Two-stage compression denoising imaging method for large-area array Geiger mode APD (avalanche photo diode)

Abstract

The invention discloses a two-stage compression denoising imaging method for a large-area array Geiger mode APD (avalanche photo diode). The method comprises the following steps: intercepting photon events in a pre-sampling time window at the front part of a data block to carry out global histogram statistics, fitting and estimating time domain characteristics of background noise based on Poisson distribution, screening and compressing full-frame data, and eliminating most noise; carrying out local joint histogram statistics and peak value extraction on the compressed data based on a pixel space neighborhood, and reconstructing a preliminary distance image and a preliminary intensity image; if the effective data in the neighborhood of the data meets the flat area condition, standard median filtering is executed; otherwise, starting a robust filtering algorithm based on absolute median difference MAD for processing. According to the method, the cooperation of front-end data efficient compression and rear-end accurate denoising and edge preserving is realized, and the method has the advantages of high compression ratio, high adaptability and excellent imaging quality, and is particularly suitable for real-time three-dimensional imaging of a large-area-array GM-APD laser radar system.

Description

A Two-Stage Compression and Denoising Imaging Method for Large-Area Geiger Mode APDs Technical Field

[0001] This invention relates to the field of photoelectric detection and image processing technology, and particularly to a data processing method suitable for large-area Geiger-mode Avalanche Photodiode (GM-APD) lidar systems. Specifically, it is a two-stage compression and denoising joint processing method based on Poisson noise statistical modeling and local spatial correlation characteristics. This method can be used in lidar, single-photon counting imaging, long-range weak target detection, and 3D reconstruction scenarios. Background Technology

[0002] Geiger-mode APDs, with their advantages of single-photon-level sensitivity and picosecond-level temporal resolution, have become the core detectors for LiDAR 3D imaging. As the size of focal plane arrays continues to expand to the thousands and even millions of pixels, the amount of photon event data generated per frame grows exponentially, posing a severe challenge to the system's data storage, transmission bandwidth, and real-time processing capabilities. Furthermore, inherent noise sources in GM-APDs during operation, such as dark counting, afterpulse, background shot noise, and inter-pixel crosstalk, can severely obliterate effective signals under weak signal detection conditions, leading to a sharp decline in image signal-to-noise ratio (SNR) and ranging accuracy.

[0003] Existing denoising methods mostly focus on one-sided processing in the time or spatial domains, which has obvious limitations: Time-domain methods, such as fixed / adaptive time gating or histogram threshold filtering, extract echo signals by setting empirical thresholds. These methods are difficult to adapt to complex and changing noise environments and are prone to two types of errors under extremely low signal-to-noise ratio conditions: first, misjudging noise as a signal and retaining too much noise; second, misjudging weak effective signals as noise and discarding them, resulting in signal loss.

[0004] Spatial domain methods, such as mean filtering, Gaussian filtering, or median filtering, can smooth noise to a certain extent, but at the cost of sacrificing spatial resolution and detail features, they are prone to causing edge blurring and loss of detail.

[0005] Learning-based methods, such as deep learning denoising algorithms based on convolutional neural networks (CNN), may perform well in specific scenarios, but their models are complex, computationally expensive, rely on a large amount of training data, and have questionable generalization capabilities, making it difficult to meet the stringent requirements of large-area GM-APD systems for high frame rates, low power consumption, and strong real-time performance.

[0006] More importantly, the aforementioned existing technologies all share a common flaw: they fail to fully utilize the statistical prior knowledge that GM-APD photon events follow a Poisson distribution, and they fail to integrate front-end data compression and back-end signal denoising in a coordinated design, resulting in a tradeoff between algorithm efficiency and performance.

