A method for determining noise events based on threshold mismatch and log-normal distribution.

Abstract

This invention discloses a method for determining noise events based on a threshold mismatch and log-normal distribution. The method includes the following steps: establishing an FPN noise model based on threshold mismatch using measured data of DVS image sensor pixels; calculating and determining the number of events and / or the distribution on the pixel array of the FPN noise in the DVS image sensor to be analyzed based on the FPN noise model and the parameters required for model calculation. Based on actual test results, this invention establishes an FPN noise model based on threshold mismatch. According to the FPN noise model, it analyzes the sources and distribution of noise, revealing the log-normal distribution characteristics of FPN noise in space. This can help researchers develop noise reduction methods based on noise sources, thereby assisting in improving the performance of dynamic vision sensors.

Description

Technical Field

[0001] This invention relates to the field of analog integrated circuit design technology, and in particular to a method for determining noise events based on threshold mismatch and log-normal distribution. Background Technology

[0002] Of the environmental information perceived by humans, visual information obtained through the human eye accounts for over 80%, and image sensors are electronic devices used to acquire image information. In the "frame"-based visual information processing mode, pixel array information is output for the entire frame regardless of whether the value of a pixel unit changes, resulting in a large amount of data redundancy.

[0003] The Dynamic Vision Sensor (DVS) abandons the concept of "frames," outputting only information about changing pixel units. Its logarithmic response stage circuit mimics the human eye's logarithmic response to light intensity, converting the input photocurrent logarithmically into an output voltage. When the light intensity voltage change exceeds a preset threshold voltage, the change in light intensity over that time period is encoded as pulse events of different polarities. It reads out dynamic information in the form of events, thus discarding static texture information. This event-driven encoding and readout method effectively reduces data redundancy.

[0004] However, the event generation mechanism remains susceptible to various noise sources. Temporal noise primarily consists of photon shot noise caused by quantum randomness and thermal noise generated by the thermal motion of charge carriers. These noises work together to form random background noise modulated by time variations and illumination intensity.

[0005] Furthermore, fixed pattern noise (FPN) originates from transistor threshold mismatch during manufacturing, resulting in spatially non-uniform responses that manifest as scene-independent response patterns. These combined noise effects complicate detailed event analysis and pose obstacles to effective noise reduction methods. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings and defects of existing technologies by providing a method for determining noise events based on threshold mismatch and log-normal distribution. By analyzing the root causes and mechanisms of log-normal FPN noise generation, a log-normal FPN noise model is established. This model is then used to determine noise events and their distribution, which is beneficial for researchers to develop denoising methods based on noise sources. This improves the spatial consistency of DVS event triggering and ultimately enhances image sensor performance.

[0007] This invention is achieved through the following technical solution: A method for determining noisy events based on threshold mismatch and log-normal distribution, comprising the following steps: Based on the measured data of DVS image sensor pixels, an FPN noise model based on threshold mismatch is established; Based on the FPN noise model and the parameters required for model calculation, the number of events and / or the distribution on the pixel array of the FPN noise of the DVS image sensor to be analyzed are calculated and determined.

[0008] Preferably, the FPN noise model based on threshold mismatch includes a model for determining the number of events that are positively correlated with the number of photoreceptor poles, the initial illumination intensity, and the rate of change of light intensity.

[0009] Preferably, the event number determination model includes: ; in, It is the rate of change of light intensity. p It is the principal pole of the photoreceptor. t 1 represents the initial light intensity I ph The time required for it to drop to 0 i Indicates the event trigger threshold. Let ln represent an event, and ln represent a logarithm.

[0010] Preferably, the dominant pole of the photoreceptor is modeled to be linearly related to the bias current, including: ; ; in, For the parasitic capacitance of the PD, The transconductance of the common-source amplifier driving transistor for the photoreceptor. Indicates the bias current. Thermoelectric voltage, This represents the subthreshold slope factor of NMOS.

