layers. This arrangement improves quantum efficiency, angular response, and low-light performance, especially at small pixel pitches.

Stacked sensors fabricate the pixel array and processing circuits on separate semiconductor layers. High-density vertical connections join the layers. This allows the upper layer to be optimized for light collection while the lower layer contains readout circuits, memory, conversion circuits, and signal processing.

More advanced structures may separate the photodiode and pixel transistors into different layers. This increases the space available for charge collection without increasing pixel pitch, although it adds manufacturing complexity and bonding requirements.

Sensitivity and Noise

Low-light performance depends on whether the useful signal remains distinguishable from noise.

Photon shot noise results from the statistical arrival of photons. It increases with the square root of the collected signal and cannot be removed by circuit design. Collecting more light through a wider aperture, longer exposure, larger effective pixel area, or higher quantum efficiency improves photon-limited performance.

Other noise sources include:

• Read noise from pixel, column, and conversion circuits

• Dark-current noise generated without incident light

• Fixed-pattern noise caused by pixel and column variations

• Random telegraph noise from transistor defects

• Quantization noise introduced during analog-to-digital conversion

The signal-to-noise ratio can be simplified as:

SNR = signal electrons ÷ total noise electrons

In bright conditions, photon shot noise often dominates. In dark conditions, read noise and dark current become more influential.

Quantum Efficiency and Optical Losses

Quantum efficiency describes the fraction of incident photons that produce collected charge. A high value allows the sensor to generate a stronger electrical signal from the same amount of light.

Actual sensitivity also depends on:

• Lens and cover-glass transmission

• Microlens efficiency

• Color-filter transmission

• Reflection at material interfaces

• Absorption depth in the silicon

• Charge-collection efficiency

• Optical crosstalk between pixels

A monochrome sensor normally collects more light than a comparable color sensor because it does not use red, green, and blue filters that reject part of the incoming spectrum.

Full-Well Capacity and Conversion Gain

Full-well capacity, or FWC, is the approximate number of electrons a pixel can store before its response saturates or becomes nonlinear. A higher FWC allows the pixel to record stronger illumination and preserve highlight detail.

Smaller pixels generally have less storage volume. Designers compensate through deeper photodiodes, optimized doping, vertical transfer gates, stacked structures, and additional charge-storage nodes.

Conversion gain describes how much the pixel output voltage changes for each collected electron. High conversion gain produces a larger voltage from a weak signal and helps reduce input-referred read noise. Low conversion gain provides more voltage headroom and supports a larger charge range.

Many sensors use dual-conversion-gain operation. High gain is selected for dark regions, while low gain is used for brighter signals. The two measurements can be combined to improve dynamic range.

Dynamic Range

Dynamic range is the ratio between the largest usable signal and the smallest detectable signal:

Dynamic range = 20 log10(full-well capacity ÷ read noise)

For example, a pixel with a usable FWC of 20,000 electrons and a read noise of 2 electrons has an idealized dynamic range of:

20 log10(20,000 ÷ 2) = 80 dB

Practical performance may be lower because of lens flare, fixed-pattern noise, converter limits, analog headroom, dark current, and image-processing constraints.

Image sensors use several methods to extend dynamic range:

• Multiple exposures with different integration times

• Dual-conversion-gain readout

• Split-pixel designs with high- and low-sensitivity regions

• Overflow capacitors for additional charge storage

• Multiple signal samples during one exposure

• Staggered exposure timing

• Nonlinear pixel responses

Multiple-exposure HDR can create motion artifacts because the scene is captured at different times. Single-exposure methods reduce this problem but require additional pixel area or circuit complexity.

Motion Capture and Shutter Type

Most consumer CMOS sensors use a rolling shutter. Rows are exposed or read at slightly different times. Rapid motion can therefore bend straight lines, skew rotating objects, or create bands under flickering light.

A global-shutter sensor exposes all pixels during the same time interval. It is preferred for:

• Industrial machine vision

• Robotic guidance

• Barcode inspection

• Motion analysis

• High-speed manufacturing

• Vehicle perception

• Scientific measurement

Global-shutter pixels require additional storage structures. These structures can reduce light-sensitive area or saturation capacity unless the sensor uses BSI or stacked construction.

Main Sensor Performance Metrics

No single specification determines image quality.