[0007] Therefore, in order to address the application bottlenecks of large-area GM-APD, there is an urgent need for a joint processing method that can deeply integrate the physical model of noise and the statistical characteristics of signals, achieve efficient data compression in the early stage of imaging, and perform precise noise reduction in the later stage, so as to meet the real-time requirements of the system while ensuring imaging quality. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a two-stage compression and denoising imaging method for large-area Geiger mode APDs. This method employs a two-stage processing strategy of "coarse screening compression" and "fine reconstruction denoising." In the early stage, it aims to quickly remove most of the noise data for efficient compression. In the later stage, it utilizes local spatial statistical information to accurately reconstruct the three-dimensional image, ultimately achieving a synergistic optimization of data volume, computational efficiency, and imaging signal-to-noise ratio. This method is suitable for engineering applications with stringent requirements for power consumption and real-time performance.

[0009] The technical solution of this invention is: a two-stage compression and denoising imaging method for large-area Geiger mode APDs, comprising the following steps: acquiring the raw photon event stream output by the GM-APD detector, generating corresponding raw image data, and dividing it into multiple data blocks according to a fixed time frame period, each data block containing multiple frames of complete raw image data; from the data blocks, acquiring multiple frames of raw image data within the first pre-sampling time window, performing global time histogram statistics on the photon events of pixels in the raw image data, and estimating the noise curve based on Poisson distribution fitting, obtaining the difference histogram to dynamically determine the effective signal range gate; using the effective signal range gate for filtering and compression, obtaining each The original image data of each frame corresponds to the coarse-screened image data; all coarse-screened image data in the same data block are equally divided into m denoised image groups; local joint time histogram statistics and peak extraction are performed based on the spatial neighborhood of pixels to reconstruct m-frame distance images D and m-frame intensity images I; the m-frame distance images D are optimized: for each pixel in the distance image D, if the effective data in its neighborhood meets the flat region condition, standard median filtering is performed; otherwise, a robust filtering algorithm based on absolute median difference is used to process it, traversing each pixel to obtain the optimized distance image D'; the reconstructed intensity image I and the optimized distance image D' are output as the final reconstruction result image.

[0010] Preferably, the method for filtering and compressing to obtain the coarse-screened image data corresponding to each frame of original image data is as follows: 21) Extract the first pre-sampling time window from the currently processed data block. 22) Use the original image data from multiple frames within the front-end image data as the front-end image data; 3) Perform global time histogram accumulation on the photon events of all pixels in the front-end image data to generate a global time histogram. 23) Based on global time histogram By fitting the Poisson distribution curve using the maximum likelihood estimation method, a noise curve characterizing the time distribution of background noise is obtained. 24) From the global time histogram and noise curve Calculate the difference histogram 25) Based on the aforementioned difference histogram mean and standard deviation ,use Criteria set dynamic thresholds 26) From the difference histogram In the selection, the intensity value continuously exceeds the dynamic threshold. 27) Traverse all the original image data of the current data block, discard all photon events in the pixels whose timestamps are not within the effective signal distance gate, and obtain the coarse image data corresponding to each frame of original image data.

[0011] Preferably, a dynamic threshold is set in step 25). The method is as follows: Where K ranges from 1 to 3.

[0012] Preferably, the method for reconstructing m-frame distance images D and m-frame intensity images I specifically includes: 31) dividing all coarse-screened image data in the same data block into m denoised image groups; 32) for each pixel in the coarse-screened image data, aggregating all photon events of the pixel itself and its 8 spatially neighboring pixels, and statistically analyzing the aggregated photon events by timestamp to generate local joint time histograms corresponding to each pixel in the denoised image group. 33) Find the local joint time histogram corresponding to each pixel. The main peak value in the image is used as the timestamp corresponding to the main peak value as the precise distance value d of the pixel, thus obtaining the distance image D corresponding to the current denoised image group; 34) The local joint temporal histogram is used as the distance image D. The amplitude of the main peak is used as the intensity value I of the pixel to obtain the intensity image I corresponding to the current denoised image group; 35) Repeat steps 32) to 34) traverse the m denoised image groups in the current data block to obtain the m-frame high signal-to-noise ratio distance image D and the m-frame high signal-to-noise ratio intensity image I of the current data block.