[0011] Preferably, the bias current is modeled as a variable related to the threshold voltage, and its spatial distribution is modeled as a log-normal distribution, including: ; ; in, This is the ideal bias current under conditions of no mismatch. It is the mismatch threshold voltage. k p This is the subthreshold factor of PMOS. This represents the bias current under mismatch conditions. This represents a function with respect to the bias current. This indicates the bias current caused by threshold mismatch. The standard deviation of the mismatch.

[0012] Preferably, the FPN noise model based on threshold mismatch includes an expectation model and a standard deviation model, wherein the expectation of the event determined by the expectation model and the standard deviation of the event determined by the standard deviation model are linearly positively correlated with the square root of the initial light intensity.

[0013] Preferably, the desired model includes: ; in, Represents the expectation of a noise event. K It is a constant. Let be the standard deviation of the noise event on a logarithmic coordinate system.

[0014] Preferably, the K For the rate of change of light intensity 、 From the initial light intensity I ph Time required to drop to 0 t 1. Event threshold i The constants that determine this include: K = ( k n t 1 / 2 C PD U T |θ| ) 1 / 2 .

[0015] Preferably, the standard deviation model includes: ; in, The standard deviation of a noise event.

[0016] Preferably, the parameters required for the model calculation include at least the initial light intensity. I ph Rate of change of light intensity Process parameters, mismatch parameters, and circuit parameters; the process parameters include the subthreshold slope factor. k n Thermoelectricity U T Parasitic capacitance of PD C PD Event trigger threshold i The mismatch parameter includes the standard deviation σ of the threshold voltage mismatch. V th The circuit parameters include the ideal bias current. Ibias0 Light intensity change time t 1.

[0017] Based on actual test results, this invention establishes an FPN noise model based on threshold mismatch. According to the FPN noise model, the source and distribution of noise are analyzed, revealing the log-normal distribution characteristics of FPN noise in space. This provides assistance to researchers in developing noise reduction methods based on noise sources, thereby providing support for improving the performance of dynamic vision sensors. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the pixel array structure of an event-based DVS image sensor.

[0019] Figure 2 This is a schematic diagram of the DVS pixel unit structure.

[0020] Figure 3 It is a spatial mismatch distribution map of DVS pixel event triggering under bright light.

[0021] Figure 4 It represents the spatial distribution of events under different initial illumination conditions with 100% change in light intensity, as well as the relationship between model parameters and illumination intensity.

[0022] Figure 5 It represents the spatial distribution of events under different proportions of light intensity change at a constant initial light intensity, as well as the relationship between model parameters and the rate of change of light intensity. Detailed Implementation

[0023] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0024] Due to differences in wafer fabrication processes, transistor mismatch occurs in the pixel array of DVS pixels, a mismatch typically modeled as threshold mismatch. Threshold mismatch not only causes non-uniformity between pixels but also generates FPN noise. Therefore, based on the characteristics of biomimetic vision imaging, this invention proposes a log-normal FPN noise model based on threshold mismatch. This log-normal FPN noise exhibits an asymmetric log-normal distribution in space. To specifically study the impact of transistor mismatch on DVS pixel event triggering, this invention starts with the bias MOS transistor of the pixel unit photoreceptor, analyzing the spatial distribution characteristics of threshold mismatch in the bias MOS transistor and the resulting spatial distribution of bias current. The influence of these two factors on the spatial distribution characteristics of event triggering is investigated, leading to the proposal of a log-normal FPN noise model based on threshold mismatch. Based on this model, FPN noise events and their distribution can be calculated and determined. In other words, by analyzing the source and distribution of noise using the proposed FPN noise model, the spatial log-normal distribution characteristics of FPN noise are revealed.

[0025] In an exemplary embodiment of this application, the method for determining noise events based on threshold mismatch and log-normal distribution includes the following steps: Based on the measured data of DVS image sensor pixels, an FPN noise model based on threshold mismatch is established; Based on the FPN noise model and the parameters required for model calculation, the number of events and / or the distribution on the pixel array of the FPN noise of the DVS image sensor to be analyzed are calculated and determined.

[0026] The log-normal FPN noise model based on threshold mismatch of the present invention is proposed after data analysis of the principle of FPN noise generation in DVS image sensor pixels and the spatial performance of FPN noise. It also proposes the control variables of the corresponding FPN noise model, providing a calculation basis for subsequent determination of the number of FPN noise events and / or the distribution on the pixel array.

