Ulsan National Institute of Science and Technology in partial fulfillment of. requirements for the Master of Science degree. -Sug Chung: Thesis Committee Member #2. Chapter 2: Image sensor noise reset.. kTC noise in 3-T image sensor .. kTC noise in 4-T image sensor .. kTC noise in I-ToF sensor. With his passion for image sensing and for BIASians, I can take a step forward of knowledge on what I have learned.
Also, from his own experience and dedication, he gave me fewer words about the relationship than an engineer. Suhyun, who is always passionate about her work, makes me passionate about work.
Abstract
Introduction
Low Noise Image Sensor
However, in realistic case, the received light may be overwhelmed by noise from the I-ToF sensor itself. When the signal is low under the low light intensity, the readout noise from image sensor will be dominant, not photon shot noise which is proportional to the square root of the number of incident photons. Breaking this limit below the readout noise dominant portion lowers the red dashed line.
It is also good for the I-ToF sensor that the depth accuracy is increased, which can lead to a more accurate depth image. That is why the role of noise reduction is becoming increasingly important in dealing with readout noise by using adequate circuit techniques.
Noise Image Sensor
2-1. Fixed pattern noise
All these fixed pattern noise components are also shown in the image as patterns as in figure 1-6. Due to its constant property, we can know the exact value of FPN quantity at any point by measurement or image. To eliminate this FPN, we can do the digital processing step which is known as ISP image signal processing.
So many images sampled from the dark state can average the value of each FPN at each point.
2-2. Temporal noise
This property of flicker noise allows us to easily suppress this noise by double sampling. But this is not an effective way to read all pixel arrays in the sensor. Read the reset value immediately after the integration is completed or just before the readout signal value.
Since the double sampling frequency is much faster than the corner frequency, we can treat the jitter noise almost as a DC value. By using high-resolution ADCs, the noise level due to quantization noise becomes lower than the noise level from other temporal noise contributors.
Source # of Noise e- Relative value
Reset Noise in Image Sensor
Pixel operation
Since we have modeled photodiode as a capacitor, we can sample the voltage by applying voltage to the gate of the reset transistor. And this plot shows that we cannot sample the clear reset voltage because of the random thermal noise. Each time we try to sample the reset voltage, we are actually sampling the different reset voltages in different time which is expressed as this VRST ≠ V(t1) ≠ V(t2) ≠ V(t3).
Conventional reset operation 2-1. Hard reset
16 In the steady state, the applied voltage to the reset transistor would follow this condition VGS – Vth = α, and VDS = 0. In this state, no more current flows through the reset transistor and photodiode. Initially, a hard reset can be done via the current on the photodiode via the reset transistor.
This allows the current to have a short reset time of a few nanoseconds, and it also ejects any remaining electrons in the photodiode, no matter how many electrons are left. To analyze hardware reset noise, an equivalent small-signal model is required, as shown in Figure 2-4. Since the reset transistor RST operates in the triode region, the reset transistor RST has a resistance in the triode region.
The equivalent small signal therefore has thermal noise voltage source Vn, noiseless resistance and photodiode capacitance CPD. First, we need to determine the transfer function of Vn and VPD in frequency domain. By simple voltage division of noiseless R and CPD, we can get the noise transfer function as follows [1].
As we already knew that the resistor noise voltage is 4kTR, what we need to do is to integrate the square of this transfer function over the entire frequency domain. It also has quite high noise power due to the small capacitance of the photo diode. This is why reset thermal noise is the most dominant term in the image sensor compared to other thermal noise contributors.
2-2. Soft reset
18 Normally the reset operation would terminate until the photodiode capacitor and reset transistor RST have come to rest. To meet this time constraint, the soft reset is performed in non-steady state within the allowed time, as shown in Figure 2-6. As mentioned earlier, the noise analysis of soft reset would be performed in the time domain since the reset operation is performed in a non-steady state where time varies.
Then we can establish time domain equation that only includes the drain current, sampled noise voltage and reset voltage of the photodiode and neglects other current source [1]. The results show that the sampled noise effect in photodiode is stopped at kT/C. Implementation of the soft reset must take into account which one is dominant with image delay problem.
