Random telegraph signal noise analysis method for image sensor

Abstract

The image sensor random telegraph signal noise analysis method comprises the following steps: constructing a standard test database; analyzing each pixel signal in the standard test database by adopting different RTS noise detection analysis methods to obtain RTS parameters; comparing the RTS parameters obtained by each analysis method with data tags in a standard test database, and calculating the accuracy of the RTS parameters of each analysis method; for each RTS parameter, judging whether the accuracy rate of the RTS parameter in each analysis method is greater than or equal to a threshold value or not, and accordingly determining an analytic method which can be acquired by each RTS parameter; and inputting an image sensor pixel signal to be analyzed, judging whether the image sensor pixel signal is an RTS noise pixel signal or not by adopting an acquirable analysis method, if not, ending, if so, analyzing each characteristic parameter by adopting the acquirable analysis method, taking a statistical result of analysis results of different acquirable analysis methods as a final result, and determining each characteristic parameter according to the final result.

Description

Technical Field

[0001] The invention relates to the field of image sensor performance testing, and in particular to a method for analyzing random telegraph signal noise of an image sensor. Background Art

[0002] As a core key component, image sensors are widely used in meteorological satellites, communication satellites, earth resources satellites, and automotive automatic driving modules. As a core component, its reliability directly determines the operating life and working performance of the entire equipment. Due to the increase in usage time or harsh working environment, such as the presence of a large number of high-energy radiation particles or high temperature environment in space, the performance of image sensors will gradually degrade. One of the most important failure modes is the appearance of a large number of random telegraph signal (RTS) noise. RTS noise will cause the pixel output grayscale value to not remain constant, but to fluctuate randomly between two or more discrete levels over time.

[0003] In order to eliminate the interference of RTS noise on the image, it is necessary to detect and analyze the RTS noise from the data collected by the image sensor. Currently, the following methods have been developed to detect RTS noise:

[0004] The most commonly used is the multi-level RTS automatic detection method based on time domain edge detection technology. This method was first proposed in the IEEE article "Multilevel RTS in Proton Irradiated CMOS Image Sensors Manufactured in a Deep Submicron Technology" in 2009. This detection method first convolves the pixel signal in the time domain with a filter to obtain the gradient signal in the time domain, and then obtains the filter threshold through the signal and noise variance, and uses this threshold to filter and shape the gradient signal, divide the RTS signal segments of different orders, and then reconstruct. However, this method does not remove white noise in advance, which will increase the difficulty of subsequent signal processing and affect the accuracy of RTS noise detection. The gradient threshold selection of this detection method is only obtained by calculating the signal variance, which greatly reduces the scope of application of the detection method. Therefore, in order to improve the accuracy and scope of application of RTS noise detection and make it easier to analyze RTS noise, it is urgent to develop a convenient, fast, accurate and widely applicable RTS noise detection and analysis system.

[0005] In addition, the commonly used RTS noise detection and analysis methods include visual counting method, threshold-based method, RTN parameter extraction based on histogram Gaussian fitting, detection based on histogram Gram-Charlier series expansion, statistical characteristics analysis, noise scatter plot, etc.

[0006] Among them, the biggest disadvantage of the visual counting method and the threshold-based method is that they rely on the operator to give an external threshold input for judgment. Although the judgment speed is fast, it cannot be run automatically. Although the two histogram-based methods have a fast judgment speed and can be run automatically, there is no way to determine the time constant of RTS and cannot reconstruct the RTS signal. In addition, the histogram-based judgment method is easily affected by the low-frequency drift of the signal, resulting in misjudgment and reduced detection efficiency. RST judgment requires a long sampling time and often has the problem of low-frequency drift.

[0007] Patent CN104853117B discloses a Matlab-based image sensor RTS noise detection and analysis system, which uses a method of filtering white noise first and then performing multi-level RTS automatic detection based on time domain edge detection technology, which improves the accuracy of RTS noise detection and the analysis of all parameters to a certain extent. However, in actual use, since RST noise is often coupled with white noise, signal low-frequency drift and hot pixels, the analyzed data is still questionable.

[0008] In the 2008 Journal of Xidian University, a method for measuring random telegraph signal noise of nano-MOS devices based on virtual instruments was proposed in the article "Measurement and Analysis of RTS Noise of Nano-MOS Devices". The relative amplitude of random telegraph signal noise was extracted by point-by-point difference and Gaussian function fitting method based on the time series of random telegraph signal noise. The time constant of random telegraph signal noise was then extracted by digital filtering and exponential function fitting method. This method is much faster than traditional methods and requires only 1 / 10 of the sampling points of traditional methods under the same accuracy requirements. When the time constant of random telegraph signal is small, the measurement accuracy is several to dozens of times that of traditional methods. However, due to the nonlinear changes and measurement errors of samples caused by sampling and quantization, the point-by-point difference value needs to be extracted by function fitting method to extract the point with the highest probability. This method is also easily affected by low-frequency drift of the signal.