[0027] The parameters required for the model calculation include at least the initial light intensity. I ph Rate of change of light intensity Process parameters, mismatch parameters, and circuit parameters; the process parameters include the subthreshold slope factor. k n Thermoelectricity U T Parasitic capacitance of PD C PD Event trigger threshold i The mismatch parameter includes the standard deviation σ of the threshold voltage mismatch. V th The circuit parameters include the ideal bias current.I bias0 Light intensity change time t 1.

[0028] The following analysis, combining the event-based DVS image sensor pixel array structure and DVS pixel unit structure, explains the model modeling of the embodiments of this application.

[0029] Figure 1 A schematic diagram of an event-based DVS pixel array structure is provided. This event-based DVS pixel array structure consists of pixel array 1, DVS pixel units 2, high-speed interface circuit 3, ESP pipeline interface 4, readout control circuit 5, row selector 6, and Y-direction address encoder 7. For example, pixel array 1 is a 1280*720 array. Pixel array 1 senses changes in light intensity and converts these changes into events, forming a sparsely encoded output. When a single pixel unit generates an event, it transmits its Y-direction address to the Y-direction address encoder 7 via a request signal, and then transmits the signal to the row selector 6. After the response signal returns to the row containing the pixel unit, the pixel unit in that row outputs the event information. The information is read out as parallel data through the high-speed interface circuit 3, and the parallel data is read out serially through the Event Signal Processing (ESP) pipeline interface 4. The timing signals during the readout process are implemented by the readout control circuit 5. However, due to the differences in the manufacturing process of pixel array 1 on the wafer, there will be slight differences in the pixel units. These differences will cause event triggering mismatch between pixel arrays, which in turn will lead to spatial FPN noise.

[0030] Figure 2 A schematic diagram of the event-based pixel structure is given. The pixel structure consists of photodiode 21, logarithmic diode 22, bias transistor 23, capacitor 24, capacitor 25, operational amplifier 26, threshold comparator 28 (ON comparator), threshold comparator 32 (OFF comparator), current source 29, transistor 30, transistor 31, transistor 33, transistor 35, and the parasitic capacitance of the PD. C PD constitute.

[0031] Among them, photodiode 21, logarithmic diode 22, bias transistor 23, transistor 35, and the parasitic capacitance of PD. C PD The photosensitive sensor is composed of a differential capacitor amplifier consisting of operational amplifier 26, capacitors 24 and 25, and transistor 33. The threshold comparator 28 (ON comparator) and threshold comparator 32 (OFF comparator) constitute the comparator. The reset module consists of current source 29, transistor 30, and transistor 31.

[0032] The specific connection is as follows: one end of photodiode 21 is grounded, and the other end is connected to the parasitic capacitance of PD. C PD One end of the transistor, the source of the logarithmic transistor 22, and the gate of the transistor 35; parasitic capacitance. C PD The other end is grounded, the drain of logarithmic transistor 22 is connected to VDD, and its gate (outputs the logarithmic output voltage of the photoreceptor) is connected to VDD. V pr The capacitor 24 of the differential capacitor amplifier is connected to the drain of bias transistor 23, the drain of transistor 35, and one end of capacitor 24. The source of transistor 35 is grounded, and the source of bias transistor 23 is connected to VDD. The other end of differential capacitor amplifier 24 is connected to one end of capacitor 25, the input of op-amp 26, and the drain of transistor 33. The other end of capacitor 25 is connected to the output of op-amp 26 and the source of transistor 33, and is connected to the input of comparator. The gate of reset transistor 33 is connected to the reset voltage. The outputs of threshold comparator 28 (ON comparator) and threshold comparator 32 (OFF comparator) of the comparator output the comparison result, i.e., the output of the ON / OFF event.

[0033] In this application, in the reset module, the power supply 29 outputs current, and the source of transistor 30 is connected to the output terminal of the current source and one end of a capacitor, i.e., connected to the control terminal. V rst The other end of the capacitor is connected to VDD, the drain of transistor 30 is connected to the drain of transistor 31, the source of transistor 31 is grounded, and the gates of transistors 30 and 31 are connected to the response signals RA and CA, respectively.