[0013] Preferably, the method for optimizing the m-frame distance image D is as follows: 41) Select any pixel in the distance image D and determine whether the current pixel simultaneously satisfies the first filtering condition and the second filtering condition. If it simultaneously satisfies the first filtering condition and the second filtering condition, proceed to step 42) to perform standard median filtering; otherwise, proceed to step 43) to process using a robust filtering algorithm based on absolute median difference; 42) Obtain the effective distance values ​​of all pixels within the neighborhood window of the current pixel, calculate the median M of the obtained effective distance values ​​as the final distance value of the current pixel, and return to step 41) to traverse. After all pixels are traversed, proceed to step 44); the neighborhood window is a region of p x p pixels centered at the current pixel, where p is a positive integer greater than or equal to 3; 43) obtain the effective distance values ​​of all pixels within the neighborhood window of the current pixel, determine the median M and the median absolute deviation MAD of the effective distance values ​​from the obtained effective distance values ​​and form a robust set, obtain the median of the effective distance values ​​belonging to the robust set within the neighborhood window, and use it as the final distance value of the pixel, return to step 41); after traversing all pixels, proceed to step 44); 44) after traversing all pixels, obtain the optimized distance image D'.

[0014] Preferably, the first and second filtering conditions in step 41) are specifically as follows: First filtering condition: The number of pixels with non-zero effective distance values ​​within the neighborhood window is greater than or equal to a first threshold N1; the first threshold N1 is not less than half the total number of pixels in the neighborhood window; Second filtering condition: The standard deviation of the effective distance values ​​of all pixels within the neighborhood window is less than or equal to the second threshold. Second threshold The value range is 1 to 3.

[0015] Preferably, step 43) involves obtaining the median of the effective distance values ​​belonging to the robust set within the neighborhood window and using it as the final distance value of the pixel. Specifically, this involves: obtaining pixels within the neighborhood window where the current pixel's effective distance value is not zero, identifying these pixels as effective pixels; obtaining the effective distance values ​​of all effective pixels, and sorting them by size to obtain the median, which is then used as the median M of the effective pixels; obtaining the difference between the effective distance value of each effective pixel and the median M, and sorting them by size to obtain the median, which is then used as the median of the absolute deviation (MAD); within the neighborhood window, obtaining the values ​​belonging to the robust set... The effective distance values ​​are sorted by size, and the median is used as the final distance value of the current pixel; where k is a proportional constant.

[0016] Preferably, the value of the proportionality constant k ranges from 0.5 to 1.5.

[0017] Compared with the prior art, the advantages of the present invention are mainly reflected in the following aspects: 1. Extremely high data compression efficiency: Through the coarse screening stage based on the noise statistical model, about 80%-95% of the noise data can be removed at the front end of the data processing link, which greatly alleviates the pressure on back-end transmission and storage, and provides a basis for system miniaturization and reduction of bandwidth costs.

[0018] 2. Denoising is based on sound physical principles and is highly adaptive: It innovatively uses a presampling window to estimate the Poisson distribution characteristics of background noise, which is consistent with the physical working principle of GM-APD. This makes denoising and distance gate extraction more scientific and adaptive, and overcomes the shortcomings of traditional fixed threshold methods that suffer from a sharp drop in performance under complex environments.

[0019] 3. High real-time performance: By adopting the strategies of "staged processing" and "frame-by-frame fitting", the computational resources are concentrated on key steps, avoiding the large amount of computation of complex algorithms such as CNN, meeting the real-time processing requirements of large-area array high frame rate systems, and the algorithm is easy to implement on FPGA or DSP.

[0020] 4. Significantly improved imaging quality: The second stage utilizes local spatial correlation for peak extraction, effectively aggregating sparse photon signals and suppressing random noise. With the addition of intelligent post-processing filtering, the final output range and intensity images have higher signal-to-noise ratios, better spatial coherence, and better detail retention.

[0021] 5. Strong engineering practicality and deployability: The method has a clear process and strong modularity, and is very easy to implement through hardware acceleration using FPGA or ASIC. It is very suitable for integration into vehicle-mounted, airborne or portable GM-APD lidar systems with strict requirements for power consumption, size and real-time performance. Attached Figure Description

[0022] Figure 1 is a schematic diagram of the overall process of a two-stage compression and denoising imaging method for large-area Geiger mode APD.