[0009] The challenges encountered by existing image sensor random telegraph signal noise analysis methods are:

[0010] 1) RTS noise is usually coupled with white noise, low-frequency drift of the signal, and hot pixels. The signal is complex, and a large amount of data is required to fully characterize various types of RTS noise. In the literature, only ideal RTS noise models or a very small amount (about 50 groups) of real test data are usually analyzed and explained. The samples commonly used in the literature are difficult to fully cover real scenarios, which poses a challenge to the effectiveness of the performance test of the RTS analysis method;

[0011] 2) There are many existing methods for analyzing random telegraph signal noise of image sensors, each with its own advantages and disadvantages, and no method can meet all the analysis needs for RTS noise;

[0012] 3) There is a lack of a unified evaluation method to identify the capabilities of various methods in different directions, so as to provide theoretical support for the random telegraph signal noise analysis method coordinated by multiple analysis methods;

[0013] 4) Machine learning is developing rapidly and has shown its unique advantages in data analysis in many fields. The introduction of machine learning in the field of random telegraph signal noise analysis of image sensors also has great potential. However, due to the large number of feature labels in random telegraph signal noise analysis and the complex connections between different feature labels, there is currently no suitable evaluation system for the analysis model trained by machine learning, making it difficult to train high-quality models. Summary of the invention

[0014] The object of the present invention is to provide an image sensor random telegraph signal noise analysis method, realize random telegraph signal noise analysis coordinated with multiple RTS noise detection and analysis methods, integrate the advantages of multiple RTS noise detection and analysis methods, and improve the analysis accuracy.

[0015] In order to achieve the above-mentioned purpose, the present invention provides an image sensor random telegraph signal noise analysis method, comprising: 1) constructing a standard test database based on different types of long-time pixel signals collected and sorted out through extensive experiments; 2) using existing different types of RTS noise detection and analysis methods to analyze each pixel signal in the standard test database to obtain the RTS parameters of the pixel signal; the RTS parameters include whether it is an RTS noise pixel signal, the maximum step amplitude, the average step interval time, the RTS level order, the signal drift value, and the pixel signal reconstructed after denoising and containing only RTS noise; 3) analyzing and comparing the RTS parameters of the pixel signal obtained by each RTS noise detection and analysis method with the data labels in the standard test database, and calculating the accuracy between the RTS parameters of the pixel signal obtained by each RTS noise detection and analysis method and the data labels in the standard test database; 4) for each RTS parameter, judging each RTS noise Whether the accuracy of the RTS parameter of the acoustic detection and analysis method is ≥ the threshold value, if so, the RTS noise detection and analysis is determined to be the analysis method of the RTS parameter, if not, the RTS noise detection and analysis is not used to analyze the RTS parameter, and the credible RTS noise detection and analysis method for each RTS parameter is determined accordingly; 5) input the image sensor pixel signal to be analyzed, and use the credible RTS noise detection and analysis method determined in step 4) to determine whether the image sensor pixel signal to be analyzed is an RTS noise pixel signal, if so, execute step 6), if not, output the image sensor pixel signal as a non-RTS noise pixel signal, and end the process; 6) for each characteristic parameter, use the credible RTS noise detection and analysis method for the characteristic parameter determined in step 4) to analyze the image sensor pixel signal to be analyzed, take the statistical results of the analysis results of different credible RTS noise detection and analysis methods as the final result, and determine each characteristic parameter accordingly.

[0016] The above-mentioned image sensor random telegraph signal noise analysis method, wherein the step 1) includes: 1-1) collecting and collating different types of long-time pixel signals as initial data through extensive experiments; 1-2) manually determining whether each pixel signal in the initial data is an RTS noise pixel signal, if so, executing step 1-4), if not, executing step 1-3); 1-3), using the determination result of whether it is an RTS noise pixel signal as a data label, and adding the non-RTS noise pixel signal and its data label to a standard test database; 1-4) using existing different types of RTS noise detection and analysis methods to analyze the RTS noise pixel signal, and obtaining characteristic parameters of the RTS noise pixel signal, wherein the characteristic parameters include the maximum step amplitude, the average step interval time, the RTS level order, the signal drift value, and the pixel signal reconstructed after denoising and containing only RTS noise; 1-5) based on the characteristic parameters obtained in step 1-4), determining the characteristic parameters of the RTS noise pixel signal through manual screening and review, and combining the determination result of whether it is an RTS noise pixel signal as a data label; 1-6) adding the RTS noise pixel signal and its data label to the standard test database.

