Stacked two-dimensional code display, reading and checking management system based on reinforcement learning

Abstract

This invention relates to the field of information recognition and data processing technology, specifically a stacked QR code display, reading, and verification management system based on reinforcement learning. It includes modules for image acquisition, consumption estimation, strategy scheduling, dynamic code display, and asynchronous verification. The system utilizes image spectral feature analysis to construct expected decoding computing power consumption and monitors the system's remaining computing power in real time. Its core principle is to perform hierarchical dimensionality reduction based on business priorities when computing power is insufficient, shielding non-critical areas and only decoding key levels to generate access credentials. Then, it uses a background statistical model to asynchronously infer and reconstruct undecoded data. This invention combines environmental perception and hardware thermal exhaustion coefficients to achieve accurate mapping from theoretical models to physical states, effectively avoiding gate deadlock and response timeouts in high-throughput scenarios, and significantly improving system robustness.

Description

Technical Field

[0001] This invention relates to the field of information recognition and data processing technology, specifically to a stacked QR code display, reading and verification management system based on reinforcement learning. Background Technology

[0002] In high-throughput scenarios such as multinational mega-transportation hubs, high-density stacked QR codes containing multiple encrypted rights serve as optical data carriers, carrying massive amounts of data such as identity authentication, ticket verification, and commercial points. Reading devices need to scan and parse the access medium in real time. Existing QR code display and reading systems typically adopt a full-data parsing mode, which involves uniformly enhancing the collected carrier image and decoding the data at all levels. Only after all information is extracted and verified is the access command triggered. This approach can ensure the integrity of information under ideal conditions, but when faced with complex optical interference such as physical wear of the medium, strong light reflection, or shadow occlusion, the floating-point operations required for image decoding increase exponentially. However, under long-term high-load operation, terminal devices are prone to chip thermal exhaustion, leading to dynamic fluctuations in the available computing power. Existing rigid decoding strategies lack awareness of the correlation between image frequency domain noise characteristics and the current computing power status. When the computing demand exceeds the computing power supply, full-level decoding is still forcibly executed, which can easily lead to computing timeouts, response delays, or even system deadlocks, causing physical access congestion.

[0003] Therefore, how to establish a dynamic balance mechanism between computing consumption and computing resources under extreme conditions such as fluctuations in hardware computing power and damage to image quality, and how to ensure the real-time performance and ledger accuracy of high-concurrency access through adaptive adjustment of decoding strategies and asynchronous verification of data, is a technical problem that urgently needs to be solved. Summary of the Invention

[0004] In view of the shortcomings of the prior art, the purpose of this invention is to provide a stacked QR code display, reading, and verification management system based on reinforcement learning, which can solve the technical problems existing in the prior art. Specifically, the purpose of this invention can be achieved through the following technical solutions: The image acquisition module is configured to perform real-time scanning of the optically stacked data carrier on the passage medium to acquire the original carrier image; The consumption prediction module is configured to perform spectral feature analysis on the original carrier image, extract frequency domain noise features, and combine them with the preset decoding model complexity to construct the expected decoding computing power consumption of the current original carrier image. The strategy scheduling module is configured to monitor the current system's computing power reserve in real time, compare the expected decoding computing power consumption with the computing power reserve, and if the expected decoding computing power consumption is greater than the computing power reserve, generate a hierarchical dimensionality reduction instruction based on the predefined business rule priority. The dynamic code display module is configured to respond to hierarchical dimensionality reduction instructions. According to the masking target indicated in the instructions, it performs pixel replacement or masking processing on the corresponding non-critical hierarchical information areas in the original carrier image, and performs targeted decoding on the retained critical hierarchical information to generate local access credentials and verification marks. The asynchronous verification module is configured to perform inference reconstruction and consistency compensation based on statistical probability models on the undecoded non-critical level information in the background system, based on the flag to be verified, to complete the final ledger verification.