[0023] Figure 2 is a schematic diagram of the first-stage coarse denoising algorithm of the present invention.

[0024] Figure 3 is a schematic diagram of the second-stage fine denoising algorithm of the present invention.

[0025] Figure 4 is a schematic diagram of the depth map post-processing algorithm of the present invention.

[0026] Figure 5 shows the experimental results of a two-stage compression denoising imaging method in one embodiment of the present invention. Detailed Implementation

[0027] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, but the implementation of the present invention is not limited thereto. This embodiment is carried out based on the technical solution of the present invention, and provides detailed implementation methods and processes, but the scope of protection of the present invention is not limited to the following embodiments.

[0028] A two-stage compression and denoising imaging method for large-area Geiger mode APDs, as shown in Figure 1, includes the following steps: S1: Data segmentation and preprocessing: Acquire the raw photon event stream output by the GM-APD detector and generate the corresponding raw image data. In one embodiment of the present invention, the resolution of the raw image data is 128 x 128. Divide the raw image data into continuous data blocks according to a fixed time frame period. Each data block contains N frames of complete raw image data. In one embodiment of the present invention, N is 1000, that is, each data block contains 1000 frames of raw image data.

[0029] S2: Based on the noise statistical model, coarse screening and compression are performed to obtain coarse screened image data corresponding to each frame of original image data; in one embodiment of the present invention, 10,000 frames of original image data correspond to 10,000 frames of coarse screened image data.

[0030] S2.1: Presampling: Extracting a presampling time window from the currently processed data block. The original image data of all frames corresponding to the photon event data within the window is used as the front-end image data; in one embodiment of the present invention, the presampling time window... The first 100 frames of image data in the current data block are used as the front-end image data.

[0031] S2.2: Global Statistics: Accumulate photon events for all pixels in the front-end image data to generate a global time histogram. In one embodiment of the present invention, a global temporal histogram is generated corresponding to the first 100 frames of image data in the data block. .

[0032] S2.3: Noise Modeling: Based on the Global Time Histogram By fitting the Poisson distribution curve using the maximum likelihood estimation method, a noise curve characterizing the time distribution of background noise is obtained. That is, the Poisson distribution fitting in step S2.3 is not performed in every frame, but once every M frames (M≥1), and the noise curve obtained from the most recent fitting is directly reused in the intermediate frames. This significantly reduces computational complexity. In the time histogram, the horizontal axis represents time points, and the vertical axis represents the cumulative number of pixels at the same time point.

[0033] S2.4: Noise Suppression: Based on the global time histogram and noise curve Calculate the difference histogram In one embodiment of the present invention, the first 100 frames of original image data in the data block correspond to a "difference histogram". (That is, the portion of Figure 5(c) where the intensity value exceeds the noise fitting curve.) S2.5: Dynamic threshold calculation: based on the aforementioned difference histogram. Statistical properties (such as mean) Standard deviation ),use Criteria set dynamic thresholds ,Right now ;in, and The difference histograms are respectively mean and standard deviation K is a constant between 1 and 3.

[0034] S2.6: Distance gate extraction: Obtain the difference histogram obtained in step 2.4. In the middle, the intensity value continuously exceeds the dynamic threshold. The time interval is used as the effective signal distance gate; S2.7: Data compression: Traverse all the original image data of the current data block, discard all photon events in the pixels of the original image data whose timestamps are not within the effective signal distance gate, realize the initial compression and coarse screening of the data, and obtain the coarse screened image data corresponding to each frame of original image data.

[0035] S3: Based on spatial correlation statistics, fine reconstruction and denoising are performed to obtain m-frame high signal-to-noise ratio distance image D and m-frame high signal-to-noise ratio intensity image I of the current data block.