[0017] The above-mentioned image sensor random telegraph signal noise analysis method, wherein the initial data comes from the field data collected from pixel signals in real image sensors of different models, different displacement radiation damage, and different aging damage; at least 5000 groups of long-term pixel signals are collected as initial data, each group of pixel signals takes at least 1000 points, covering 30 minutes in the time domain, and more than half of them are manually judged as RTS noise pixel signals.

[0018] In the above-mentioned image sensor random telegraph signal noise analysis method, the pixel signal in the standard test database is also coupled with one or more noises of hot pixels, white noise, signal low-frequency drift, and burr signals.

[0019] The above-mentioned image sensor random telegraph signal noise analysis method, wherein the existing different types of RTS noise detection and analysis methods include a multi-level RTS automatic detection method based on time domain edge detection technology, a threshold-based method, RTN parameter extraction based on histogram Gaussian fitting, detection based on histogram Gram-Charlier series expansion, statistical characteristic analysis, and noise scatter plot.

[0020] In the above-mentioned image sensor random telegraph signal noise analysis method, in the step 3), when calculating the accuracy of the "RTS level order", the accuracy of each order is first calculated, and then the smallest one is taken as the accuracy of the "RTS level order".

[0021] In the above-mentioned image sensor random telegraph signal noise analysis method, in the step 3), the accuracy of the "pixel signal reconstructed after denoising and containing only RTS noise" is described by calculating the degree of overlap between the reconstructed pixel signal obtained by the RTS noise detection and analysis method and the reconstructed pixel signal in the standard test database.

[0022] The above-mentioned image sensor random telegraph signal noise analysis method, wherein in the step 3), if the number of non-RTS noise pixel signals with data labels in the standard test database is A, the RTS noise detection and analysis method analyzes and believes that among the A pixel signals, Aa pixel signals are non-RTS noise pixel signals, and Ab pixel signals are RTS noise pixel signals; the number of RTS noise pixel signals with data labels in the standard test database is B, the RTS noise detection and analysis method analyzes and believes that among the B pixel signals, Ba pixel signals are non-RTS noise pixel signals, and Bb pixel signals are RTS noise pixel signals;

[0023] When the RTS noise detection and analysis method gives a result of a non-RTS noise pixel signal, its accuracy AccJudge1 is:

[0024] AccJudge1=Aa / (Aa+Ba)

[0025] When the RTS noise detection and analysis method gives a result of an RTS noise pixel signal, its accuracy AccJudge2 is:

[0026] AccJudge2=Bb / (Ab+Bb)

[0027] Finally, the accuracy of the RTS noise detection and analysis method in analyzing whether it is an RTS noise pixel signal is:

[0028] AccJudge=min{AccJudge1, AccJudge2}.

[0029] In the above-mentioned image sensor random telegraph signal noise analysis method, in said step 6), the analysis results of different adopted RTS noise detection and analysis methods are subjected to voting to obtain the final result, and the weight of each RTS noise detection and analysis method is determined by the accuracy rate.

[0030] In the above-mentioned image sensor random telegraph signal noise analysis method, in said step 6), the analysis results of different adopted RTS noise detection and analysis methods are analyzed by using statistical methods such as mean, median or mode to determine the final result.

[0031] Compared with the prior art, the beneficial technical effects of the present invention are:

[0032] 1) Build a standard test database of RTS noise pixel signal coupled white noise, signal low-frequency drift, burr signal and hot pixel. The signal complexity can fully characterize various RTS noise pixel signals, fully cover real scenes, and effectively test the performance of RTS analysis solutions;

[0033] 2) The present invention provides a unified evaluation method to identify the capabilities of various RTS noise detection and analysis methods in different directions, thereby realizing a random telegraph noise analysis method coordinated with multiple RTS noise detection and analysis methods, reducing the amount of calculation while ensuring the accuracy of the calculation;

[0034] 3) Typical RTS noise detection and analysis methods (such as Figure 4 ) both start the analysis from the composition of the signal, eliminate the noise factors that affect the parameter analysis from the bottom up through mathematical means, and balance the accuracy of the analysis method and the amount of calculation through the manual selection of some key thresholds. After the calculation method is determined, the analysis path is fixed, and when facing unknown new noise factors, it will cause serious deviations in the parameter analysis; the random telegraph noise analysis method coordinated by multiple RTS noise detection and analysis methods can self-drive the iterative update of the detection and analysis method by continuously adding new RTS noise detection and analysis methods and updating the standard test database;

[0035] 4) The present invention can determine the data labels in the random telegraph noise analysis to provide a suitable evaluation for the analysis model trained by machine learning, thereby supporting its high-quality and effective training;

[0036] 5) The present invention solves the current difficulty in RTS noise analysis and can provide a more in-depth study of various process and radiation generation defects of image sensors. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] The image sensor random telegraph signal noise analysis method of the present invention is given by the following embodiments and drawings.

