Adaptive photon denoising three-dimensional imaging method under sparse or strong noise counting

Abstract

The invention belongs to the technical field of single-photon imaging laser radars, and discloses a self-adaptive photon denoising three-dimensional imaging method under sparse or strong noise counting, which is used for extracting echo area counting in a self-adaptive manner under different noise conditions and accurately positioning the echo counting according to the specificity degree of a counting value in a neighborhood, so that the accuracy of the echo area counting is improved. Laser flight time is obtained to obtain distance information, and all pixels are traversed and processed to realize accurate reconstruction of a target three-dimensional image. According to the method, the low-signal-background-ratio detection conditions of sparse echo counting and strong noise counting can be adaptively processed, and the method has better adaptability to various different detection environments and detection applications; echo positioning is carried out according to the time domain distribution characteristic difference of echo counting and noise counting in the counting histogram, the method can automatically adapt to the low signal background ratio condition of sparse counting and strong noise counting, and the method has good robustness for noise.

Description

Technical Field

[0001] This invention belongs to the field of single-photon imaging lidar technology, and relates to an adaptive photon denoising three-dimensional imaging method under sparse or noisy counting. Background Technology

[0002] LiDAR, as a non-contact active detection method, can obtain high spatial resolution distance and reflectivity information of targets by actively emitting lasers. Therefore, LiDAR has been widely used in various detection scenarios such as natural environment monitoring, geographic mapping, and autonomous driving. The basic principle of LiDAR is to actively emit laser pulses towards the target and detect the laser echo that returns along the original path after diffuse reflection from the target. After photoelectric conversion, the flight time of the laser to the target and the intensity of the echo are determined, thereby obtaining the target's distance and reflectivity information.

[0003] Traditional 3D imaging lidar uses detectors operating in linear mode to obtain the complete waveform of the echo. However, linear mode requires high power for the transmitted pulse, and the high energy consumption limits the working distance of the lidar.

[0004] In recent years, single-photon lidar based on echo photon counting has attracted widespread attention. Single-photon lidar can respond to photon-level energy and has extremely high sensitivity, thus enabling detection with lower power consumption and higher laser pulse repetition rate, increasing working distance and spatial resolution. However, due to its extremely high sensitivity, it is also easily affected by dark counting of the detection device and external background light, resulting in a large portion of the detection response being noise. In actual detection, hundreds or thousands of echo response counts are often required to obtain an accurate echo photon count histogram, leading to increased imaging time costs.

[0005] In practical applications of mobile platforms and rapid scanning imaging, power consumption and detection time are limited, and it is difficult to improve the signal-to-noise ratio by increasing the number of echo accumulations. Therefore, accurate and rapid echo positioning in the low signal-to-background ratio (SBR) histogram obtained by single-photon lidar detection to achieve precise distance imaging has become one of the key technologies for single-photon lidar detection applications.

[0006] Traditional echo localization algorithms include cross-correlation based on single pixel histograms and maximum likelihood methods. These algorithms are fast, but they have high requirements for signal-to-background ratio and are severely affected by noise. When the echo count is insufficient, the imaging effect is very poor. In existing photon-efficiency type echo localization algorithms, spatial correlation is used to enhance the echo count. The count level is modeled as the sum of the system impulse response and the constant level of noise. Then, the counting process is regarded as a Poisson process. The likelihood function is constructed by combining the count data obtained from observation and then the maximum likelihood is solved. The unconditioned problem of echo localization is transformed into a constrained optimization problem [1], [2], [3], [4]. However, due to the discrete digital response characteristics of the detector, it only responds to the presence or absence of photons during detection and has no ability to distinguish the number of photons. Therefore, the count value does not represent the actual number of photons reached. Moreover, the detector device has a dead time, which leads to the loss of photon counts. Under high illumination conditions, the time distribution of the count will be distorted, and the photon counts at different times are no longer independent of each other. Therefore, they no longer follow the Poisson distribution. Therefore, these algorithms ignore the dead time and model the counting process as a Poisson process, which is not accurate. When the detector dead time is long, there will be a large deviation from the real situation. Therefore, the Poisson observation model of counting may introduce errors into the imaging and is greatly affected by noise, and its universality is insufficient in different application scenarios. In addition, there is also an echo localization method that traverses the counting histogram through a sliding window and uses the position containing the most counts in the window as the estimate of the echo position [5]. However, this method has high requirements for the signal-to-background ratio and cannot perform echo localization well when the echo count is sparse or when the echo count is submerged by noise counts. Existing algorithms also have regulations on the response counting histogram to be processed. Usually, sparse response counting or strong backscattering is required in advance, and then different processing strategies are adopted under these premises. If the actual input counting histogram is inconsistent with the assumed situation, it will increase the computational cost and time cost of the algorithm, and even cause the algorithm to fail, thus further limiting its adaptability.