[0036] S3.0: Divide all coarsely screened image data in the same data block into m denoised image groups; S3.1: Neighborhood aggregation: For each pixel in the obtained coarsely screened image data, aggregate all photon events of itself and its 8 spatial neighboring pixels; S3.2: Local statistics: Statistically analyze the aggregated photon events according to timestamps to generate a local joint time histogram corresponding to each pixel in each denoised image group. The information in each pixel of the time histogram is the effective distance value.

[0037] S3.3: Signal Extraction: Find the local joint time histogram for each pixel. The dominant peak value in the image is used as the precise distance value d for that pixel, corresponding to the timestamp (i.e., time position) of the dominant peak value; this yields the high signal-to-noise ratio distance image D corresponding to the current denoised image group; S3.4: Intensity generation: the local joint time histogram is used as the distance value d for that pixel. The amplitude of the main peak is used as the intensity value I of the pixel to obtain the high signal-to-noise ratio intensity image I corresponding to the current denoised image group; S3.5: Image generation: Repeat steps 3.1-3.4, traverse the m denoised image groups in the current data block to obtain the m-frame high signal-to-noise ratio distance image D and the m-frame high signal-to-noise ratio intensity image I of the current data block.

[0038] In one embodiment of the present invention, 10,000 frames of image data are compressed into 100 high signal-to-noise ratio intensity images I and 100 high quality distance images D by coarse screening compression and fine reconstruction denoising.

[0039] S4: Robust Filtering Optimization Based on Conditional Judgments S4.1: Branch Filtering Process: If a pixel in image D satisfies two filtering conditions, it is determined that the pixel is in a flat region, and proceeds to step S4.2a; otherwise, it is determined that the pixel may be in an edge or feature region, and the robust filtering algorithm is activated, proceeding to step S4.2b; thus, while efficiently suppressing noise, the image edge features are preserved to the extreme. S4.2a: Obtain the effective distance values ​​of all pixels within the neighborhood window of the current pixel, and calculate the median M of the obtained effective distance values ​​as the final distance value of the current pixel; the neighborhood window is a region of p x p pixels centered on the current pixel, where p is a positive integer greater than or equal to 3; S4.2b: Obtain... The effective distance values ​​of all pixels within the neighborhood window of the current pixel are used to determine the median M and median absolute deviation MAD of the effective distance values, forming a robust set. The median of the effective distance values ​​belonging to the robust set within the neighborhood window is then obtained and used as the final distance value for the pixel. Specifically: within the neighborhood window of the current pixel, pixels with non-zero effective distance values ​​are considered valid pixels; the effective distance values ​​of all valid pixels are obtained, and the median is sorted by size and used as the median M of the valid pixels; the difference between the effective distance value of each valid pixel and the median M is obtained, and the median is sorted by size and used as the median absolute deviation MAD; within the neighborhood window, the values ​​belonging to the robust set are considered... The effective distance values ​​within the range are sorted by size, and the median is used as the final distance value for the current pixel; where k is a scaling constant. The scaling constant k ranges from 0.5 to 1.5.

[0040] First filtering condition: For each pixel in the distance image D, determine whether the number of effective distance values ​​of pixels within the neighborhood window (preferably a 3×3 window) that are not zero is greater than or equal to a first threshold N1; the first threshold N1 in step S4.1 is not less than half the total number of pixels in the neighborhood window. In one embodiment of the present invention, the neighborhood window is 3×3, and N1 equals 5.

[0041] Second filtering condition: The standard deviation of the effective distance values ​​of all pixels within the neighborhood window (preferably a 3×3 window) is less than or equal to the second threshold. Second threshold The second threshold is set based on the system's ranging accuracy and noise level. The value range is 1 to 3.

[0042] S4.3: Image Output: After traversing all pixels and completing the above filtering process, the optimized distance image D' is obtained.

[0043] This invention fully utilizes the Poisson statistical properties and spatial correlation of signals in GM-APD to achieve synergy between efficient front-end data compression and precise back-end denoising and edge preservation. It boasts advantages such as high compression ratio, strong adaptability, and excellent imaging quality, making it particularly suitable for real-time 3D imaging in large-area GM-APD lidar systems. This method can be applied to single-photon lidar, 3D imaging, long-range target detection, or optical imaging systems in low-light environments.