[0038] Figure 1 The flowchart of the method for analyzing random telegraph signal noise of an image sensor according to an embodiment of the present invention is shown.

[0039] Figure 2 Pixel signals of different types of typical RTS noise in the standard test database in an embodiment of the present invention.

[0040] Figure 3 These are pixel signals of non-RTS noises of different types coupled with other noises in the standard test database in an embodiment of the present invention.

[0041] Figure 4 This is a flow chart of a multi-level RTS automatic detection method based on time domain edge detection technology.

[0042] Figure 5 Schematic diagram of an image sensor pixel signal to be analyzed, a pixel signal reconstructed after denoising and containing only RTS noise, and a curve based on RTS noise determination in an embodiment of the present invention. DETAILED DESCRIPTION

[0043] The following will be combined Figure 1 to Figure 5 The image sensor random telegraph signal noise analysis method of the present invention is further described in detail.

[0044] Figure 1 Shown is a flow chart of a method for analyzing random telegraph signal noise of an image sensor according to an embodiment of the present invention.

[0045] See also Figure 1 The image sensor random telegraph signal noise analysis method of this embodiment includes:

[0046] 1) Build a standard test database;

[0047] 1-1) Collect and organize different types of long-term pixel signals as initial data through extensive experiments;

[0048] The initial data comes from the field data collected from pixel signals of real image sensors of different models, different displacement irradiation damage, and different seasoning damage;

[0049] In this embodiment, at least 5000 groups of long-time pixel signals are collected as initial data, each group of pixel signals has at least 1000 points, and covers 30 minutes in the time domain;

[0050] 1-2) Manually determine whether each pixel signal in the initial data is an RTS noise pixel signal, if so, execute step 1-4), if not, execute step 1-3);

[0051] According to whether it contains RTS noise, the initial data can be divided into two categories: one is the pixel signal containing RTS noise, which is defined as the RTS noise pixel signal (such as Figure 2 ), and the other is a pixel signal without RTS noise, which is defined as a non-RTS noise pixel signal;

[0052] Whether it is an RTS noise pixel signal or a non-RTS noise pixel signal, it may be coupled with hot pixels, white noise, signal low-frequency drift, burr signals, etc. Figure 3 Shown are non-RTS noise pixel signals of coupled hot pixels, white noise, signal low-frequency drift, and glitch signals;

[0053] In this embodiment, there are 2500 groups of pixel signals that are manually determined to be RTS noise;

[0054] 1-3), the determination result of whether it is an RTS noise pixel signal is used as a data label, and the non-RTS noise pixel signal and its data label are added to the standard test database;

[0055] 1-4) Analyze the RTS noise pixel signal using different existing RTS noise detection and analysis methods to obtain characteristic parameters of the RTS noise pixel signal, wherein the characteristic parameters include the maximum step amplitude, the average step interval time, the RTS level order, the signal drift value, and the pixel signal containing only RTS noise after denoising and reconstructing;

[0056] For RTS noise pixel signals coupled with hot pixels, white noise, signal low-frequency drift, burr signals, etc., denoising (i.e. removing hot pixels, white noise, signal low-frequency drift, burr signals, etc.) is performed to reconstruct a pixel signal containing only RTS noise as a characteristic parameter of the RTS noise pixel signal;

[0057] Each RTS noise pixel signal is analyzed by using different existing RTS noise detection and analysis methods. Thus, a certain characteristic parameter of a certain RTS noise pixel signal may have values ​​calculated by multiple RTS noise detection and analysis methods.

[0058] The existing different RTS noise detection and analysis methods, such as the multi-level RTS automatic detection method based on time domain edge detection technology (such as Figure 4 ), threshold-based methods, RTN parameter extraction based on histogram Gaussian fitting, detection based on histogram Gram-Charlier series expansion, statistical characteristics analysis, noise scatter plot, etc.;

[0059] 1-5) Based on the characteristic parameters obtained in step 1-4), the characteristic parameters of the RTS noise pixel signal are determined by manual screening and review, and the result of the determination of whether it is an RTS noise pixel signal is used as a data label;