Parameter	What It Measures	Practical Effect
Pixel pitch	Center-to-center pixel spacing	Affects resolution, sensitivity, and sensor size
Quantum efficiency	Collected electrons per incident photon	Higher values improve sensitivity
Full-well capacity	Maximum stored charge	Higher values preserve bright detail
Read noise	Electronic uncertainty during readout	Lower values reveal weaker signals
Dynamic range	Strongest usable signal relative to noise floor	Wider range preserves shadows and highlights
Dark current	Thermally generated charge	High values cause long-exposure noise and hot pixels
Conversion gain	Voltage change per electron	Affects low-light sensitivity and signal capacity
Crosstalk	Signal entering neighboring pixels	Reduces color accuracy and spatial detail
Frame rate	Images captured per second	Determines motion-recording capability
Shutter type	Exposure timing across the array	Determines rolling-shutter distortion

The correct balance depends on the application. Smartphones prioritize compact size, resolution, low power, and computational imaging. Industrial cameras prioritize timing accuracy and low motion distortion. Scientific sensors prioritize low noise, linearity, cooling, and calibrated output.

Balancing Pixel Size and Resolution

The competition between pixel size and resolution is an optimization problem rather than a simple race toward smaller dimensions. Sensor designers must balance spatial resolution, sensitivity, full-well capacity, crosstalk, optics, data rate, power consumption, and manufacturing yield.

Pixel Pitch, Sensor Area, and Resolution

For a fixed sensor area, reducing pixel pitch increases the number of pixels. Halving the pitch in both directions allows approximately four times as many pixels to fit within the same area.

The tradeoff is that each pixel receives light over a smaller physical region and generally stores less charge. Increasing the total sensor area can preserve larger pixels and high resolution, but it requires a larger image circle, lens assembly, package, data path, and power budget.

Smartphones face strict limits on camera thickness and lens diameter. Small pixels allow higher resolution without increasing module size at the same rate.

Optical Limits on Resolution

A sensor cannot record detail that the lens does not resolve. Lens aberrations, diffraction, focus error, motion blur, and optical filtering can all reduce useful resolution.

The approximate diameter of the central diffraction pattern is:

Airy disk diameter = 2.44 × wavelength × f-number

At a wavelength of 550 nm and an aperture of f/2.0, the Airy disk diameter is about 2.7 μm. Pixels smaller than this value can still improve sampling, cropping, stabilization, and demosaicing, but the gain in actual scene detail may be smaller than the increase in nominal megapixels.

Why Smaller Pixels Have Lower Per-Pixel SNR

A smaller pixel may have good quantum efficiency, but it collects fewer total photons because of its reduced area.

If one pixel collects 1,600 photons, its photon shot noise is approximately:

√1,600 = 40

Its shot-noise-limited SNR is:

1,600 ÷ 40 = 40

A pixel with one-quarter of the area may collect about 400 photons:

√400 = 20

Its SNR becomes:

400 ÷ 20 = 20

Combining four smaller pixels can recover much of the larger-pixel signal, but real performance also depends on read noise, fill factor, isolation losses, color filters, and the stage at which the signals are combined.

Pixel Binning

Pixel binning combines neighboring pixel signals into a lower-resolution output with a higher effective signal level.

Grouped color-filter arrays place several neighboring pixels under filters of the same color. Common arrangements include:

• 2 × 2 four-cell arrays

• 3 × 3 nine-cell arrays

• 4 × 4 sixteen-cell arrays

A 108-megapixel sensor using 3 × 3 grouping can produce an output near 12 megapixels. A 200-megapixel sensor using 4 × 4 grouping can produce an output near 12.5 megapixels.

Binning can occur in the charge, analog-voltage, conversion, or digital domain. Earlier combination can reduce the effect of repeated read noise, while digital summation includes the read noise of each contributing pixel.

Binning improves low-light output, reduces data volume, and supports video modes that do not require full sensor resolution. It does not correct poor optics, focus error, motion blur, or insufficient scene illumination.

Full-Resolution Capture and Remosaicing

Grouped color-filter arrays do not sample color in the same pattern as a standard Bayer sensor. Producing a full-resolution image therefore requires remosaicing, which estimates missing color information.