[0044] Example: This example uses a large-area GM-APD lidar system with a pixel size of 128 × 128 pixels. The system operates at a wavelength of 1064nm, a laser repetition frequency of 10kHz, a single-frame imaging time of 10ms, and a data block containing data from 10 consecutive frames. The imaging cycle is... The timestamp precision is 100ms (time bin width) and is 1ns.

[0045] S1: Data Segmentation and Preprocessing: The system continuously collects data according to the frame period. Within every 10 frames (100ms), all photon events generated by all pixels (including timestamps and pixel coordinates) are combined into an independent data block, which is then handed over to the subsequent processing flow.

[0046] S2: First Stage - Coarse Screening and Compression Based on Noise Modeling, as shown in Figure 2: S2.1 (Presampling): Taking n=10, the presampling time window length is... Extract all photon events within the first 1ms of the current data block, i.e., the first frame of the data block.

[0047] S2.2 (Global Statistics): All photon events of the entire array (128 × 128 pixels) within this 1ms are accumulated with a bin width of 1ns to generate a global time histogram containing 3000 time bins. .

[0048] S2.3 (Noise Modeling): The maximum likelihood estimation method is used for... By fitting a Poisson distribution curve, the noise rate parameter is obtained. And generate noise fitting curves. To reduce computational load, this example sets up a fitting operation every M=10 frames, with intermediate frames directly reusing the most recent fitting result. .

[0049] S2.4 (Noise Suppression): Calculate the difference histogram .

[0050] S2.5 (Dynamic Threshold Calculation): Calculation mean and standard deviation .according to Criteria for setting dynamic thresholds .

[0051] S2.6 (Distance Gate Extraction): Search All consecutively exceeding The time interval is determined as the effective signal distance gate.

[0052] S2.7 (Data Compression): Traverse the entire current 10ms data frame and discard all photon events whose timestamps are not within the distance gate. This step removes approximately 85% of the data, achieving significant compression.

[0053] S3: Second Stage - Fine Reconstruction and Denoising Based on Spatial Domain Correlation Statistics, the process is shown in Figure 3: S3.1 (Neighborhood Aggregation): For the compressed data, each pixel in the array to be processed... It aggregates all photon events within itself and its 8 spatial neighboring pixels (i.e., a 3×3 pixel window).

[0054] S3.2 (Local Statistics): Regenerate the local joint time histogram for all photon events aggregated into this 3×3 window. .

[0055] S3.3 / S3.4 (Signal Extraction and Intensity Generation): Searching The dominant peak value is used. The time position corresponding to this dominant peak value is taken as the distance value of that pixel. The amplitude of the main peak value is used as the intensity value of the pixel. .

[0056] S3.5 (Image Generation): Traverse all 128×128 pixels, repeating steps S3.1 to S3.4 to finally generate a preliminary distance image. And the intensity is like I.

[0057] S4: Robust filtering optimization based on conditional judgment: Optimizing the initially obtained range profile A robust filtering process based on conditional judgment is performed, as shown in Figure 4. The specific steps are as follows: a 3×3 filtering window is used.

[0058] S4.1 (Conditional Judgment): For each pixel to be processed in the distance image Determine whether the number of valid distance values ​​within its 3×3 neighborhood is greater than or equal to the first threshold N1 = 5, and whether the standard deviation of these valid distance values ​​is less than or equal to the second threshold. .

[0059] S4.2 (Branch Filtering): S4.2a (Flat Region Processing): If pixels If both of the above conditions are met, the pixel is determined to be in a flat region. The standard median of all valid distance values ​​within this neighborhood is taken as the final distance value for that pixel.

[0060] S4.2b (Edge / Feature Region Processing): If pixel If any of the above conditions are not met, it is determined that it may be located in an edge or feature region. Therefore, a robust filtering algorithm based on median absolute difference (MAD) is enabled, and the specific process is as follows: 1. Calculate the median M of all valid distance values ​​in the neighborhood.