[0007] In order to automatically adapt to different noise environments and adaptively handle histogram cases of sparse counting and strong noise counting, thereby enhancing the versatility and robustness of the echo localization algorithm to noise, this invention proposes an adaptive photon denoising three-dimensional imaging method under sparse or strong noise counting.

[0008] Reference document description:

[0009] [1]Joshua Rapp, Vivek K Goyal, A Few Photons Among Many: Unmixing Signaland Noise for Photon-Efficient Active Imaging[J]. IEEE Transactions on Computational Imaging, 2017, 3(3): 445-459.

[0010] [2] Dongeek Shin, Ahmed Kirmani, Vivek K Goyal, et al. Photon-EfficientComputational 3-D and Reflectivity Imaging With Single-Photon Detectors[J].. IEEE Transactions On Computational Imaging, 2015, 1(2): 112-125.

[0011] [3]Juli\'an Tachella,Yoann Altmann,Ximing Ren,et al.Bayesian3DReconstruction ofComplex Scenes from Single-Photon Lidar Data[J].SIAMJ.Imaging Sciences,2019,12(2):521-550.

[0012] [4] Patent CN 114488186A, published on May 13, 2022, full text.

[0013] [5] Patent CN 104914446A, published on September 16, 2015, full text. Summary of the Invention

[0014] (I) Purpose of the Invention

[0015] The purpose of this invention is to provide an adaptive photon denoising 3D imaging method for sparse or noisy counting, based on the aforementioned background technology, in order to adaptively perform echo localization under different noise conditions and adapt to different counting histogram conditions, thereby enhancing the adaptability and stability to different target scenes. This method adaptively extracts echo region counts under different noise conditions, accurately locates the echo counts based on the specificity of the count values ​​in the neighborhood, obtains the laser flight time to obtain distance information, and traverses and processes all pixels to achieve accurate reconstruction of the target 3D image.

[0016] (II) Technical Solution

[0017] To address the aforementioned technical problems, this invention provides an adaptive photon denoising three-dimensional imaging method for sparse or noisy counting, characterized by comprising the following steps:

[0018] Step 1: Use a photonic lidar to emit laser pulses, detect and acquire the counting histogram at each pixel, process the acquired counting histogram, and use the counting histograms of pixels in the spatial neighborhood of the current pixel to perform echo counting enhancement on the counting histogram of the current pixel.

[0019] Step 2: For the enhanced counting histogram, use a sliding window to traverse and statistically obtain the zero-value threshold ThrZero for each counting position, and regard the count values ​​less than or equal to the corresponding zero-value threshold as zero values; then use a sliding window to traverse the counting histogram processed according to the zero-value threshold, count the number of zero count values ​​in the window and the average count in the window for each counting position, and substitute them into the definition of the exponential function to obtain the clustering parameter Eau of the count values, and set the clustering threshold ThrEau;

[0020] Step 3: Divide the enhanced count histogram length T into multiple equal-length non-overlapping sub-time intervals, and select intervals that may be echoes based on the clustering parameter Eau and the threshold ThrEau. For the count values ​​at positions where the clustering is greater than the threshold ThrEau, enhance them non-uniformly based on the clustering.

[0021] Step four: Use a Gaussian function to define the relative difference between count values ​​in the time neighborhood, with the difference value between 0 and 1; within each retained echo interval, calculate the total difference of count values ​​in the time neighborhood to characterize the specificity of the count in the neighborhood.

[0022] Step 5: Within the remaining possible echo intervals, the echo position is optimized by using the specificity of the count value in the time neighborhood as the objective function, thereby obtaining the laser pulse flight time and calculating the distance.