[0060] Since each RTS noise pixel signal is analyzed by a plurality of RTS noise detection and analysis methods, a characteristic parameter of a certain RTS noise pixel signal may have multiple values, and the accuracy of these multiple values ​​is different; based on these multiple values, the RTS noise pixel signal (initial data) and experience, the characteristic parameter of the RTS noise pixel signal is finally determined through manual screening and review as a data label; the finally determined characteristic parameter may be a value calculated by a certain RTS noise detection and analysis method, or may not be a value calculated by the RTS noise detection and analysis method, but a value determined manually based on the RTS noise pixel signal (initial data) and experience;

[0061] 1-6) Add the RTS noise pixel signal and its data label to the standard test database;

[0062] 2) Using different existing RTS noise detection and analysis methods to analyze each pixel signal in the standard test database to obtain the RTS parameters of the pixel signal;

[0063] The different existing RTS noise detection and analysis methods, such as a multi-level RTS automatic detection method based on time domain edge detection technology, a threshold-based method, an RTN parameter extraction method based on histogram Gaussian fitting, a detection method based on histogram Gram-Charlier series expansion, statistical characteristic analysis, noise scatter plot, etc.;

[0064] Each pixel signal in the standard test database is analyzed using different existing RTS noise detection and analysis methods;

[0065] The RTS parameters include whether it is an RTS noise pixel signal, the maximum step amplitude, the average step interval time, the RTS level order, the signal drift value, and the pixel signal reconstructed after denoising and containing only RTS noise;

[0066] 3) Analyze and compare the RTS parameters of the pixel signal obtained by each RTS noise detection and analysis method with the data labels in the standard test database, and calculate the accuracy between the RTS parameters of the pixel signal obtained by each RTS noise detection and analysis method and the data labels in the standard test database;

[0067] Each RTS noise detection and analysis method analyzes all pixel signals in the standard test database, and analyzes and compares the analysis results (i.e., RTS parameters) of all pixel signals in the standard test database by the RTS noise detection and analysis method with the data labels in the standard test database, and calculates the accuracy of the RTS noise detection and analysis method for each RTS parameter. The accuracy reflects the analysis capability of the RTS noise detection and analysis method for the RTS parameter. Only when the accuracy meets certain conditions can the RTS noise detection and analysis method be used to obtain the RTS parameter.

[0068] (a) For the RTS parameter "Whether it is an RTS noise pixel signal", the accuracy calculation method is as follows:

[0069] If the number of pixel signals with data labels of non-RTS noise in the standard test database is A, the RTS noise detection and analysis method considers that among the A pixel signals, Aa pixel signals are non-RTS noise pixel signals, and Ab pixel signals are RTS noise pixel signals; if the number of pixel signals with data labels of RTS noise in the standard test database is B, the RTS noise detection and analysis method considers that among the B pixel signals, Ba pixel signals are non-RTS noise pixel signals, and Bb pixel signals are RTS noise pixel signals;

[0070] When the RTS noise detection and analysis method gives a result of a non-RTS noise pixel signal, its accuracy AccJudge1 is:

[0071] AccJudge1=Aa / (Aa+Ba)

[0072] When the RTS noise detection and analysis method gives a result of an RTS noise pixel signal, its accuracy AccJudge2 is:

[0073] AccJudge2=Bb / (Ab+Bb)

[0074] The smaller of AccJudge1 and AccJudge2 is taken as the accuracy of the RTS noise detection and analysis method to analyze whether it is an RTS noise pixel signal, to prevent the influence of the analysis method being biased towards all RTS or all non-RTS. Therefore, the final accuracy of the RTS noise detection and analysis method to analyze whether it is an RTS noise pixel signal is:

[0075] AccJudge=min{AccJudge1, AccJudge2};

[0076] (b) For “RTS level order”, the accuracy is calculated as follows:

[0077] The RTS level order is a discrete number. When calculating its accuracy, the accuracy of each order is calculated first, and then the smallest one is taken to prevent misjudgment caused by overfitting of a certain order by the test program;

[0078] Since the current engineering pays more attention to the pixel signals of 2nd, 3rd and 4th order RTS levels, the pixel signals of 5th order or above RTS levels have basically lost the credibility of the pixel signal results after repair. Therefore, for the RTS level order, the calculation accuracy method is as follows: if the number of RTS noise pixel signals with data labels of 2nd order in the standard test database is C, the RTS noise detection and analysis method believes that among these C pixel signals, Ca pixel signals are 2nd order RTS noise pixel signals, Cb pixel signals are 3rd order RTS noise pixel signals, Cc pixel signals are 4th order RTS noise pixel signals, and Cd pixels are RTS noise pixel signals of more orders;

[0079] The number of RTS noise pixel signals with data labels of 3rd order in the standard test database is D. The RTS noise detection and analysis method believes that among these D pixel signals, Da pixel signals are 2nd order RTS noise pixel signals, Db pixel signals are 3rd order RTS noise pixel signals, Dc pixel signals are 4th order RTS noise pixel signals, and Dd pixel signals are more order RTS noise pixel signals.