[0061] 2. Calculate the absolute deviation of each valid distance value from the median M.

[0062] 3. Calculate the median (MAD) of these absolute deviations.

[0063] 4. Set dynamic range The proportionality constant k takes the value of 1.

[0064] 5. From the effective distance values ​​in the neighborhood, select all values ​​that fall within the above dynamic range to form a "robust set".

[0065] 6. Calculate the median of this "robust set" as the pixel value. The final distance value.

[0066] S4.3 (Image Output): After traversing all pixels and completing the above branch filtering process, the final high-quality, high-fidelity distance image D is obtained.

[0067] Figure 5 shows the experimental results of the two-stage compression denoising imaging method in one embodiment of the present invention. Figure 5(a) is the target grayscale image used to verify the present invention. Figure 5(b) is the peak depth map without any processing, serving as a reference image for the present invention. Figure 5(c) is the fitting curve between the global histogram obtained from the first-stage presampling and the Poisson noise. Figure 5(d) is the target intensity map obtained from the second-stage fine denoising. Figure 5(e) is the target depth map obtained from the second-stage fine denoising. Figure 5(f) is the final target depth map obtained by post-processing the second-stage target depth map.

[0068] While the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the invention. Any person skilled in the art can make possible variations and modifications to the technical solutions of the present invention using the disclosed methods and techniques without departing from the spirit and scope of the invention. Therefore, any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention, without departing from the content of the technical solutions of the present invention, shall fall within the protection scope of the present invention. Where there is no conflict, the embodiments of this application and the technical features thereof can be combined with each other.

[0069] The contents not described in detail in this specification are common knowledge to those skilled in the art.

Claims

1. A two-stage compression and denoising imaging method for large-area Geiger mode APDs, characterized in that, The steps include: acquiring the raw photon event stream output by the GM-APD detector, generating corresponding raw image data, and dividing it into multiple data blocks according to a fixed time frame period, with each data block containing multiple frames of complete raw image data; acquiring multiple frames of raw image data within the first pre-sampling time window from the data block, performing global time histogram statistics on the photon events of pixels in the raw image data, and estimating the noise curve based on Poisson distribution fitting, obtaining the difference histogram to dynamically determine the effective signal range gate; using the effective signal range gate for filtering and compression to obtain the coarse-screened image data corresponding to each frame of raw image data; and dividing all coarse-screened image data in the same data block into m denoised image groups. Local joint temporal histogram statistics and peak extraction are performed based on the spatial neighborhood of pixels to reconstruct m-frame distance images D and m-frame intensity images I. The m-frame distance images D are then optimized: for each pixel in the distance image D, if the effective data in its neighborhood meets the flat region condition, standard median filtering is performed; otherwise, a robust filtering algorithm based on absolute median difference is used to process the data, and the optimized distance image D' is obtained by traversing each pixel. The reconstructed intensity image I and the optimized distance image D' are then output as the final reconstruction result image.

2. The two-stage compression and denoising imaging method for large-area Geiger mode APD according to claim 1, characterized in that, The method for filtering and compressing to obtain the coarse-screened image data corresponding to each frame of original image data is as follows: 21) Extract the first presampling time window from the currently processed data block. 22) Use the original image data from multiple frames within the front-end image data as the front-end image data; 3) Perform global time histogram accumulation on the photon events of all pixels in the front-end image data to generate a global time histogram. ； 23) Based on global time histogram By fitting the Poisson distribution curve using the maximum likelihood estimation method, a noise curve characterizing the time distribution of background noise is obtained. 24) From the global time histogram and noise curve Calculate the difference histogram 25) Based on the aforementioned difference histogram mean and standard deviation ,use Criteria set dynamic thresholds 26) From the difference histogram In the selection, the intensity value continuously exceeds the dynamic threshold. 27) Traverse all the original image data of the current data block, discard all photon events in the pixels whose timestamps are not within the effective signal distance gate, and obtain the coarse image data corresponding to each frame of original image data.