Remosaicing can preserve additional detail under bright conditions, but it may create:

• False color

• Moiré

• Edge artifacts

• Reduced color resolution

• Noise amplification

• Higher processing demand

Full-resolution modes therefore work best in well-lit scenes with limited motion. Binned modes are usually better for low light, video, and faster readout.

Crosstalk and Pixel Isolation

As pixel pitch decreases, light and charge are more likely to enter neighboring pixels. This reduces color accuracy and spatial resolution.

Optical crosstalk occurs when light passes through the optical stack at an angle and reaches an adjacent photodiode. Electrical crosstalk occurs when generated carriers diffuse through the silicon before collection.

Common countermeasures include:

• Deep-trench isolation

• Metal light shields

• Improved microlens alignment

• Optimized color filters

• Controlled photodiode depth

• Back-side surface treatment

• Revised doping profiles

These structures improve isolation but increase process complexity. Poor trench surfaces can also introduce defects and dark current.

Preserving Charge Capacity

Small pixels require new methods to maintain usable charge storage. Designers may use:

• Deeper vertical photodiodes

• Optimized potential wells

• Vertical transfer gates

• Separate transistor and photodiode layers

• Overflow capacitors

• Dual-conversion-gain readout

These methods must be evaluated for charge-transfer efficiency, image lag, blooming, linearity, temperature dependence, dark current, and manufacturing variation.

Pixel Architecture Comparison

Pixel Approach	Main Benefit	Main Limitation	Typical Use
Large native pixel	Strong light collection and high FWC	Lower resolution for a fixed area	Scientific, industrial, low-light cameras
Small native pixel	High pixel density and compact size	Lower signal per pixel	Smartphones and compact cameras
2 × 2 grouped array	Flexible full-resolution and binned modes	Requires remosaicing	Mobile main and telephoto cameras
3 × 3 grouped array	Strong nine-pixel low-light grouping	More complex reconstruction	High-resolution mobile sensors
4 × 4 grouped array	Very high nominal resolution	High bandwidth and processing demand	200-megapixel mobile sensors
Split-pixel HDR	Simultaneous bright- and dark-region capture	Larger or more complex pixels	Automotive and surveillance
Stacked pixel	More space for photodiode and logic	Higher manufacturing complexity	Advanced mobile sensors

High-Resolution Video

A 4K UHD frame contains about 8.3 megapixels, far fewer than the total count of a modern high-resolution mobile sensor.

Video can be produced through:

• Pixel binning

• Line skipping

• Pixel skipping

• Full-array oversampling

• Cropped readout

Line skipping reduces data demand but may create aliasing. Cropping speeds readout but narrows the field of view. Full-array oversampling improves detail but consumes more bandwidth and power. Binned readout usually provides a practical balance between sensitivity and processing load.

Video performance is also limited by rolling-shutter time, processor throughput, memory bandwidth, interface speed, heat, battery capacity, autofocus, and HDR timing.

When Higher Resolution Helps

More megapixels can provide:

• Additional cropping flexibility

• Greater digital zoom range

• Improved stabilization margins

• Better oversampling

• Fine detail under controlled lighting

• More spatial samples for inspection

They provide less value when:

• The lens cannot resolve the added detail

• The scene is too dark

• Motion blur exceeds the pixel dimensions

• Heat limits processing

• Storage or transmission capacity is restricted

• Noise reduction removes the additional detail

The pixel-size race will continue, but future progress will depend increasingly on stacked construction, improved optical materials, advanced charge-storage structures, faster readout, and application-specific operating modes rather than pitch reduction alone.

Beyond the Human Eye

Conventional cameras record two-dimensional visible-light intensity and color. Advanced CMOS sensors can also measure distance, wavelength, polarization, timing, and local changes within a scene.

Depth-Sensing Methods

Common depth-sensing methods include:

• Stereo vision

• Structured light

• Indirect time of flight

• Direct time of flight

• LiDAR scanning

• Focus-based estimation

Each method differs in range, resolution, lighting requirements, power consumption, and processing load.

Stereo Vision

Stereo vision uses two cameras separated by a known baseline. The system matches features in the two images and measures their displacement, or disparity.

Depth can be approximated by:

Depth = focal length × baseline ÷ disparity

A wider baseline improves long-range accuracy but increases module width. A narrow baseline supports compact products but provides less disparity for distant objects.

