Chapter 24. Correlated Double Sampling: Subtracting the Reset Noise from Every Pixel |
Summary |
Every time a CCD pixel's charge is read, the output amplifier's floating diffusion node must be reset to a known reference voltage. That reset action, unavoidable as it is, injects a random fluctuation known as reset noise or kTC noise into the signal path. Correlated double sampling, or CDS, is the circuit technique that removes this noise by measuring the floating diffusion twice: once immediately after reset, before the signal charge arrives, and once after the charge has been transferred. The difference between these two measurements is the true signal, with the reset noise largely cancelled because it appears identically in both samples. This chapter explains how CDS works in principle, why it became indispensable in CCD scanners and scientific imagers, and how its benefits manifest across industries ranging from barcode scanning and document digitization to medical imaging, astronomy, and industrial inspection. The chapter closes with a detailed summary of CDS's role in the CCD signal chain. |
|
1. The Problem CDS Solves: Reset Noise in the Floating Diffusion |
To understand why correlated double sampling exists, one must first understand what happens at the very end of the CCD's charge-transfer chain. After photons have been converted to electrons, after those electrons have been shifted row by row and pixel by pixel through the silicon, they arrive at a small region called the floating diffusion. This node is essentially a capacitor with a certain physical capacitance. The charge packet deposited onto it changes its voltage by an amount proportional to the number of electrons. That voltage is then buffered by a source-follower amplifier and sent off-chip for further processing. |
Before the next charge packet arrives, the floating diffusion must be reset to a known reference voltage. A reset transistor briefly connects the diffusion to a supply rail, charging it back up. When the transistor turns off, the charge on the floating diffusion is frozen in place. But here is the catch: the act of turning off a transistor is not perfectly deterministic. Thermal agitation of electrons in the channel means that the exact amount of charge left behind varies randomly from one reset to the next. This randomness manifests as a voltage fluctuation on the floating diffusion, and it is commonly called kTC noise, reset noise, or thermal reset noise. |
The magnitude of this noise depends on the capacitance of the floating diffusion and the temperature. Smaller capacitance gives larger voltage noise for the same charge fluctuation, which is one reason why CCD designers pay close attention to the geometry and doping of the output node. In early CCD scanners and scientific cameras, reset noise was often the dominant noise source, limiting the ability to distinguish faint signals from the noise floor. |
The crucial insight that makes CDS possible is this: the reset noise voltage, once frozen onto the floating diffusion, stays there. It affects the reference level that the amplifier sees before the signal charge arrives. When the signal charge is subsequently transferred onto the floating diffusion, the same noise voltage is still present, now superimposed on the signal-induced voltage change. If one could measure the reference level and the signal level in quick succession, and subtract the former from the latter, the reset noise would cancel out almost perfectly. That is precisely what correlated double sampling does. |
|
2. How CDS Works in Principle: Two Samples, One Difference |
The name 'correlated double sampling' describes the technique precisely. Two samples are taken of the output amplifier's voltage: one at the reset reference level, and one at the image signal level. The two samples are 'correlated' because they share the same reset noise component. By subtracting one from the other, the shared noise is removed. |
In a typical CCD output chain, the sequence unfolds as follows. First, the floating diffusion is reset. The amplifier output settles to the reset reference voltage, which includes the random kTC noise for that particular pixel. A sampling switch closes briefly, storing this reference voltage on a capacitor. Next, the charge packet from the horizontal shift register is transferred onto the floating diffusion. The amplifier output changes by an amount proportional to the signal charge. A second sampling switch closes, storing this signal voltage on another capacitor. Finally, a difference amplifier or switched-capacitor network computes the difference between the two stored voltages. |
The timing of the two samples matters a great deal. The reference sample must be taken after the reset transient has settled but before any signal charge arrives. The signal sample must be taken after the charge transfer is complete but before the next reset begins. In practical CCD scanners, these timing windows are determined by the horizontal clock phases and the reset clock, all of which must be carefully sequenced so that the CDS circuit samples at the right moments. |
One important subtlety is that CDS does not only remove reset noise. Because the two samples are taken close together in time, any low-frequency noise from the output amplifier, such as flicker noise or drift in the bias conditions, also tends to be correlated between the two samples and is therefore partially cancelled. This makes CDS valuable not only for its reset-noise suppression but also as a general noise-shaping technique that improves the quality of the analog signal before digitization. |
In the earliest implementations, CDS was performed with analog sample-and-hold circuits using discrete switches and capacitors. As CCD scanners evolved, the function migrated into integrated analog front-end chips, where the sample-and-hold stages, the difference amplifier, and the analog-to-digital converter are all fabricated on the same die. More recent designs perform the subtraction in the digital domain after digitizing both samples, but the principle remains the same. |