[0080] The number of RTS noise pixel signals with data labels of 4th order in the standard test database is E. The RTS noise detection and analysis method believes that among these E pixel signals, Ea pixel signals are 2nd order RTS noise pixel signals, Eb pixel signals are 3rd order RTS noise pixel signals, Ec pixel signals are 4th order RTS noise pixel signals, and Ed pixel signals are RTS noise pixel signals of more orders;

[0081] The number of pixel signals with data labels of more-order RTS noise in the standard test database is F. The RTS noise detection and analysis method believes that among these F pixel signals, Fa pixel signals are 2nd-order RTS noise pixel signals, Fb pixel signals are 3rd-order RTS noise pixel signals, Fc pixel signals are 4th-order RTS noise pixel signals, and Fd pixel signals are more-order RTS noise pixel signals.

[0082] When the RTS noise detection and analysis method gives a result of 2nd-order RTS level, its accuracy AccLv2 is:

[0083] AccLv2=Ca / (Ca+Da+Ea+Fa)

[0084] When the RTS noise detection analysis method gives the result of the third-order RTS level, its accuracy AccLv3 is:

[0085] AccLv3=Db / (Cb+Db+Eb+Fb)

[0086] When the RTS noise detection analysis method gives the result of the 4th order RTS level, its accuracy AccLv4 is:

[0087] AccLv4=Ec / (Cc+Dc+Ec+Fc)

[0088] The accuracy of different RTS level orders analyzed by the RTS noise detection and analysis method are AccLv2, AccLv3, and AccLv4, and the smallest one is taken as the accuracy of the "RTS level order" analyzed by the RTS noise detection and analysis method. Therefore, the accuracy of the "RTS level order" analyzed by the RTS noise detection and analysis method is:

[0089] AccLv=min{AccLv2, AccLv3, AccLv4}

[0090] (c) For the RTS parameter “pixel signal reconstructed after denoising and containing only RTS noise”, the accuracy is calculated as follows:

[0091] The pixel signal S2(t) obtained by the RTS noise detection and analysis method and reconstructed after denoising and containing only RTS noise is subtracted from the pixel signal S1(t) obtained by the standard test database and reconstructed after denoising and containing only RTS noise, and the integral value of the absolute value is taken. The maximum value S of the integral value and S1(t) is calculated. 1MAX The ratio of the integral over the time is described (another way is to use the correlation coefficient to describe it), where the greater the accuracy AccReSign is, the stronger the reconstruction ability of the RTS noise detection and analysis method is;

[0092] AccReSign=1-[∫|S1(t)-S2(t)|dt] / (∫S 1MAX dt);

[0093] 4) For each RTS parameter, determine whether the accuracy of each RTS noise detection and analysis method for the RTS parameter is ≥ the threshold value. If so, determine that the RTS noise detection and analysis is the analysis method for the RTS parameter. If not, the RTS noise detection and analysis is not used to analyze the RTS parameter, and determine the RTS noise detection and analysis method that can be adopted for each RTS parameter accordingly;

[0094] In this embodiment, for the RTS parameter "whether it is an RTS noise pixel signal", the credible RTS noise detection and analysis methods include the detection method based on the histogram Gram-Charlier series expansion, statistical characteristic analysis, and noise scatter plot; the credible RTS noise detection and analysis methods for "maximum step amplitude" include the noise scatter plot, the RTN parameter extraction method based on the histogram Gaussian fitting, and the multi-level RTS automatic detection method based on the time domain edge detection technology; the credible RTS noise detection and analysis methods for "RTS level order" include the RTN parameter extraction method based on the histogram Gaussian fitting, the multi-level RTS automatic detection method based on the time domain edge detection technology, the detection method based on the histogram Gram-Charlier series expansion, and the noise scatter plot;

[0095] 5) Inputting the image sensor pixel signal to be analyzed, using the credible RTS noise detection and analysis method determined in step 4) to determine whether the image sensor pixel signal to be analyzed is an RTS noise pixel signal, if so, executing step 6), if not, outputting the image sensor pixel signal as a non-RTS noise pixel signal, and ending the process;

[0096] 6) For each characteristic parameter, the image sensor pixel signal to be analyzed is analyzed using the credible RTS noise detection and analysis method for the characteristic parameter determined in step 4), and the statistical results of the analysis results of different credible RTS noise detection and analysis methods are taken as the final results, and each characteristic parameter is determined accordingly;