[0023] Step 6: Process each pixel using steps 1 to 5, and construct a 3D image based on the obtained distance values.

[0024] Furthermore, in step one, the echo counting enhancement formula is:

[0025]

[0026] In the formula, V (i,j) w represents a square region centered on the currently processed pixel (i, j). (k，z) ∈[0,1] represents the pixel weight, Y (k,z)Y' is a count histogram of the neighboring pixels (k, z) of pixel (i, j). (i,j) This is the count histogram for the enhanced pixel (i, j).

[0027] Furthermore, in step two, the process of obtaining the zero-value threshold ThrZero is as follows: traverse the counting histogram through a sliding window, and take the largest count value within the window as the zero-value threshold ThrZero at the location of the sliding window; if the count value is less than or equal to the zero-value threshold ThrZero, it is regarded as zero in subsequent processing, otherwise it is retained.

[0028] Furthermore, in step two, the calculation process for the clustering parameter Eau is as follows:

[0029]

[0030] Eau (i,j),t =Ea (i,j),t / max({Ea (i,j),k |k=1,2,...,T})

[0031] In the formula, Eau (i,j),t Wn represents the normalized clustering parameter at the t-th counting position of pixel (i, j); t This represents the time range contained at the t-th count position of the sliding window center in the histogram; Nzero (i,j),t Let b be the number of zero count values ​​within the window when the center of the sliding window is at the t-th counting position in the histogram, where b>0 is a constant.

[0032] Furthermore, in step two, the value of Eau is in the range of (0,1], and the aggregation threshold ThrEau is set to 0.5, so as to remove most of the noise counting area while avoiding the removal of the real echo counting area.

[0033] Furthermore, in step three, let T be divided into s subintervals, then we have:

[0034]

[0035] Among them, T l This represents the l-th sub-interval obtained after partitioning. Since the laser pulse time width Tp is only on the order of nanoseconds, the external noise level can be considered constant at this time order. Therefore, the length L of the sub-time interval is... sub Setting it to 10 times Tp will highlight the echo counting specificity.

[0036] Further, in step three, possible echo intervals are selected and retained based on clustering: sub-time intervals with clustering greater than a set threshold ThrEau are retained and considered as possible echo intervals; otherwise, they are considered noise intervals. The count values ​​within the retained possible echo intervals are non-uniformly enhanced based on clustering to highlight the specificity of the echo count within the neighborhood. Yen' represents the enhanced count data, and the enhancement process is as follows:

[0037] (1) If Eau (i,j),t ≥ThrEau (i,j),t To enhance, Yen' (i,j),t =Y' (i,j),t ·(1+Eau (i,j),t );

[0038] (2) If Eau (i,j),t <ThrEau (i,j),t Without enhancement, Yen' (i,j),t =Y' (i,j),t ;

[0039] Furthermore, in step four, due to the random nature of the count value histogram, directly comparing the differences between two count values ​​is insufficient to obtain effective information. Therefore, this invention compares the count value at the current time position within the time neighborhood. Mean of small neighborhood counts at other time locations The relative differences between them.

[0040] Furthermore, the relative difference Dp defined in step four for the Gaussian function is as follows:

[0041]

[0042] t c ,t o =1,2,...,L sub

[0043] In the formula, This indicates that within the l-th sub-time interval of pixel (i,j), the t-th time interval... c Count value With the tth o Average number of counts The relative degree of difference; the greater the difference, the better. The larger the value, the better. st >0 is a parameter for adjusting the sensitivity to differences.

[0044] Furthermore, within the l-th sub-time interval, The total relative difference TDp relative to all other locations in the neighborhood is:

[0045]

[0046] Because the sub-time intervals are relatively short, the noise levels within them also vary little and can be considered constant. Therefore, when calculating the relative difference value, there will be relative constraints between the noise counts within the sub-time intervals, and their overall relative difference is small. However, the count in the echo region can be considered as the sum of the echo count and the noise count. The echo count causes abrupt changes in the echo region count value, and the overall difference between the echo count and the noise count is larger, indicating strong specificity within the time neighborhood.