|
3. Why CDS Matters for Barcode Scanning and Document Imaging |
Barcode scanners were among the first mass-market products to depend on CCD sensors, and they remain one of the clearest illustrations of why CDS is not an optional refinement but a practical necessity. |
Consider a handheld barcode scanner in a retail environment. The scanner illuminates a label with a red LED or laser line, and a linear CCD array captures the reflected light across the width of the barcode. The contrast between dark bars and light spaces produces a corresponding pattern of charge packets in the CCD. The scanner must decide, for each pixel along the scan line, whether it saw a bar or a space. This is a threshold decision, and the threshold must be reliable across a wide range of label quality, ambient lighting, and label distance. |
If reset noise were present at the level of several hundred microvolts on the floating diffusion, and the signal difference between a bar and a space were only a few millivolts in weak lighting, the noise could easily push a pixel across the decision threshold incorrectly. The result would be a misread or a failed scan. CDS suppresses the reset noise so that the dominant noise sources become the photon shot noise and the amplifier's own input-referred noise, both of which are better behaved and more predictable. In practical terms, CDS is what allows a scanner to read a poorly printed label at an awkward angle without requiring excessive illumination or exotic optics. |
The same logic applies to document scanners and fax machines, which use linear CCD arrays to digitize pages line by line. A document scanner must reproduce fine text and halftone images faithfully. Reset noise, if left uncorrected, would appear as a fixed-pattern-like graininess that varies from line to line, degrading the perceived sharpness of text and introducing visible artifacts in smooth gradients. Because the scanner reads thousands of lines per page, even a small amount of uncorrelated reset noise would accumulate into a noticeable degradation of image quality. CDS, by making each line's pixel values more consistent with their neighbors, preserves the spatial coherence that the human eye expects in a scanned document. |
In both barcode and document scanning, the CDS circuit operates at relatively modest speeds by modern standards, but the noise requirements are stringent because the signal levels are small and the decisions are binary. The technique's ability to remove a noise source that would otherwise be irreducible makes it a foundational element of the scanner's analog front end. |
|
4. CDS in Medical Imaging: Endoscopy, Radiography, and Microscopy |
Medical imaging places demands on CCD sensors that go far beyond those of consumer scanning, and CDS becomes even more critical in these contexts. |
In digital radiography, a scintillator converts X-rays into visible light, which is then imaged by a CCD array. The X-ray dose delivered to the patient must be kept as low as reasonably achievable, which means the number of visible photons reaching each CCD pixel is limited. Every noise source in the readout chain therefore matters. Reset noise, if uncorrected, would force the system designer to increase the X-ray dose to maintain diagnostic image quality. By removing reset noise through CDS, the imaging chain can operate closer to the photon-shot-noise limit, allowing lower doses for the same contrast-to-noise ratio. This is not a marginal improvement; in fluoroscopy, where real-time X-ray video is used during procedures, the cumulative dose to the patient and staff is a primary safety concern, and every decibel of noise reduction in the readout chain translates into a tangible reduction in exposure. |
Endoscopy presents a different challenge. A modern endoscopic camera must fit a CCD array, its drive electronics, and its output amplifier into a package small enough to pass through a body cavity. The floating diffusion capacitance in such a device is tiny, which, as noted earlier, makes reset noise larger for a given charge fluctuation. CDS is therefore indispensable in endoscopy, where the small signal from a dimly lit mucosa must be read out reliably without adding noise that would obscure tissue detail. The circuit must also be compact and low-power, which has driven the integration of CDS into custom analog front-end ASICs designed specifically for endoscopic imagers. |
In microscopy, particularly fluorescence microscopy and live-cell imaging, the situation is again different. Fluorescent samples emit very few photons, and the biologist often needs to observe faint structures over time without damaging the sample with excessive illumination. Electron-multiplying CCDs, or EMCCDs, have become the sensor of choice for this work because they provide gain before the readout noise is added, effectively allowing single-photon detection. But even in an EMCCD, the initial charge-to-voltage conversion at the floating diffusion is subject to reset noise. CDS remains part of the readout chain, cleaning up the reset artifact before the signal is amplified and digitized. Without CDS, the very first stage of the readout would inject noise that no amount of subsequent gain could remove. |
|
5. Astronomy and Space Science: Where CDS Becomes Mission-Critical |
Nowhere is the importance of CDS more dramatically illustrated than in astronomy, where CCD sensors have become the workhorse detectors for telescopes on the ground and in space. |
An astronomical CCD camera may integrate for minutes or hours on a faint galaxy or a distant supernova. During that time, the sensor is cooled to temperatures that would make most electronics stop working, often below minus one hundred degrees Celsius, precisely to suppress dark current. The read noise of the output amplifier is the other major noise source, and it is measured in electrons. Modern scientific CCDs achieve read noise of a few electrons root mean square, a remarkable figure that is possible only because the readout chain is carefully optimized at every stage. |