Stereo performance declines on smooth surfaces, repeating patterns, reflective objects, transparent materials, and poorly lit scenes. It is used in robotics, industrial measurement, mapping, and vehicle perception.

Time-of-Flight Sensing

A time-of-flight, or ToF, system illuminates a scene and measures information related to the return time of the reflected light. It can generate depth data without depending on visible texture.

ToF systems are used for:

• Gesture recognition

• Face authentication

• Room mapping

• Object measurement

• Robot navigation

• In-cabin monitoring

• Autofocus assistance

Indirect Time of Flight

Indirect ToF, or iToF, emits modulated light and measures the phase difference between the transmitted and reflected waveforms.

Distance can be expressed as:

Distance = c × phase shift ÷ 4πf

where c is the speed of light and f is the modulation frequency.

Higher modulation frequencies can improve precision but reduce the unambiguous range. Multiple frequencies may be used to resolve phase wrapping.

iToF supports dense depth maps and high frame rates, but its accuracy can be affected by ambient light, signal attenuation, multipath reflections, calibration error, and phase ambiguity.

Direct Time of Flight and SPADs

Direct ToF, or dToF, measures the arrival time of reflected photons. Many dToF sensors use single-photon avalanche diodes, or SPADs.

A SPAD is biased above its breakdown voltage. One photo-generated carrier can trigger an avalanche, which is then stopped by a quenching circuit. Timing or histogram circuits record the photon arrival distribution over repeated optical pulses.

dToF can support longer range and precise timing, but designers must manage:

• SPAD dead time

• Dark count rate

• Optical crosstalk

• Afterpulsing

• Timing-circuit area

• High event data rates

• Laser-safety limits

Stacked construction allows the SPAD array to sit above digital timing and processing circuits, improving detector density.

iToF and dToF Comparison

Characteristic	Indirect ToF	Direct ToF
Measurement basis	Phase difference	Photon flight time
Common detector	Demodulating photodiode	SPAD
Main strength	Dense depth maps	Longer range and precise timing
Main limitation	Multipath and phase ambiguity	Detector and timing complexity
Common use	Gesture, industrial depth, cabin sensing	LiDAR, mobile depth, long-range sensing

The correct method depends on range, field of view, depth precision, resolution, ambient light, emitter power, and safety requirements.

Ultraviolet and Infrared Imaging

Silicon sensors can detect visible light and part of the near-infrared spectrum. Standard color cameras normally use an infrared-cut filter because unwanted infrared affects color reproduction.

Modified silicon sensors support:

• Night surveillance

• Eye tracking

• Vein imaging

• Industrial alignment

• Agricultural inspection

• Fluorescence imaging

Ultraviolet imaging requires suitable detector surfaces, filters, lenses, and coatings because ordinary optical materials may absorb ultraviolet wavelengths.

Short-Wave Infrared Imaging

Silicon becomes ineffective beyond wavelengths near 1.1 μm. Short-wave infrared sensors therefore use materials such as indium gallium arsenide, germanium, or specialized quantum-dot structures.

SWIR imaging is used for:

• Wafer inspection

• Food sorting

• Moisture detection

• Recycling

• Agricultural analysis

• Laser-beam profiling

• Pharmaceutical inspection

These detectors often require separate fabrication and bonding to a silicon readout circuit, increasing cost and pixel pitch.

Multispectral and Hyperspectral Imaging

A standard color sensor measures broad red, green, and blue bands. Multispectral systems capture several selected bands, while hyperspectral systems record many narrow wavelength bands.

The resulting data cube contains spatial position and spectral intensity. Materials that appear similar in ordinary color images may have distinct spectral signatures.

Applications include:

• Crop monitoring

• Mineral identification

• Medical analysis

• Food inspection

• Environmental monitoring

• Artwork examination

• Chemical sorting

Increasing the number of spectral bands reduces the light available in each band and raises data-processing requirements.

Polarization Sensors

Polarization sensors measure the orientation of the electric field in reflected light. A common design places micro-polarizers at 0, 45, 90, and 135 degrees above neighboring pixels.

The sensor can calculate intensity, polarization angle, and degree of linear polarization.