[0034] The photodiode 21 converts the light intensity signal into a photocurrent. I ph The photocurrent is converted by the logarithmic diode in the logarithmic diode 22. I ph The voltage is converted into a logarithmic voltage and connected to a common-source stage composed of transistors 23 and 35 for amplification; the differential capacitor amplifier is composed of operational amplifier 26, capacitor 24 and capacitor 25, which samples and amplifies the input signal, i.e., photovoltage, and its amplification factor is determined by capacitors 24 and 25.

[0035] Two threshold comparators, 28 and 32, are used to perform ON / OFF event detection. Specifically, if the current differential capacitor output voltage... V diff If the voltage exceeds the threshold voltage of threshold comparator 28, the pixel unit outputs an ON event. If the current differential capacitor output voltage... V diff If the threshold voltage of the threshold comparator 32 is exceeded, the pixel unit outputs an OFF event.

[0036] After the event occurs, the reset structure controlled by current source 29 will perform a four-way handshake protocol, and transistors 30 and 31 will jointly control the reset voltage. V rst The levels of the response signals. After the row and column responses are completed, both the response signals RA (row response signal) and CA (column response signal) change from low to high. V rst Pulled down to low level. Differential capacitor amplifier through V rst The controlled transistor 33 has completed its reset; the next event recognition is now enabled.

[0037] Due to differences in the wafer fabrication process, bias transistor 23 will experience transistor mismatch in the pixel array, which is generally modeled as threshold mismatch. This threshold mismatch exhibits a normal distribution characteristic in the pixel array. Clearly, the threshold mismatch of bias transistor 23 will cause bias current... I bias The mismatch. Because bias transistor 23 operates in the subthreshold region, the bias current... I bias With threshold voltage V th The exponential change in bias current. I bias The variables modeled as being related to the threshold voltage are spatially distributed according to a log-normal distribution. The model parameters are shown in equations (1) and (2) below.

[0038] (1) (2) in I bias0 It is the ideal bias current under the condition of no mismatch, Δ V th It is the mismatch threshold voltage, where k p This is the subthreshold factor of PMOS. U T This is thermal voltage. I bias0 The exponential term related to the threshold voltage will account for a small change in the threshold voltage σ( V th ) Magnified to be significant I bias Mismatch.

[0039] Figure 3A spatial mismatch distribution map of DVS pixel event triggering under bright light is presented, illustrating the behavior of the FPN noise proposed in this invention, including the macroscopic manifestation of FPN noise on the pixel array and the model characteristics of the event quantity. Through modeling analysis, the statistical distribution of the event quantity follows a log-normal distribution. (Reference) Figure 3 The curve represents the model fitting data. Due to the high light intensity, the bias current... I bias This directly affects the principal pole of the photoreceptor, and the two are modeled as a linear relationship, as shown in equation (3).

[0040] (3) (4) in, C PD For the parasitic capacitance of the PD, g m1 For the transconductance of transistor 35, k n For NMOS, the subthreshold factor is the bias current. I bias The linear model of the dominant pole leads to bandwidth variation and I bias The same statistical behavior propagates, and this mechanism directly results in differences in the photoreceptor bandwidth of different pixels in the pixel array, thus causing a mismatch in the number of events triggered by pixel units. The number of events is modeled as a positive correlation with the poles, the initial illumination intensity, and the rate of change of light intensity, i.e., the event determination model of this application, as shown in Equation (5).

[0041] (5) in, It is the rate of change of light intensity. p It is the principal pole of the photoreceptor. t 1 represents the initial light intensity I ph The time required for it to drop to 0 i This represents the threshold of the event. In practical measurements, this difference manifests as spatial FPN noise. By placing the chip under uniform illumination and varying the illumination intensity, we modeled and analyzed the number of events and the corresponding number of pixel units on the pixel array. The model revealed the spatial log-normal distribution characteristics of the FPN noise.