[0047] Furthermore, in step five, all possible echo intervals preserved based on clustering are traversed and processed. The relative difference degree TDp within the neighborhood is used as the objective function for constrained optimization to obtain the echo location. The constrained optimization problem for the echo location is as follows:

[0048]

[0049] l = 1, 2, ..., s

[0050] t c =1,2,...,L sub

[0051] In the formula, This indicates the sub-time interval number where the echo obtained from the optimization solution is located. Indicates the echo at the 1st Location within a specific time interval, TV (i,j) This represents the smoothing penalty term introduced during optimization.

[0052] Therefore, the optimal distance estimate for pixel (i,j) is as follows:

[0053]

[0054] In the formula, This represents the estimated absolute position of the echo at pixel (i,j) in the counting histogram. This represents the distance estimate at pixel (i,j).

[0055] Furthermore, in step six, the target distance image is obtained by traversing and processing each pixel.

[0056] (III) Beneficial Effects

[0057] The adaptive photon denoising three-dimensional imaging method for sparse or noisy counting provided by the above technical solution has the following beneficial effects:

[0058] (1) The present invention can adaptively handle the low signal-to-background ratio detection situation of sparse echo counting and strong noise counting, and has better adaptability to various detection environments and applications.

[0059] (2) The present invention first enhances echo counting by accumulating the pixel count histogram in the spatial neighborhood based on the spatial correlation between pixels; then, it simultaneously considers the clustering characteristics of the count level and the count value on the time axis to separate the echo and noise count regions, retaining only the possible echo regions, thereby limiting the error and reducing the amount of data to be processed.

[0060] (3) The present invention enhances the echo region count non-uniformly based on the aggregation, further weakening the influence of noise count; then the relative difference between count values ​​in the time neighborhood is defined by the Gaussian function, and the specificity of the count is characterized by the total difference of the count in its time neighborhood. In the preserved possible echo region, the count position with the strongest specificity in the neighborhood is used as the estimate of the echo position, thereby obtaining the estimated value of the distance value.

[0061] (4) When performing distance estimation, this invention does not impose restrictions or make prior assumptions on the noise count level. It performs echo localization based on the difference in the temporal distribution characteristics of echo count and noise count in the count histogram. It can automatically adapt to the low signal-to-background ratio of sparse count and strong noise count, and has good robustness to noise. Attached Figure Description

[0062] Figure 1 This is a flowchart of an adaptive photon denoising three-dimensional imaging method under sparse or noisy counting, according to an embodiment of the present invention.

[0063] Figure 2 The results of echo localization using the method of this invention are illustrated in the counting histograms under different signal-to-background ratios (SBR).

[0064] Figure 3 Comparison of the imaging performance of the method of the present invention with several other algorithms under strong noise counting conditions;

[0065] Figure 4 In the case of sparse counting, the imaging effect of the method of the present invention is compared with several other algorithms. Detailed Implementation

[0066] To make the objectives, contents, and advantages of the present invention clearer, the specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples.

[0067] like Figure 1 As shown, the adaptive photon denoising three-dimensional imaging method under sparse or noisy counting in this embodiment includes the following steps:

[0068] Step 1: Use a photonic lidar to emit laser pulses and detect and acquire the count histogram at each pixel. Process the acquired count histograms. When processing the count histogram of each pixel, use the count histograms of pixels in the spatial neighborhood of the current pixel to accumulate the count histogram of the current pixel with weights. This will supplement and enhance the echo count of the currently processed pixel, so as to initially improve the SBR.

[0069] The enhancement method for boosting the count is as follows:

[0070]

[0071] In the formula, V (i,j) w represents a square region centered on the currently processed pixel (i, j). (k，z) ∈[0,1] represents the pixel weight, Y (k,z) Y' is a count histogram of the neighboring pixels (k, z) of pixel (i, j). (i,j) This is the count histogram for the enhanced pixel (i, j).

[0072] Pixel weights can be set using a Gaussian 3D pulse shape, making the histogram of the central pixel dominant, while the other pixels in the spatial neighborhood serve as auxiliary pixels, thus reducing the impact of different pixels having different echo positions.