Applications include:

• Glare reduction

• Transparent-material inspection

• Mechanical stress detection

• Scratch and defect inspection

• Reflection separation

• Biological imaging

The polarizer layer reduces transmitted light, while the grouped sampling pattern lowers effective spatial resolution.

Event-Based Vision Sensors

A conventional sensor records complete frames at fixed intervals. An event-based sensor reports only local brightness changes that exceed a selected threshold.

Each event normally includes:

• Pixel coordinates

• Change direction

• Timestamp

This method provides low latency, reduced motion blur, wide dynamic range, and less redundant data in static scenes.

Event sensors are used for:

• High-speed tracking

• Drone navigation

• Robotic control

• Industrial motion analysis

• Eye tracking

• Vibration measurement

They do not directly produce conventional full-color frames, so additional processing may be needed for image reconstruction.

On-Sensor Processing

Stacked sensors can place memory and digital processing beneath the pixel array. This allows selected operations to occur before the data leaves the sensor.

Possible functions include:

• Motion detection

• Region-of-interest selection

• Exposure control

• Object-presence detection

• Depth calculation

• Event extraction

• Compression

• Privacy masking

Local processing reduces bandwidth and may improve privacy by transmitting only the required result instead of a complete image. The tradeoff is greater chip complexity, heat generation, software effort, and verification demand.

Practical Sensor Selection

Sensor selection should begin with the measurement requirement rather than the megapixel count.

A smartphone camera usually needs a compact BSI sensor, low-power readout, grouped-pixel modes, HDR, and fast autofocus. A factory camera may require global shutter, external triggering, and deterministic timing. A depth camera requires an emitter, optical filtering, calibration, and safety analysis. A scientific camera may prioritize cooling, low noise, linear response, and calibrated output.

Key selection questions include:

• What wavelength range must be detected

• What is the smallest required feature size

• How much light reaches the sensor

• What dynamic range is required

• How fast does the subject move

• Is rolling-shutter distortion acceptable

• What resolution and frame rate are needed

• Is direct depth measurement required

• What bandwidth is available

• What power and thermal limits apply

• Is calibrated raw data needed

• What temperature range must be supported

• What privacy restrictions apply

• What sensor and optics cost can the system support

The future of CMOS image sensing will not depend only on producing images that resemble human vision. New sensors will increasingly combine visible imaging, depth measurement, infrared response, spectral analysis, polarization detection, and event sensing.

The value of these systems will depend on how accurately they convert physical information into usable data while remaining within limits for size, power, bandwidth, cost, privacy, and reliability.

Conclusion

CMOS image sensor performance depends on the combined design of the photodiode, pixel circuitry, optical stack, readout architecture, and processing system. Smaller pixels support higher resolution and compact camera modules, but they also reduce per-pixel light collection and charge capacity, requiring improved isolation, binning, conversion-gain control, and stacked structures. At the same time, CMOS sensors are expanding beyond visible image capture by measuring depth, wavelength, polarization, and rapid scene changes. Selecting the appropriate sensor therefore requires balancing resolution, sensitivity, dynamic range, motion performance, wavelength response, bandwidth, power consumption, cost, and the specific measurement needs of the final system.

Frequently Asked Questions [FAQ]

1. Why is reducing pixel size challenging in CMOS image sensors?

Smaller pixels increase resolution but collect less light and store less charge, which can reduce sensitivity and dynamic range. Technologies such as BSI, stacked sensors, deep-trench isolation, and pixel binning help offset these effects.

2. How does pixel binning improve image quality?

Pixel binning combines neighboring pixels into one larger effective pixel. This improves low-light sensitivity, reduces noise, and supports high-quality video while allowing the sensor to capture full-resolution images when lighting is sufficient.

3. What determines the dynamic range of a CMOS image sensor?

Dynamic range depends mainly on full-well capacity and read noise. A higher full-well capacity and lower read noise allow the sensor to capture more shadow and highlight detail in the same image.

4. What is the difference between indirect ToF and direct ToF sensors?

Indirect ToF measures phase differences in reflected light and is commonly used for short- to medium-range depth sensing. Direct ToF measures photon travel time, providing greater range and higher depth accuracy.

5. Why do modern CMOS image sensors support infrared and depth sensing?

Infrared and depth sensing allow CMOS image sensors to perform tasks beyond photography, including facial recognition, industrial inspection, robotics, autonomous driving, and 3D mapping.