CDS is a non-negotiable part of that optimization. When an astronomer measures the brightness of a star, the measurement is ultimately a count of electrons. If reset noise contributed even a few tens of electrons of uncertainty per pixel per readout, the measurement of a faint source would be swamped. By sampling the reset level and the signal level and subtracting, CDS reduces the reset contribution to a level where the amplifier's own noise and the photon statistics dominate. This is what allows astronomers to combine multiple exposures coherently and to detect sources that are millions of times fainter than the noise floor of a single exposure. |
The Hubble Space Telescope's CCD cameras, and more recently the detectors on the Euclid mission and the Vera Rubin Observatory's enormous camera array, all rely on this principle. In space, the stakes are even higher because the instrument cannot be serviced or upgraded after launch. The CDS circuit must work reliably for years in a radiation environment, with no opportunity for correction. The fact that CCDs with CDS readout have been the baseline for so many successful missions is a testament to the maturity and robustness of the technique. |
There is also a subtlety in astronomy that makes CDS particularly valuable: the signal levels are often so low that the CCD is operated in a mode where the read noise is the dominant uncertainty for faint pixels. In this regime, even a small reduction in read noise translates directly into a significant improvement in the limiting magnitude of the observation. CDS does not eliminate the amplifier's noise, but it eliminates the reset noise that would otherwise be the largest single contributor, allowing the amplifier noise to be minimized by other means, such as correlated double sampling with longer integration times or the use of electron-multiplying gain. |
|
6. Industrial Inspection and Machine Vision: Speed, Precision, and Reliability |
Industrial machine vision systems use linear and area CCD sensors for tasks that range from verifying the presence of a label on a pharmaceutical bottle to measuring the width of a steel strip to within a fraction of a millimeter. In these applications, the CCD sensor is often the last line of defense against a defective product reaching a customer, and the readout chain must be both fast and precise. |
In a high-speed inspection line, a linear CCD camera might read thousands of lines per second. Each line requires a reset and a readout for every pixel, and the CDS circuit must keep pace. At these speeds, the timing of the two samples becomes tighter, and the settling behavior of the output amplifier becomes more critical. The CDS circuit must sample the reset level before the signal charge arrives and sample the signal level before the next reset begins, all within the horizontal clock period. Designers of industrial CCD cameras spend considerable effort optimizing this timing so that the CDS subtraction remains accurate even at the maximum line rate. |
The benefit of CDS in this context is not merely noise reduction in the abstract. It is the ability to set a reliable threshold for defect detection. Suppose a machine vision system is inspecting a continuous sheet of material for dark spots. The threshold between 'acceptable' and 'defect' is set based on the expected signal from a dark spot against the background. If reset noise shifts pixel values randomly from line to line, the threshold must be set conservatively, and small defects may be missed or false alarms may be triggered by noise. CDS makes the per-pixel values more repeatable, allowing the threshold to be set closer to the true defect signal. In an industry where a missed defect can mean a product recall and a false alarm can mean unnecessary scrap, this improvement in signal fidelity has direct economic value. |
Linear CCDs are also used in spectroscopic instruments for industrial process monitoring, where the sensor measures the intensity of light at different wavelengths after it has passed through or interacted with a sample. Here, CDS contributes to the stability of the baseline, which is essential for detecting small changes in absorbance or emission. A drifting or noisy baseline limits the detection limit of the instrument, and CDS is one of the tools that keeps the baseline stable. |
|
7. The Evolution of CDS Implementation: From Discrete Components to System-on-Chip |
The history of CDS in CCD scanners mirrors the broader history of analog signal processing: a technique that began with discrete components and careful hand-tuning has gradually been absorbed into highly integrated, largely digital systems. |
In the 1980s and early 1990s, a CCD scanner's analog front end might have consisted of a discrete preamplifier, a pair of sample-and-hold circuits built from operational amplifiers and analog switches, and a difference amplifier, all laid out on a printed circuit board. The CDS timing was generated by a custom logic circuit or a microcontroller, and the designer's job was to ensure that the sample pulses landed in the right places relative to the CCD clocks. This required a deep understanding of the CCD's output waveform, including the feedthrough from the reset clock and the settling behavior of the source follower. |
As integration improved, CDS moved into dedicated analog front-end integrated circuits. These chips combined the sample-and-hold stages, the difference amplifier, a programmable gain amplifier, and an analog-to-digital converter in a single package. The CDS timing became a programmable feature, adjustable by the system designer to match the particular CCD's characteristics. This integration reduced component count, improved matching between the two sampling paths, and made CDS accessible to a wider range of scanner designs. |