[0005] As a further aspect of the present invention, the method for obtaining the original carrier image includes: The optical scanning device is controlled to perform multiple exposure acquisitions within a preset time window to obtain multiple initial images with different exposure levels. Pixel-level fusion of multiple initial images is performed to generate a high dynamic range image; Edge detection and texture analysis are performed on high dynamic range images to identify areas of physical wear and areas of light interference. Based on the distribution of physical wear areas and light interference areas, weighted denoising processing is performed on the high dynamic range image to obtain the original carrier image.

[0006] As a further aspect of the present invention: the method for constructing the expected decoding computational power consumption of the current original carrier image includes: A two-dimensional fast Fourier transform is performed on the original carrier image to obtain the image spectrogram; Calculate the proportion of high-frequency components and spectral entropy of the image spectrogram as frequency domain noise features; Obtain the preset baseline computing power consumption model; The frequency domain noise characteristics are input into the benchmark computing power consumption model to map the theoretical number of calculation cycles required in the full-level decoding mode; Obtain the current system's chip temperature and memory fragmentation index; The chip temperature and memory fragmentation index are normalized and weighted summed to obtain the hardware thermal exhaustion coefficient. Multiply the theoretical number of calculation cycles by the hardware heat exhaustion factor to obtain the expected decoding computing power consumption.

[0007] As a further aspect of the present invention: the method for generating hierarchical dimensionality reduction instructions includes: Obtain predefined hierarchical structure information in the optical stacked data carrier, wherein the hierarchical structure information includes multiple data levels and their corresponding service types; Assign priority weights to each data level based on predefined business rule priorities; Calculate the remaining computing power minus the expected decoding computing power consumption. If the value is less than zero, then the absolute value of the value is determined as the computing power gap. Based on the size of the computing power gap, the data levels to be masked are selected sequentially from low priority weight to high priority weight until the expected decoding consumption of the remaining data levels is less than or equal to the computing power margin. The identified data levels to be masked are marked as masking targets, and hierarchical dimensionality reduction instructions containing the masking targets are generated.

[0008] As a further aspect of the present invention, a method for region masking of non-critical hierarchical information in the original carrier image includes: Based on the masking target in the hierarchical dimensionality reduction instruction, the spatial distribution region of the corresponding non-critical layer information in the original carrier image is determined. Construct a mask matrix corresponding to the spatial distribution area; The mask matrix is ​​applied to the original carrier image, and the pixel values ​​of non-critical layer information are set to zero or replaced with preset background noise values ​​to generate a dimension-reduced carrier image. Methods for targeted decoding of key hierarchical information include: The reduced-dimensional carrier image is input into a pre-trained adversarial generative network model, and super-resolution reconstruction and deblurring are performed on the region where the key layer information is located to generate enhanced key layer features. An error correction decoding algorithm is used to analyze the enhanced key layer features and extract core business data as local access credentials.

[0009] As a further aspect of the present invention, a method for probabilistically reconstructing and compensating for undecoded non-critical level information in a background system includes: Receive local access credentials and verification tags, and parse the verification tags to identify the types of non-critical information that have been masked. Retrieve historical transaction records and user historical behavior data associated with local access passes; Based on historical transaction records and user historical behavior characteristics data, the posterior probability of non-critical level information meeting compliance requirements is calculated using Bayesian inference models or Markov chain models. If the posterior probability is greater than the preset compliance threshold, non-critical level information will be automatically completed, a temporary credit certificate will be generated, and the ledger to be verified will be updated to complete the consistency compensation. If the posterior probability is less than or equal to the compliance threshold, an anomaly alert will be generated and the relevant rights and interests will be frozen, pending manual review.

[0010] As a further aspect of the present invention: the method for real-time monitoring of the current system's computing power reserve includes: Real-time acquisition of the current operating frequency, current temperature, and current task queue length of the decoding chip; Based on the current temperature, query the preset temperature-frequency derating curve to determine the maximum allowable operating frequency. Calculate the occupied computing power based on the current task queue length; Convert the current maximum allowed operating frequency into total available computing power, and subtract the occupied computing power from the total available computing power to obtain the computing power reserve.