[0073] The spatial neighborhood region V used when enhancing pixel (i, j) (i,j) The size can be determined by the overall image size (Nr×Nc) and the SBR (Signal-to-Noise Ratio). Generally, the side length of the square neighborhood can be set to (1 / 50~1 / 100)×min(Nr, Nc). If the pixel neighborhood is too small, it may not be able to effectively enhance the counting data in the case of sparse photons with a low SBR, thus affecting the localization effect; if the pixel neighborhood is too large, too many pixels at different distances may be selected for enhancement, resulting in blurred target edges in the imaging result. Due to spatial correlation characteristics, the distances of the neighborhood regions near a pixel are close, that is, the echo positions in the counting histogram are close, and the echo counts are concentrated in a very small time region, while the noise counts are evenly distributed across the entire time axis. Therefore, accumulating the counting histograms of pixels within the neighborhood is equivalent to improving the signal-to-noise ratio. Enhancing the neighborhood pixels can highlight the features of the echo region and can also "fill in" "empty pixels" where there are no echo counts.

[0074] Step two involves using a sliding window to iterate through the enhanced counting histogram and obtain the zero-value threshold ThrZero for each counting position. Count values ​​less than or equal to the corresponding zero-value threshold are considered zero. Then, the sliding window iterates through the counting histogram processed according to the zero-value threshold, and the clustering parameter Eau is obtained by substituting the definition of an exponential function. The clustering parameter at each counting position is calculated, and a clustering threshold ThrEau is set. Thus, each counting position has a corresponding clustering parameter, which can characterize the temporal distribution characteristics of the counts at that position.

[0075] Due to varying noise levels and detection conditions, the count values ​​in the obtained count histogram may be very sparse, consisting only of 0 and 1 (sparse counting), or they may be filled with a large number of non-zero noise counts (strong noise counting). When calculating the clustering parameter for histograms of count values ​​under different conditions, simply using whether the count value is zero is meaningless and lacks versatility. To automatically adapt to different count histogram conditions, a sliding window is first used to obtain the zero-value threshold ThrZero at each position. When calculating the clustering parameter, the zero-value threshold is used to determine whether to treat the count value as zero. When the count is sparse, zero is used as the threshold; when the count is "dense," a larger number is used as the threshold. This way, regardless of the number or size of the count values, the clustering characteristics of the count values ​​can be adaptively reflected.

[0076] In step two, a sliding window is used to traverse the counting histogram to obtain the zero threshold at each position, and then the clustering parameter of the count value is obtained. The sliding window width Tw used in these two steps is determined based on the pulse width Tp. The echo count is mainly clustered in a small interval of width Tp in time. Therefore, in order to highlight the clustering characteristics at the echo, Tw can be taken as Tp ~ 2×Tp. The sliding window width is slightly larger than the pulse, which is conducive to finding the echo region.

[0077] The specific method for obtaining the zero-value threshold ThrZero through counting statistics is as follows: select the count value with the highest proportion within the sliding window as the zero-value threshold ThrZero for the corresponding counting position. When the count value is less than or equal to the zero-value threshold ThrZero at the corresponding position, it is considered a zero count value in subsequent steps; otherwise, it is retained.

[0078] The clustering parameter Eau combines two features: the magnitude of the count level and the degree of clustering of the counts over time. The definition is as follows:

[0079]

[0080] Eau (i,j),t =Ea (i,j),t / max({Ea (i,j),k |k=1,2,...,T})

[0081] In the formula, Eau (i,j),t Wn represents the normalized clustering parameter at position t in the histogram of pixel (i, j). t Nzero represents the range contained at the t-th position of the sliding window center in the histogram. (i,j),t This represents the number of counts considered as zero within the window. b > 0 is a constant. Changing the value of parameter b can change the sensitivity to the number of counts considered as zero within the window. The larger b is, the more stringent the measure of clustering becomes, which means that more non-zero counts are required within the sliding window to increase the clustering parameter. b = 0.5 can be set.

[0082] The clustering threshold ThrEau determines the range of clustering parameters to be selected, considering regions with clustering parameters greater than the threshold as potential echo regions. Decreasing the clustering parameter threshold ThrEau increases the probability of selecting and retaining true echo intervals, but also increases the probability of mistaking noise counts for echo intervals. Increasing the clustering parameter threshold ThrEau can better filter noise counts, but in complex detection scenarios, there may be situations where noise counts are highly clustered while echo intervals are less clustered; increasing the threshold ThrEau may cause true echo intervals to be overlooked.