[0097] Preferably, the analysis results of different adopted RTS noise detection and analysis methods are voted to obtain the final result, and the weight of each RTS noise detection and analysis method is determined by the accuracy rate;

[0098] Preferably, for the analysis results of different adopted RTS noise detection and analysis methods, a statistical method such as mean, median or mode is used to determine the final result;

[0099] In this embodiment, the detection method based on the histogram Gram-Charlier series expansion, statistical characteristic analysis, and noise scatter Figure 3 Among the results given by the analysis method, two results are judged as RTS noise pixel signals, that is, the image sensor pixel signals to be analyzed are considered to be RTS noise pixel signals;

[0100] The median (i.e., the median) of the maximum step amplitudes obtained from the three analysis methods, namely, the noise scatter plot, the RTN parameter extraction method based on histogram Gaussian fitting, and the multi-level RTS automatic detection method based on time domain edge detection technology, is taken as the final result;

[0101] RTN parameter extraction method based on histogram Gaussian fitting, multi-level RTS automatic detection method based on time domain edge detection technology, detection method based on histogram Gram-Charlier series expansion, noise scatter point Figure 4 The RTS level order obtained by the analysis method takes the mode as the final result. If there is no mode, the mean is taken as the final result. For example, three of the four results are 2nd order and one is 3rd order, then the final result is 2nd order.

[0102] Figure 5 Schematic diagram of an image sensor pixel signal to be analyzed, a pixel signal reconstructed after denoising and containing only RTS noise, and a curve based on RTS noise determination in an embodiment of the present invention. Figure 5 The blue part in the upper middle part is the input image sensor pixel signal to be analyzed, and the red part is the pixel signal reconstructed after denoising and containing only RTS noise; Figure 5 The lower middle part is the RTS noise judgment basis curve for characteristic parameters obtained by using a multi-level RTS automatic detection method based on time domain edge detection technology. The possible location of the step on the pixel signal is found through windowed convolution filtering to obtain the black curve. It is then judged in turn whether a step occurs at that location. If it is a step point, it is marked in red. After determining the step position and the amplitude of each step, the reconstructed signal can be obtained. In the figure, the reconstructed signal has basically completely removed all noise except the RTS noise, and only contains pixel signals and RTS noise.

[0103] Pixel signals containing RTS noise are usually coupled with white noise, signal low-frequency drift and hot pixels. The signals are complex and require a large amount of data to fully characterize various types of RTS noise pixel signals. In existing technical literature, only ideal RTS noise models or a very small amount (about 50 groups) of real test data are usually analyzed and explained. Commonly used samples are difficult to fully cover real scenes, which poses a challenge to the effectiveness of performance testing of RTS noise detection and analysis methods. The present invention collects and organizes initial data through extensive experiments, collects at least 5,000 groups of pixel signals, of which 2,500 groups are determined to be RTS noise pixel signals, and each group of pixel signals takes at least 1,000 points, covering 30 minutes in the time domain.

[0104] The present invention takes the statistical results of the analysis results of different adopted RTS noise detection and analysis methods as the final result, is a random telegraph signal noise analysis method coordinated with multiple RTS noise detection and analysis methods, and integrates the advantages of multiple RTS noise detection and analysis methods.

Claims

1. A method for analyzing random telegraph signal noise of an image sensor, characterized in that: include: 1) Construct a standard test database based on different types of long-term pixel signals collected and collated from extensive experiments; 2) Using different existing RTS noise detection and analysis methods to analyze each pixel signal in the standard test database to obtain the RTS parameters of the pixel signal; The RTS parameters include whether it is an RTS noise pixel signal, the maximum step amplitude, the average step interval time, the RTS level order, the signal drift value, and the pixel signal reconstructed after denoising and containing only RTS noise; 3) Analyze and compare the RTS parameters of the pixel signal obtained by each RTS noise detection and analysis method with the data labels in the standard test database, and calculate the accuracy between the RTS parameters of the pixel signal obtained by each RTS noise detection and analysis method and the data labels in the standard test database; 4) For each RTS parameter, determine whether the accuracy of each RTS noise detection and analysis method for the RTS parameter is ≥ the threshold value. If so, determine that the RTS noise detection and analysis is the analysis method for the RTS parameter. If not, the RTS noise detection and analysis is not used to analyze the RTS parameter, and accordingly determine the RTS noise detection and analysis method that can be adopted for each RTS parameter; 5) Inputting the image sensor pixel signal to be analyzed, using the credible RTS noise detection and analysis method determined in step 4) to determine whether the image sensor pixel signal to be analyzed is an RTS noise pixel signal, if so, executing step 6), if not, outputting the image sensor pixel signal as a non-RTS noise pixel signal, and ending the process; 6) For each characteristic parameter, the image sensor pixel signal to be analyzed is analyzed using the RTS noise detection and analysis method that can be used for the characteristic parameter determined in step 4), and the statistical results of the analysis results of different credible RTS noise detection and analysis methods are taken as the final result, and each characteristic parameter is determined accordingly.