[0011] As a further aspect of the present invention: the system also includes an evolution analysis module, which is used for: Collect frequency domain noise features extracted at different time periods and the corresponding decoding failure rates; Cluster analysis of frequency domain noise characteristics identifies specific noise patterns that cause the actual decoding consumption to deviate from the expected decoding computing power consumption by more than a preset threshold or lead to an increase in the decoding failure rate. Specific noise patterns are labeled as anti-wear characteristics; Based on the adversarial wear characteristics, the weight parameters of the benchmark computing power consumption model are iteratively updated to improve the accuracy of subsequent decoding computing power consumption prediction.

[0012] Compared with existing technologies, this invention introduces an environmentally aware dynamic evaluation mechanism and a hardware thermal exhaustion coefficient to accurately construct the expected decoding computing power consumption, effectively solving the problem that static weighted indicators in existing technologies cannot cope with nonlinear environmental noise and hardware thermal state interference. Specifically, this invention uses Fast Fourier Transform to extract image frequency domain noise features and combines an enhanced noise gain regression model to establish a mapping relationship between frequency domain features and theoretical calculation cycles. More importantly, the system collects chip temperature and memory fragmentation index in real time to construct a hardware thermal exhaustion coefficient, transforming the theoretical calculation cycle into an actual consumption expectation that conforms to the current physical state. This accurate mapping from theoretical model to physical environment enables the system to perceive computing power risks at the millisecond level, avoiding gate deadlock or response timeouts caused by overestimation of computing power, and significantly improving the robustness of the system in high-throughput scenarios. Attached Figure Description

[0013] The present invention will be further explained below with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart of the stacked QR code display, reading and verification management system based on reinforcement learning, which is part of the present invention. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0015] Please see Figure 1 This invention is a reinforcement learning-based stacked QR code display, reading, and verification management system, comprising: The image acquisition module is configured to perform real-time scanning of the optically stacked data carrier on the passage medium to acquire the original carrier image; The consumption prediction module is configured to perform spectral feature analysis on the original carrier image, extract frequency domain noise features, and combine them with the preset decoding model complexity to construct the expected decoding computing power consumption of the current original carrier image. The strategy scheduling module is configured to monitor the current system's computing power reserve in real time, compare the expected decoding computing power consumption with the computing power reserve, and if the expected decoding computing power consumption is greater than the computing power reserve, generate a hierarchical dimensionality reduction instruction based on the predefined business rule priority. The dynamic code display module is configured to respond to hierarchical dimensionality reduction instructions. According to the masking target indicated in the instructions, it performs pixel replacement or masking processing on the corresponding non-critical hierarchical information areas in the original carrier image, and performs targeted decoding on the retained critical hierarchical information to generate local access credentials and verification marks. The asynchronous verification module is configured to perform inference reconstruction and consistency compensation based on statistical probability models on the undecoded non-critical level information in the background system, based on the flag to be verified, to complete the final ledger verification.