[0083] In this embodiment, the value of Eau is in the range of (0,1], and the aggregation threshold ThrEau can be set to 0.5, so as to remove most of the noise counting area while avoiding the removal of the real echo counting area.

[0084] Step 3: Divide the length T of the enhanced counting histogram into multiple equal-length, non-overlapping sub-time intervals. Assuming T is divided into s sub-intervals, then we have... Where T l This represents the l-th sub-interval obtained after partitioning. Because the laser pulse duration Tp is only on the order of nanoseconds, the change in external noise is relatively slow; therefore, the external noise level can be considered constant at this time interval. Thus, the length L of the sub-interval is... sub Setting the value to 10 times Tp highlights the echo counting specificity. Based on the clustering parameter Eau and the threshold ThrEau, intervals that are likely echoes are selected for retention: sub-time intervals with clustering greater than the threshold are retained as possible echo intervals; otherwise, they are considered noise intervals. Furthermore, the count values ​​within the retained possible echo intervals are non-uniformly amplified based on clustering to highlight the echo counting specificity within the neighborhood.

[0085] The count value is non-uniformly enhanced based on its clustering property, increasing the value according to the magnitude of the clustering parameter. Strong clustering may result in a large increase in the position of the echo count, while low clustering means the noise count area remains largely unchanged. This prevents the enhancement of noise counts from affecting echo localization.

[0086] (1) If Eau (i,j),t ≥ThrEau (i,j),t Enhance Yen' (i,j),t =Y' (i,j),t ·(1+Eau (i,j),t );

[0087] (2) If Eau (i,j),t <ThrEau (i,j),t No enhancement Yen' (i,j),t =Y' (i,j),t

[0088] Step four: To facilitate the normalization analysis of the relative differences between count values, a Gaussian function is used to define the relative differences between count values ​​within the time neighborhood, ensuring that the difference value is between 0 and 1. Within each retained echo interval, the total difference of the count values ​​within the neighborhood is calculated to characterize the specificity of the count within the neighborhood.

[0089] Due to the random nature of the count histogram, directly comparing the differences between two count values ​​is insufficient to obtain effective information. Therefore, this invention compares the count value at the current time position. Mean of small neighborhood counts at other time locations The relative differences between them.

[0090] The relative difference Dp between counts in the time neighborhood, as defined by the Gaussian function, is as follows:

[0091]

[0092] t c ,t o =1,2,...,L sub

[0093] In the formula, This indicates that within the l-th sub-time interval of pixel (i,j), the t-th time interval... c Count value With the tth o Average number of counts The relative degree of difference; the greater the difference, the better. The larger the value, the greater the σ. st >0 is a parameter that adjusts the sensitivity to differences; σ st The smaller the value, the more sensitive it is to minute differences; σ can be set. st=1.

[0094] Because the sub-time intervals are relatively short and the noise level changes within them are also relatively small, there are relative constraints between the noise counts within the sub-time intervals when calculating the relative difference value, and their overall relative difference is small. However, the echo region count is the sum of the echo count and the noise count. The echo count causes abrupt changes in the echo region count value, resulting in a larger overall difference between the echo count and the noise count, indicating strong neighborhood specificity.

[0095] Step 5: Within all possible echo intervals, optimize the echo position by using the specificity of the count value within the neighborhood as the objective function, thereby obtaining the laser pulse flight time to calculate the distance, and then process the next pixel.

[0096] The objective function is optimized within the possible echo region selected based on the clustering parameter to obtain the sub-time interval location l where the echo count is located, and the location t of the echo peak within that interval. c .

[0097] The constraint optimization problem regarding the echo location is as follows:

[0098]

[0099] l = 1, 2, ..., s

[0100] t c =1,2,...,L sub

[0101] In the formula, l (i,j) echo The t represents the sub-time interval number where the echo obtained from the optimization solution is located. (i,j) echo Indicates the echo at the lth (i,j) echo Location within a specific time interval, TV (i,j) This represents the smoothing penalty term introduced during optimization.

[0102] Therefore, the optimal distance estimate for pixel (i, j) is as follows:

[0103]

[0104] In the formula, This represents the estimated absolute position of the echo at pixel (i, j) in the counting histogram. This represents the distance estimate at pixel (i, j).