2. The image sensor random telegraph signal noise analysis method according to claim 1, characterized in that: The step 1) comprises: 1-1) Collect and organize different types of long-term pixel signals as initial data through extensive experiments; 1-2) Manually determine whether each pixel signal in the initial data is an RTS noise pixel signal, if so, execute step 1-4), if not, execute step 1-3); 1-3), the determination result of whether it is an RTS noise pixel signal is used as a data label, and the non-RTS noise pixel signal and its data label are added to the standard test database; 1-4) Analyze the RTS noise pixel signal using different existing RTS noise detection and analysis methods to obtain characteristic parameters of the RTS noise pixel signal, wherein the characteristic parameters include the maximum step amplitude, the average step interval time, the RTS level order, the signal drift value, and the pixel signal containing only RTS noise after denoising and reconstruction; 1-5) Based on the characteristic parameters obtained in step 1-4), the characteristic parameters of the RTS noise pixel signal are determined by manual screening and review, and the result of the determination of whether it is an RTS noise pixel signal is used as a data label; 1-6) Add the RTS noise pixel signal and its data label to the standard test database.

3. The image sensor random telegraph signal noise analysis method as claimed in claim 2, characterized in that: The initial data comes from field data collected from pixel signals in real image sensors of different models, different displacement radiation damage, and different aging damage; at least 5,000 groups of long-term pixel signals are collected as initial data, each group of pixel signals has at least 1,000 points, covering 30 minutes in the time domain, and more than half of them are manually judged as RTS noise pixel signals.

4. The image sensor random telegraph signal noise analysis method as claimed in claim 2, characterized in that: The pixel signals in the standard test database are also coupled with one or more of the following noises: hot pixels, white noise, signal low-frequency drift, and burr signals.

5. The image sensor random telegraph signal noise analysis method according to claim 2, characterized in that: The different existing RTS noise detection and analysis methods include a multi-level RTS automatic detection method based on time domain edge detection technology, a threshold-based method, RTN parameter extraction based on histogram Gaussian fitting, detection based on histogram Gram-Charlier series expansion, statistical characteristic analysis, and noise scatter plot.

6. The image sensor random telegraph signal noise analysis method according to claim 2, characterized in that: In the step 3), when calculating the accuracy of the "RTS level order", the accuracy of each order is calculated first, and then the smallest one is taken as the accuracy of the "RTS level order".

7. The image sensor random telegraph signal noise analysis method according to claim 2, characterized in that: In the step 3), the accuracy of the "pixel signal reconstructed after denoising and containing only RTS noise" is described by calculating the degree of overlap between the reconstructed pixel signal obtained by the RTS noise detection and analysis method and the reconstructed pixel signal in the standard test database.

8. The image sensor random telegraph signal noise analysis method according to claim 2, characterized in that: In the step 3), if the number of non-RTS noise pixel signals in the standard test database with data labels is A, the RTS noise detection and analysis method analyzes and believes that among the A pixel signals, Aa pixel signals are non-RTS noise pixel signals and Ab pixel signals are RTS noise pixel signals; The number of pixel signals with data labels of RTS noise in the standard test database is B. The RTS noise detection and analysis method believes that among these B pixel signals, Ba pixel signals are non-RTS noise pixel signals and Bb pixel signals are RTS noise pixel signals. When the RTS noise detection and analysis method gives a result of a non-RTS noise pixel signal, its accuracy AccJudge1 is: AccJudge1=Aa / (Aa+Ba) When the RTS noise detection and analysis method gives a result of an RTS noise pixel signal, its accuracy AccJudge2 is: AccJudge2=Bb / (Ab+Bb) Finally, the accuracy of the RTS noise detection and analysis method in analyzing whether it is an RTS noise pixel signal is: AccJudge = min{AccJudge1, AccJudge2}.

9. The image sensor random telegraph signal noise analysis method according to claim 2, characterized in that: In the step 6), the analysis results of different adopted RTS noise detection and analysis methods are voted to obtain the final result, and the weight of each RTS noise detection and analysis method is determined by the accuracy.

10. The image sensor random telegraph signal noise analysis method according to claim 2, characterized in that: In the step 6), the final result is determined by using statistical methods such as mean, median or mode for the analysis results of different RTS noise detection and analysis methods.

Images