3. The two-stage compression and denoising imaging method for large-area Geiger mode APD according to claim 2, characterized in that, In step 25), a dynamic threshold is set. The method is as follows: Where K ranges from 1 to 3.

4. A two-stage compression and denoising imaging method for large-area Geiger mode APD according to any one of claims 1-3, characterized in that, The method for reconstructing the m-frame distance image D and the m-frame intensity image I is as follows: 31) Divide all coarse-screened image data in the same data block into m denoised image groups; 32) For each pixel in the coarse-screened image data, aggregate all photon events of the pixel itself and its 8 spatial neighboring pixels, and statistically analyze the aggregated photon events according to their timestamps to generate local joint time histograms corresponding to each pixel in the denoised image group. 33) Find the local joint time histogram corresponding to each pixel. The main peak value in the image is used as the timestamp corresponding to the main peak value as the precise distance value d of the pixel, thus obtaining the distance image D corresponding to the current denoised image group; 34) The local joint temporal histogram is used as the distance image D. The amplitude of the main peak is used as the intensity value I of the pixel to obtain the intensity image I corresponding to the current denoised image group; 35) Repeat steps 32) to 34) traverse the m denoised image groups in the current data block to obtain the m-frame high signal-to-noise ratio distance image D and the m-frame high signal-to-noise ratio intensity image I of the current data block.

5. A two-stage compression and denoising imaging method for large-area Geiger mode APDs according to claim 4, characterized in that, The method for optimizing the m-frame distance image D is as follows: 41) Select any pixel in the distance image D and determine whether the current pixel simultaneously satisfies the first and second filtering conditions. If it satisfies both conditions, proceed to step 42) to perform standard median filtering; otherwise, proceed to step 43) to process the pixel using a robust filtering algorithm based on absolute median difference. 42) Obtain the effective distance values ​​of all pixels within the neighborhood window of the current pixel, and calculate the median M of the obtained effective distance values ​​as the final distance value of the current pixel. Return to step 41) After traversing all pixels, proceed to step 44); The neighborhood window is a region of p x p pixels centered at the current pixel, where p is a positive integer greater than or equal to 3; 43) Obtain the effective distance values ​​of all pixels within the neighborhood window of the current pixel. Based on the obtained effective distance values, determine the median M and the median absolute deviation MAD of the effective distance values ​​and construct a robust set. Obtain the median of the effective distance values ​​belonging to the robust set within the neighborhood window, and use it as the final distance value of the pixel. Return to step 41) After traversing all pixels, proceed to step 44); 44) After traversing all pixels, the optimized distance image D' is obtained.

6. A two-stage compression and denoising imaging method for large-area Geiger mode APDs according to claim 5, characterized in that, The first and second filtering conditions in step 41) are as follows: First filtering condition: The number of effective distance values ​​of pixels in the neighborhood window that are not zero is greater than or equal to the first threshold N1; the first threshold N1 is not less than half of the total number of pixels in the neighborhood window; Second filtering condition: The standard deviation of the effective distance values ​​of all pixels within the neighborhood window is less than or equal to the second threshold. Second threshold The value range is 1 to 3.

7. A two-stage compression and denoising imaging method for large-area Geiger mode APDs according to claim 6, characterized in that, Step 43) The method for obtaining the median of the effective distance values ​​belonging to the robust set within the neighborhood window and using it as the final distance value of the pixel is as follows: within the neighborhood window of the current pixel, the pixels with effective distance values ​​that are not zero are taken as effective pixels; the effective distance values ​​of all effective pixels are obtained, and the median is obtained by sorting by size, which is used as the median M of the effective pixels. Obtain the difference between the effective distance value and the median M for each effective pixel, sort them by size, and obtain the median as the median MAD of the absolute deviation; within the neighborhood window, obtain the value belonging to the robust set. The effective distance values ​​are sorted by size, and the median is used as the final distance value of the current pixel; where k is a proportional constant.

8. A two-stage compression and denoising imaging method for large-area Geiger mode APDs according to claim 7, characterized in that, The value of the proportionality constant k ranges from 0.5 to 1.5.

Images