[0105] Furthermore, each pixel is processed to obtain the target distance image.

[0106] To verify the effectiveness of this invention in echo localization within the histogram under low SBR conditions, low SBR counting histograms were generated for both sparse and high-noise counting scenarios. The counting histograms contained 4800 counting positions, with the laser pulse width occupying 10 positions and the pulse echo peak value counting position set to 1560. For the high-noise counting scenario, SBR = 20 / 4014 = 0.0049 and SBR = 30 / 4014 = 0.0074 were set, and counting histograms were generated. For the sparse counting scenario, SBR = 8 / 120 = 0.0664 and SBR = 4 / 60 = 0.0664 were set, and counting histograms were generated. Figure 2 As shown in the figure, (a) and (b) represent the counting histograms and echo localization results for the case of strong noise counting; (c) and (d) represent the counting histograms and echo localization results for the case of sparse counting. The results show that the method of the present invention achieves accurate echo localization in both sparse counting and strong noise counting cases with extremely low SBR conditions.

[0107] Furthermore, to test the photon counting 3D imaging effect, pixel-by-pixel counting data was generated for distance image reconstruction testing, and the results were compared with existing algorithms. For noisy counting, SBR was set to 2 / 200 = 0.01; for sparse counting, SBR was set to 0.4 / 1 = 0.4. The imaging results are as follows... Figures 3-4 Show, Figure 3 This is the imaging result under strong noise counting conditions. Figure 4 The image shows the imaging results for the sparse counting case. The GT depth in the figure represents the reference depth image. The results show that the method of this invention achieves better imaging for both sparse and noisy counting cases, and provides more accurate and complete reconstruction of details and edges in the scene.

[0108] As can be seen from the above technical solution, this invention proposes an adaptive photon denoising 3D imaging method under sparse or noisy photon counting based on the temporal distribution characteristics of photon counting. It utilizes spatial correlation to enhance echo counting and considers the temporal aggregation characteristics of echo counting to separate echo and noise regions in the counting histogram. Experimental results show that under extremely low SBR conditions in both sparse and noisy counting scenarios, the method accurately achieves 3D imaging of the scene and completely and accurately preserves edge details, demonstrating strong versatility for different photon counting conditions, good robustness to noise, and suitability for photon counting imaging under low SBR conditions.

[0109] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. An adaptive photon denoising three-dimensional imaging method under sparse or noisy counting, characterized in that, Includes the following steps: Step 1: Use a photonic lidar to emit laser pulses, detect and acquire the counting histogram at each pixel, process the acquired counting histogram, and use the counting histograms of pixels in the spatial neighborhood of the current pixel to perform echo counting enhancement on the counting histogram of the current pixel. Step 2: For the enhanced counting histogram, use a sliding window to traverse and statistically obtain the zero-value threshold ThrZero for each counting position, and regard the count values ​​less than or equal to the corresponding zero-value threshold as zero values; then use a sliding window to traverse the counting histogram processed according to the zero-value threshold, count the number of zero count values ​​in the window and the average count in the window for each counting position, and substitute them into the definition of the exponential function to obtain the clustering parameter Eau of the count values, and set the clustering threshold ThrEau; Step 3: Divide the enhanced count histogram length T into multiple equal-length non-overlapping sub-time intervals, and select intervals that may be echoes based on the clustering parameter Eau and the threshold ThrEau. For the count values ​​at positions where the clustering is greater than the threshold ThrEau, enhance them non-uniformly based on the clustering. Step four: Use a Gaussian function to define the relative difference between count values ​​in the time neighborhood, with the difference value between 0 and 1; within each retained echo interval, calculate the total relative difference of the count values ​​in the time neighborhood to characterize the specificity of the count in the neighborhood. Step 5: Within the remaining possible echo intervals, the echo position is optimized by using the specificity of the count value in the time neighborhood as the objective function, thereby obtaining the laser pulse flight time and calculating the distance. Step 6: Process each pixel using steps 1 to 5, and construct a 3D image based on the obtained distance values.

2. The adaptive photon denoising three-dimensional imaging method under sparse or noisy counting as described in claim 1, characterized in that, In step one, the echo counting enhancement formula is: In the formula, V (i,j) w represents a square region centered on the currently processed pixel (i, j). (k，z) ∈[0,1] represents the pixel weight, Y (k,z) Y' is a count histogram of the neighboring pixels (k, z) of pixel (i, j). (i,j) This is the count histogram for the enhanced pixel (i, j).