More recently, the trend has been toward digital CDS, in which the reset level and the signal level are both digitized and the subtraction is performed numerically. This approach has the advantage of avoiding analog mismatch between the two sample-and-hold paths, and it allows the subtraction to be performed with essentially perfect linearity. The tradeoff is that the analog-to-digital converter must have enough resolution and speed to capture both levels accurately. In modern CMOS image sensors, which often use column-parallel analog-to-digital converters, digital CDS has become the standard approach. Even in CCD scanners, where the readout is typically serial, digital CDS is increasingly common because the required ADC performance is now readily available at low cost. |
For the designer of a barcode scanner or a scientific camera, this evolution has meant that CDS is no longer something to be designed from scratch but something to be configured. The underlying principle, however, remains exactly the same as it was in the discrete era: sample twice, subtract, and let the correlated noise cancel. |
|
8. When CDS Is Not Enough: Limitations and Complementary Techniques |
Correlated double sampling is a powerful technique, but it is not a universal solution to all noise problems, and a clear-eyed account of its limitations helps explain why CCD scanners also employ other noise-reduction strategies. |
First, CDS removes reset noise and suppresses low-frequency amplifier noise, but it does not remove the amplifier's white noise. The source-follower transistor at the CCD output has its own thermal noise, and that noise is not correlated between the two samples in the same way that reset noise is. In a well-designed CDS system, the amplifier noise becomes the dominant residual noise source, and further improvements require a lower-noise amplifier or a different readout architecture. |
Second, CDS is only effective if the two samples are taken at the right times. If the reset level sample is taken too early, before the reset transient has settled, the stored reference will be incorrect. If the signal sample is taken too late, after the next reset has begun, the signal will be corrupted. The timing windows are determined by the CCD's clocking scheme and the settling behavior of the output amplifier, and they become tighter as readout speeds increase. At very high speeds, the design of the CDS timing becomes a significant engineering challenge. |
Third, CDS does not address dark current or its shot noise. Dark current is generated by thermal excitation in the silicon itself, and it accumulates during the integration time. Cooling the sensor reduces dark current, but it does not eliminate it entirely. In scientific applications, dark current is typically calibrated out by taking dark frames, but the shot noise associated with dark current remains. CDS is orthogonal to this problem. |
Finally, in some modern CMOS image sensors, the reset noise is handled differently, for example by using a pinned photodiode with a fully depleted transfer gate that allows complete charge transfer and true correlated double sampling within the pixel. The CCD's floating diffusion CDS is a specific implementation of a more general principle, and designers working with different sensor technologies must adapt the principle to their particular device physics. |
These limitations do not diminish the importance of CDS. They simply place it in context: it is one essential stage in a readout chain that must be designed as a whole, with each stage addressing the noise sources that it is best suited to handle. |
|
9. Detailed Summary: CDS as the Quiet Enabler of CCD Performance |
To conclude this chapter, it is worth drawing together the threads of the discussion into a detailed summary of what CDS does, why it matters, and how it fits into the larger story of CCD scanners. |
Correlated double sampling is a readout technique that removes the reset noise, or kTC noise, from each pixel's signal by sampling the output amplifier's voltage twice: once at the reset reference level and once at the image signal level, and subtracting the first from the second. The reset noise, which is frozen onto the floating diffusion when the reset transistor turns off, appears identically in both samples and therefore cancels in the subtraction. The technique also suppresses low-frequency noise from the output amplifier because such noise is correlated between the two closely spaced samples. |
The importance of CDS arises from the physics of the CCD output node. The floating diffusion is a small capacitor, and its reset noise can be large enough to dominate the noise budget in applications where signal levels are low and precision matters. Without CDS, a barcode scanner would struggle to read poorly printed labels under dim illumination; a document scanner would produce grainy images; a medical X-ray system would require higher doses to achieve diagnostic image quality; an astronomical camera would lose sensitivity to faint sources; and an industrial inspection system would be forced to set conservative thresholds that miss small defects. |
Across these industries, CDS manifests in different ways but serves the same fundamental purpose. In barcode scanning and document imaging, it enables reliable binary decisions and clean image reproduction at modest cost. In medical imaging, it supports low-dose imaging and compact endoscope designs. In astronomy and space science, it is a mission-critical element of the readout chain that allows CCDs to approach their theoretical noise limits. In industrial inspection, it provides the repeatability needed for automated defect detection at high speed. |
The implementation of CDS has evolved from discrete sample-and-hold circuits to integrated analog front ends to digital subtraction in modern system-on-chip designs. The principle, however, has remained unchanged since the earliest CCD scanners: sample the reference, sample the signal, and let the correlated noise cancel. For the designer of a CCD-based scanner, CDS is not an exotic refinement but a standard and indispensable stage in the signal chain, a quiet enabler of the image quality that users have come to expect from devices ranging from handheld barcode readers to the great telescopes that survey the sky. |
cs@easiersoft.com