[0042] Figure 4 The distribution of events on the pixel array under different initial illumination intensities when the light intensity changes by 100%, and the relationship between model parameters and illumination intensity are presented. The points in the figure represent actual measured data, and the curves represent model-fitted data. Changes in the principal pole of the photoreceptor affect the logarithmic voltage. Vpr The output curve of the light intensity affects the actual triggering time of the event. The model can determine the number of events generated by each pixel unit when the light intensity changes linearly. n event It is positively correlated with the square root of the initial light intensity and the square root of the principal pole.

[0043] Since the dominant pole is modeled with respect to the threshold voltage V th The number of events triggered by each pixel unit follows a log-normal distribution. Therefore, the number of pixel units corresponding to the number of events triggered by each pixel unit is modeled as a log-normal distribution on the pixel array. Furthermore, as the initial light intensity increases, the number of events generated by each pixel unit also increases. n event As the light intensity increases, the model curve shifts to the right. The model's expectation and standard deviation are linearly positively correlated with the square root of the initial light intensity. Therefore, as the light intensity increases, the number of events... n event The expected value and standard deviation are increased to obtain the expected value model and standard deviation model of this application, as shown in equations (6) and (7) below.

[0044] (6) (7) in, K = ( k n t 1 / 2 C PD U T |θ| ) 1 / 2 It is a product of , t 1 and i The constant that determines the outcome.

[0045] By fitting the model to the test data, the coefficients of determination (COD) of all fitted curves of the model exceeded 95%.

[0046] It is important to note that the model of the photoreceptor's principal pole changes when the chip is in a low-light environment. (Photocurrent) I ph Replace bias current I bias The photocurrent is considered a major factor affecting the principal pole of a photoreceptor. I ph There is no log-normal distribution characteristic in space, therefore the number of events n eventThe test data deviates from the log-normal distribution model, and the log-normal distribution characteristics are weakened.

[0047] Figure 5 The distribution of events on a pixel array under different light intensity change rates with the same initial illumination intensity, and the relationship between model parameters and the light intensity change rate, are presented. The points in the figure represent actual measured data, and the curves represent model-fitted data. The model can be used to determine the number of events triggered in each pixel unit. n event、 It is still related to the dominant pole of the photoreceptor in that pixel unit. Therefore, the number of events in the model... n event It still maintains the distribution characteristics of a log-normal distribution in space.

[0048] However, according to equation (5), when the initial illumination intensity is constant, the number of events triggered by each pixel unit is... n event、 The number of events triggered is directly proportional to the rate of change of light intensity. Different rates of change of light intensity will inevitably result in differences in the number of events triggered.

[0049] According to the models shown in equations (6) and (7), the expected value and standard deviation of the normal distribution model curve both show a monotonically increasing trend as the rate of change of light intensity increases. This corresponds to the rightward shift and dispersion of the test data as the rate of change of light intensity increases. The COD of all fitted curves of the model exceeds 95%. This fully demonstrates the reliability of the log-normal spatial fixed-mode noise model based on threshold mismatch.

[0050] Therefore, through the model, it is possible to clearly calculate the specific magnitude of this spatial fixed-mode noise caused by threshold mismatch in terms of the number of events and its specific distribution on the pixel array.

[0051] Example of pixel noise calculation: The test was conducted under uniform lighting conditions, with the following specific conditions: (1) Illumination conditions: initial light intensity is 10000 lux, and the rate of change of light intensity is... =100%.

[0052] (2) Process parameters: Using a typical 110nm CMOS process, the subthreshold slope factor of NMOS is... k n =0.8, thermal voltage U T =26mV, parasitic capacitance of PD C PD =100fF, event trigger threshold i =300mV.

[0053] (3) Mismatch parameter: Standard deviation of threshold voltage mismatch σ(V th =5mV.

[0054] (4) Circuit parameters: ideal bias current I bias0 =128nA, time of light intensity change t 0 = 10ms, due to the rate of change of light intensity =100%, therefore t 1 = 10ms.