3. The adaptive photon denoising three-dimensional imaging method under sparse or noisy counting as described in claim 2, characterized in that, In step two, the process of obtaining the zero-value threshold ThrZero is as follows: traverse the counting histogram through a sliding window, and take the largest count value within the window as the zero-value threshold ThrZero at the location of the sliding window; if the count value is less than or equal to the zero-value threshold ThrZero, it is regarded as zero in subsequent processing, otherwise it is retained.

4. The adaptive photon denoising three-dimensional imaging method under sparse or noisy counting as described in claim 3, characterized in that, In step two, the calculation process for the clustering parameter Eau is as follows: Water (i,j),t *Ea (i,j),t / max({Ea (i,j),k |k=1,2,...,T}) In the formula, Eau (i,j),t Wn represents the normalized clustering parameter at the t-th counting position of pixel (i, j); t This represents the time range contained at the t-th count position of the sliding window center in the histogram; Nzero (i,j),t Let b be the number of zero count values ​​within the window when the center of the sliding window is at the t-th counting position in the histogram, where b>0 is a constant.

5. The adaptive photon denoising three-dimensional imaging method under sparse or noisy counting as described in claim 4, characterized in that, In step two, the value of Eau is in the range of (0,1], and the aggregation threshold ThrEau is set to 0.5, so as to remove most of the noise counting area while avoiding the removal of the real echo counting area.

6. The adaptive photon denoising three-dimensional imaging method under sparse or noisy counting as described in claim 4, characterized in that, In step three, let T be divided into s subintervals, then we have: Among them, T l This represents the l-th sub-time interval obtained after division, where the laser pulse duration Tp is on the order of nanoseconds, and the length L of the sub-time interval is... sub Set to 10 times Tp.

7. The adaptive photon denoising three-dimensional imaging method under sparse or noisy counting as described in claim 6, characterized in that, In step three, the process of selecting and retaining possible echo intervals based on clustering is as follows: sub-time intervals with clustering greater than a set threshold ThrEau are retained and considered as possible echo intervals; otherwise, they are considered noise intervals. For the count values ​​within the retained possible echo intervals, non-uniform enhancement is applied based on clustering. (i,j),t The enhanced count data is represented by the non-uniform enhancement method: (1) If Eau (i,j),t ≥ThrEau (i,j),t To enhance, Yen' (i,j),t =Y' (i,j),t ·(1+Eau (i,j),t ); (2) If Eau (i,j),t <ThrEau (i,j),t Without enhancement, Yen' (i,j),t =Y' (i,j),t .

8. The adaptive photon denoising three-dimensional imaging method under sparse or noisy counting as described in claim 7, characterized in that, In step four, record the current time position count value. Mean of small neighborhood counts at other time locations The degree of relative difference between them The relative degree of difference in the definition of the Gaussian function as follows: In the formula, This indicates that within the l-th sub-time interval of pixel (i, j), the t-th time interval... c Count value With the tth o Average number of counts The relative degree of difference; the greater the difference, the better. The larger the value, the greater the σ. st >0 is a parameter for adjusting the sensitivity to differences.

9. The adaptive photon denoising three-dimensional imaging method under sparse or noisy counting as described in claim 8, characterized in that, In step four, within the l-th sub-time interval, The degree of relative difference relative to all other locations in the neighborhood. for:

10. The adaptive photon denoising three-dimensional imaging method under sparse or noisy counting as described in claim 8, characterized in that, In step five, all possible echo intervals preserved based on clustering are traversed and processed. The relative difference within the neighborhood is used as the objective function for constrained optimization to obtain the echo location. The constrained optimization formula for the echo location is as follows: In the formula, This indicates the sub-time interval number where the echo obtained from the optimization solution is located. Indicates the echo at the 1st Location within a specific time interval, TV (i,j) This represents the smoothing penalty term introduced during optimization. The distance between pixels (i, j) is as follows: In the formula, This represents the estimated absolute position of the echo at pixel (i,j) in the counting histogram. This represents the distance estimate at pixel (i,j).

Images