Normally, the 3-T image sensor integrates light and reads the signal which is converted to voltage after the photodiode is then reset and reads the reset voltage. These two operations give us (Vreset - Vsignal) which is displayed as an image on a single pixel. During integration, the 3-T image sensor has reset thermal noise as kT/C1 that is caused by the RST1 reset operation.
When we sample the correlated reset signal with reset noise, the image sensor must cue the frame memory to store the reset voltage for as long as a few milliseconds. A 4-T image sensor consists of 4 transistors and has much the same pixel noise problems as Figure 2-10. Collected electrons are transferred to the floating diffusion node called the FD node by TX.
As before the electron transfer, first reset the FD node and reset the FD node reading with kTC reset noise. And then the charges are transferred to this noise kTC and the read signal of node FD. Immediately after RST1 at the FD node TX helps transfer electrons from the PPD to the FD node.
When the modulated light enters the PPD, synchronized with the modulated light TX1 and TX2 transfer electrons to each FD1 node and FD2 node. Using the voltage difference of two FD nodes, we can detect phase difference of the modulated signal. Since we do the depth calculation based on this voltage, I-ToF sensor also suffers from the same kTC noise.
Proposed Active Reset Technique for Reset Noise Suppression
1.Active reset
Time domain noise analysis of active reset
To theoretically analyze active reset noise, we need to do a time domain noise analysis. Because the sampled kTC noise on the photodiode itself would decrease over time due to negative feedback. Then current would flow from the RG reset gate to the photodiode with no noise and current from the output node of the amplifier.
The current equation, expressed as Ohm's law, can be replaced as a subthreshold current equation, as below. 25 Because the equation has two parts, the average term and the noise term, the average voltage of equation (2) will be equal. The average term is a noiseless term that is the equation without noise current in(t). Now we can replace equation (4) with the ideal equation.
As mentioned above, the voltage term can be divided by the noise term as VPD(t) = VPD_a(t) + vpd(t). Since it is noise power, the actual reduced noise voltage is proportional to the square root of the amplifier gain. The assumption for solving the equations in the time domain is that the feedback circuit has ideal source followers.
This means that the reset noise kTC is suppressed by the negative feedback loop gain. With the results from time domain noise analysis, we can get estimation results from the expected values of typical image sensors. Compared with the previous kTC noise, the noise power has decreased as the gain of the amplifier, which is 60dB, has decreased 1000 times.
Frequency domain noise analysis of active reset
Typical way to analyze each noise contribution to the photodiode is done by superposition method. But for simplicity, the noise analysis will be performed as all noise source is merged with the noise source of the amplifier. This method cannot give us exact noise ratio of each component and gives us only approximate noise.
Since all the noise source is lumped into the amplifier noise, the noise source can be expressed in the amplifier noise current as shown in Figure 3-7. 28 Substituting equation (ii) into equation (i), we can obtain the noise transfer function from the amplifier to the photodiode. The significance of this evaluation result is that the noise of additional components is much smaller than the suppressed noise.
The amplifier, the source follower, the thermal noise of the RG reset gate contribute to a small part of the total noise in the photodiode.
Proposed pixel level active reset sensor
Measurement
Measurement setup
Measurement results 2-1. ROIC noise
2-2. Hard reset noise
2-3. Active reset noise
Since hard reset has greater noise than the active reset, hard reset shows greater voltage deviation. Hard reset noise is estimated as 840uVrms and it can be converted to noise electron using equation QN=CPD x VN in FD node. In the same way, the measured number of noise electrons is 34 electrons for hard reset and active reset is 2.6 electrons.
The equation QN = CPD x VN is a linear equation, the suppression rate of noise electrons is the same in the voltage.
Conclusion
36 Theoretically, the sampled thermal noise on the photodiode or FD node is suppressed by approximately 96% compared to the kTC noise with hard reset. And the noise from the amplifier, source follower and reset gate switch is quite small compared to suppressed noise which can be neglected. The measurement also suppresses the sampled thermal noise at the FD node by 92% compared to the measured hard reset noise.
In conclusion, thanks to the active reset, the sampled noise can be suppressed up to 96% in theory and 92% in measurement.
Further works