[0055] According to equation (1), σ( ) can be obtained through process angle simulation or analytical model. I bias =1.53. Then calculate... n event The statistical parameters are obtained by substituting the parameters into formulas (5) and (6). K =3.58×10 5 ,σ( n event =0.765 n event The expected number of events is 168, and the standard deviation is approximately 150 events. Verification showed that the test results using the calculated data obtained from the model were in perfect agreement, sufficiently demonstrating the reliability of the model presented in this application.

[0056] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. It will be apparent to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and that the present invention can be implemented in other specific forms without departing from the spirit or basic features of the present invention. Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, it is intended that all variations falling within the meaning and scope of the equivalents of the claims be included within the invention.

[0057] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for determining noise events based on a threshold mismatch log-normal distribution, characterized in that, Includes the following steps: Based on the measured data of DVS image sensor pixels, an FPN noise model based on threshold mismatch is established; Based on the FPN noise model and the parameters required for model calculation, the number of events and / or the distribution on the pixel array of the FPN noise of the DVS image sensor to be analyzed are calculated and determined.

2. The method for determining noise events based on threshold mismatch and log-normal distribution according to claim 1, characterized in that, The threshold mismatch-based FPN noise model includes a model for determining the number of events that are positively correlated with the number of photoreceptor poles, initial illumination intensity, and rate of change of illumination intensity.

3. The method for determining noise events based on threshold mismatch and log-normal distribution according to claim 2, characterized in that, The event number determination model includes: ； in, It is the rate of change of light intensity. p It is the principal pole of the photoreceptor. t 1 represents the initial light intensity I ph The time required for it to drop to 0 θ Indicates the event trigger threshold. Let ln represent an event, and ln represent a logarithm.

4. The method for determining noise events based on threshold mismatch and log-normal distribution according to claim 3, characterized in that, The dominant pole of the photoreceptor is modeled to be linearly related to the bias current, including: ； ； in, For the parasitic capacitance of the PD, The transconductance of the common-source amplifier driving transistor for the photoreceptor. Indicates the bias current. Thermoelectric voltage, This represents the subthreshold slope factor of NMOS.

5. The method for determining noise events based on threshold mismatch and log-normal distribution according to claim 4, characterized in that, The bias current is modeled as a variable related to the threshold voltage, and its spatial distribution is modeled as a log-normal distribution, including: ； ； in, This is the ideal bias current under conditions of no mismatch. It is the mismatch threshold voltage. κ p This is the subthreshold factor of PMOS. This represents the bias current under mismatch conditions. Represents a function with respect to bias current ， This indicates the bias current caused by threshold mismatch. The standard deviation of the mismatch.

6. The method for determining noise events based on threshold mismatch and log-normal distribution according to claim 5, characterized in that, The threshold mismatch-based FPN noise model includes an expectation model and a standard deviation model. The expectation of the event determined by the expectation model and the standard deviation of the event determined by the standard deviation model are linearly positively correlated with the square root of the initial light intensity.

7. The method for determining noise events based on threshold mismatch and log-normal distribution according to claim 6, characterized in that, The desired model includes: ； in, Represents the expectation of a noise event. K It is a constant. Let be the standard deviation of the noise event on a logarithmic coordinate system.

8. The method for determining noise events based on threshold mismatch and log-normal distribution according to claim 7, characterized in that, The K For the rate of change of light intensity 、 From the initial light intensity I ph Time required to drop to 0 t 1. Event threshold θ The constants that determine this include: K = ( κ n t 1 / 2 C PD U T |θ| ) 1 / 2 。 9. The method for determining noise events based on threshold mismatch and log-normal distribution according to claim 8, characterized in that, The standard deviation model includes: ； in, The standard deviation of a noise event.

10. The method for determining noise events based on threshold mismatch and log-normal distribution according to claim 1, characterized in that, The parameters required for the model calculation include at least the initial light intensity. I ph Rate of change of light intensity Process parameters, mismatch parameters, and circuit parameters; the process parameters include the NMOS subthreshold slope factor. κ n Thermoelectricity U T Parasitic capacitance of PD C PD Event trigger threshold θ The mismatch parameter includes the standard deviation σ of the threshold voltage mismatch. V th The circuit parameters include the ideal bias current. I bias0 Light intensity change time t 1.

Images