Access Official Technical PDF

Share Summary:LinkedInEmail

A high-speed CMOS image sensor has to do more than acquire images quickly. It must capture the whole scene at the same instant, preserve enough light-derived signal for useful image quality, and store frames faster than they can be read out. A demonstrated global-shutter design addresses these demands with correlated double sampling (CDS) in each pixel and dedicated on-chip analog memory. The 100 × 100-pixel sensor was tested at 200,000 frames per second (fps), with exposure times as short as 1 microsecond.

Although developed for scientific high-speed imaging rather than a specific industrial inspection line, the architecture illustrates important considerations for engineers and system integrators working with transient events: synchronization, image sensitivity, data capture and the trade-offs involved in scaling sensor resolution.

Why a global shutter matters in high-speed imaging

In a rolling-shutter sensor, rows are exposed at different times. When an object moves rapidly during that interval, the captured image can show skew or other timing-related distortions. A global-shutter sensor exposes the complete array together, making it better suited to recording a fast event without row-to-row exposure offsets.

Global shutter alone does not guarantee a high frame rate. After exposure, pixel signals still need to be transferred and digitized. In a conventional readout approach, this process can limit how quickly successive frames are captured, particularly as pixel count increases. The design described here uses burst capture: each pixel transfers its signal to on-chip memory, allowing several frames to be stored before the slower readout process begins.

How the high-speed CMOS image sensor works

Pixel-level sampling and noise control

Each pixel uses an 11-transistor circuit and two MOSFET capacitors. Its front end includes a pinned photodiode, which converts incident light into charge, and circuitry that performs CDS. In simplified terms, CDS samples a reset level and a signal level, then uses their difference to reduce reset-related noise. Performing this operation within the pixel allows the stored frame value to be read in a single readout.

The pixel measures 15.1 × 15.1 micrometres and has a fill factor above 40%. Fill factor describes the portion of a pixel area that is photosensitive; retaining a relatively large active area supports light collection, an important consideration when exposure times are short. The design also includes an in-pixel current source that can be switched off when not in use. This arrangement helps avoid output-voltage dependence on the pixel’s location in the sensor.

The pixel’s MOSFET capacitors do not occupy upper metal layers, leaving those layers available for routing. Together, the capacitor approach, pixel circuitry and layout are intended to fit CDS and signal-driving functions into a compact pixel while maintaining the stated fill factor.

Dedicated memory enables burst capture

Each pixel connects to a memory block outside the pixel array. The block contains 16 analog storage elements, implemented as MOSFET capacitors with access switches. The architecture can store up to 16 frames; the described operation uses 15 consecutive frames, with one frame reserved for calibration.

Fifty pixels share a connection line to their memory blocks, so transferring pixel values to memory takes 50 clock cycles. This parallel storage path separates rapid capture from subsequent digitization. For readout, two 100 MHz, 12-bit pipeline analog-to-digital converters (ADCs) are used. The reported readout time is 8 milliseconds.

Measured performance and design trade-offs

Testing demonstrated operation at 200,000 fps and exposure times down to 1 microsecond. The measured sensitivity was 22.3 microvolts per electron, while linearity was within ±1.9%. The reported nonuniformity error at 50% saturation was 4.1%; the work notes that software calibration could reduce this error.

Noise performance reflects the complete signal chain, not just the pixel. The sensor’s stored-memory noise was reported as 0.4 millivolts rms, and the noise read from the ADC was 0.33 millivolts rms. Testing also identified significant input-referred noise in the selected ADCs, measured at 0.7 millivolts rms. Including this contribution, the total input-referred noise was 38 electrons, compared with an expected 15 electrons. The resulting signal-to-noise ratio (SNR) was 43.9 dB, and dynamic range was 57.5 dB.

These results show why sensor selection and integration require evaluation of the entire acquisition chain. Pixel sensitivity, memory noise, ADC noise and readout timing all contribute to the final image. A fast capture rate is useful only when the resulting data retain the quality required for analysis.

Memory retention and scaling considerations

Analog memory introduces a retention concern: stored voltages can change as charge leaks away. Calculated leakage over the 8-millisecond readout interval was negligible, but testing found that leakage could become significant when memory blocks were exposed to light. For high-irradiance conditions, the memory blocks therefore need light shielding.

The tested device is a 100 × 100-pixel sensor, not a demonstrated megapixel sensor. Its architecture was designed for potential expansion to 1000 × 1000 pixels using the same building blocks and a similar structure. The proposed expansion is a design direction rather than a measured result. The work identifies reducing ADC noise and adding a light-blocking layer over the memories as steps for the expanded sensor. Under those proposed changes, the projected dynamic range would exceed 63 dB, while SNR was expected to remain the same.

Practical implications for automation and measurement

For engineers evaluating imaging of short-duration events, this design highlights several questions to include in a system specification:

  • Timing: Does the application require simultaneous exposure across the full field of view?
  • Capture and readout: Is burst storage needed because the event rate exceeds practical continuous readout speed?
  • Signal quality: Have ADC noise and other downstream contributions been included alongside pixel-level noise?
  • Operating conditions: Could light exposure affect stored analog values during the readout interval?
  • Resolution: Are scaling assumptions based on a tested device or on a proposed architecture?

A global-shutter CMOS sensor with on-chip memory can separate rapid event capture from slower frame conversion, while pixel-level CDS and a relatively high fill factor address noise and sensitivity concerns. The demonstrated results establish operation at 200,000 fps with 1-microsecond exposures in a 100 × 100-pixel device. For practical deployment, engineers should assess the complete readout chain, account for memory light sensitivity, and treat megapixel performance as a design objective until validated in hardware.

Access Official Technical PDF

About The Author