[0016] This embodiment discloses a reinforcement learning-based stacked QR code display, reading, and verification management system. This system aims to solve the traffic congestion singularity problem caused by physical wear and tear of the medium, light interference, and thermal exhaustion of decoding computing power in extremely high-throughput scenarios such as multinational mega-transportation hubs, when dealing with high-density stacked QR codes containing multiple encrypted rights. This system achieves its technical objectives through the collaborative work of the following modules: The system utilizes an image acquisition module to acquire high-quality raw carrier images in complex lighting environments. This module is configured to perform real-time scanning of access media, such as the screen of a smart terminal held by a passenger or the optically stacked data carrier on a paper voucher. Considering the possibility of strong light reflection or shadow obstruction on site, the acquisition process adopts multispectral or high frame rate mode to obtain raw carrier images. Before performing deep decoding, the consumption prediction module uses lightweight computation to predict the current decoding difficulty and computing cost. This module performs spectral feature analysis on the original carrier image, specifically by converting the image from the spatial domain to the frequency domain through fast Fourier transform, and extracting frequency domain noise features, such as the proportion of high-frequency clutter. This module combines the preset decoding model complexity, i.e. the theoretical number of floating-point operations required for full-level decoding, to construct the expected decoding computational power consumption of the current original carrier image; this expected value is not a fixed value, but a dynamic variable that is positively correlated with the noise level of the image. The strategy scheduling module acts as the command center of the system, dynamically balancing computing resources and business compliance. This module monitors the current computing power reserve of the system in real time, that is, the remaining computing power available for hardware at the current moment. The module compares the expected decoding computing power consumption with the computing power reserve. If the expected decoding power consumption exceeds the remaining power, it indicates that forcibly performing full-level decoding is likely to cause gate deadlock or response timeout, and the system triggers the circuit breaker protection mechanism; the module generates hierarchical reduction instructions based on predefined business rule priorities, such as security being greater than ticketing being greater than commercial points. Based on this, the dynamic code display module executes a lossy but compliant downgrade reading strategy; in response to the hierarchical dimensionality reduction instruction, according to the masking target indicated in the instruction, namely the low-priority non-critical data level, the module performs pixel replacement on the corresponding non-critical information area in the original carrier image, such as blacking out or masking, thereby physically blocking decoding attempts on that area and saving computing power. At the same time, the module performs targeted decoding on key hierarchical information such as the core identity layer to generate local access credentials for controlling the gate opening, and verification markers to record the hierarchical identity identifiers (IDs) that are blocked. After physical passage is completed, the asynchronous verification module repairs the consistency of data in the digital space. Based on the verification mark, the module processes the undecoded non-critical level information in the background server system. Since the original data has not been read, the module uses statistical probability models, such as Bayesian networks, combined with user historical profiles to infer and reconstruct the missing data and compensate for consistency, thus completing the final ledger verification. This embodiment achieves an innovative mechanism of edge-side dimensionality reduction and cloud-based probabilistic completion through the collaborative work of the above modules. When faced with the computing power crisis caused by optical adversarial wear, the system can perceive the risk in milliseconds and actively discard the decoding of non-critical information, ensuring the continuity of physical passage and avoiding congestion. At the same time, the asynchronous verification in the background ensures the final consistency of the digital ledger, effectively solving the deadlock problem of high-density stacked codes in extreme scenarios.

[0017] In a preferred embodiment of the present invention, the method for acquiring the original carrier image includes: controlling an optical scanning device to perform multiple exposure acquisition within a preset time window to acquire multiple initial images with different exposure levels; performing pixel-level fusion on the multiple initial images to generate a high dynamic range image; performing edge detection and texture analysis on the high dynamic range image to identify physical wear areas and light interference areas; and performing weighted denoising processing on the high dynamic range image based on the distribution of physical wear areas and light interference areas to acquire the original carrier image.

[0018] This embodiment specifies the method for acquiring the original carrier image, aiming to improve data quality from the source. The system controls the optical scanning device to perform multiple exposure acquisitions within a preset time window. In this embodiment, the time window is set to an extremely short interval, such as within 15ms. During this period, the image acquisition module controls the sensor to continuously acquire multiple frames of initial images with different exposure times, such as short exposure to capture bright details and long exposure to capture shadow details. The system performs pixel-level fusion on multiple initial images; using the HDR high dynamic range fusion algorithm, it weights and synthesizes the clearest pixels from multiple images to generate a high dynamic range image; the specific fusion formula is as follows: for each pixel position... Calculate each frame weight : in, Let be the standard deviation of the Gaussian kernel function, and let its value be... This is an empirical value, determined by: constructing a calibration dataset containing 500 sets of scenes with different light ratios based on the photosensitivity characteristics of the optical sensor selected by the system, with a step size of 1. traversal within the interval The information entropy and histogram uniformity of the fused image are calculated, and the optimal value of 70 is selected as a fixed parameter. The calculation formula is: in, For the present The image information entropy corresponding to the value, For histogram uniformity; In the above During the traversal process, the minimum and maximum values ​​of information entropy and uniformity are calculated; To prevent extremely small constants with a denominator of zero, the value is taken as... ; The weighting is based on the following principle: in low-light fusion scenarios, the richness of information content takes precedence over the smoothness and uniformity of brightness distribution. Indicates the frame number acquired during multiple exposures. This represents the total number of frames captured. Indicates the first Frame image in two-dimensional coordinates The grayscale pixel value at that location ranges from 1 to 10. Final pixel value: in, This represents the total number of frames acquired; this step effectively eliminates local blind spots caused by screen reflections and backlighting; the system performs edge detection and texture analysis on high dynamic range images; using the Sobel operator or Canny edge detection algorithm, combined with texture feature extraction, such as the gray-level co-occurrence matrix, it identifies non-data textures in the image; to ensure the accuracy of recognition and the reproducibility of the code implementation, the system specifically constructs a step size. ,direction The gray-level co-occurrence matrix is ​​calculated and the mean of its four-directional features is taken. To eliminate noise interference and unify the calculation benchmark, the image undergoes gray-level compression processing before constructing the gray-level co-occurrence matrix; the quantization level is set. Using the formula: in, For high dynamic range images in coordinates The grayscale value at that location; The total number of gray levels in an 8-bit grayscale image, i.e. Range; the original grayscale range Linear mapping to Based on the quantized image Construct a matrix, where the matrix dimension is... The system introduces a preset quantization threshold: contrast threshold. Homogeneity threshold The system performs a sliding window scan of the image; if the calculated contrast of a certain area... and homogeneity If the area is identified as a physical wear area, it is characterized by an irregular network of micron-level scratches, i.e., high contrast and low homogeneity, which is different from data areas with periodic texture features, i.e., high contrast but high homogeneity; the second is an area of ​​light interference, characterized by large areas of overexposure spots or moiré patterns. Based on the distribution of physical wear areas and illumination interference areas, weighted denoising processing is performed on high dynamic range images. The system generates a weight map based on the identified interference areas; specifically, the system constructs a noise distribution matrix with the same size as the image. , where subscript The two-dimensional coordinates of image pixels are used to define the coordinates of pixels identified as physically worn. For areas with illumination interference, in order to establish a quantization relationship between the degree of brightness overflow and the noise reduction weight, the system obtains the grayscale brightness values ​​of the pixels in that area. Set a brightness threshold for determining strong light interference. For example, for 230, the weights are calculated using a linear mapping formula: in, The grayscale value of the currently processed pixel is truncated to... Within the specified range; the rest of the range is set to 0; the noise reduction process uses an adaptive weighting formula: in, The aforementioned synthesized high dynamic range image is used as the original input. These are the pixel values ​​after Gaussian filtering. The preset mixing coefficient, The coordinates of the new image generated after denoising processing In this embodiment, based on extensive experimental testing, the grayscale pixel value at that location, the coefficient... The value range is set to value range The boundary determination criteria are as follows: the lower limit of 0.6 corresponds to the peak signal-to-noise ratio (PSNR); the inflection point where the first derivative of the gain curve begins to decay; and the upper limit of 0.85 corresponds to the structural similarity (SSIM) and the critical point where the exponential descent slope suddenly increases. Typical value The specific basis for the extensive experimental tests here is as follows: The system constructed a benchmark dataset containing 5,000 images with different light intensities (200-10,000 Lux) and physical wear levels. Peak signal-to-noise ratio (PSNR) was used as the quantitative indicator of denoising effect, and structural similarity (SSIM) was used as the quantitative indicator of texture detail preservation. Experimental data show that when When the PSNR improvement in the highlight area is less than 3dB below the effective threshold, it cannot effectively suppress spot interference; when At this point, the SSIM index of the image falls below the critical threshold of 0.85, causing high-frequency data texture blurring and a decoding success rate decrease of more than 15%; typical value The equilibrium point with the highest normalized weighted score between PSNR and SSIM was selected; specifically, this normalized weighted score... The calculation formula is: in, and The coefficients used are respectively The peak signal-to-noise ratio and structural similarity values ​​of the processed image relative to the standard reference image; where, , The weighting is based on the following: the system uses the analytic hierarchy process (AHP) to weight the factors affecting the decoding success rate, constructs a judgment matrix, and calculates the consistency ratio. To obtain structural information relative signal-to-noise ratio Importance ratio Construct a judgment matrix , of which elements Indicates the first The first indicator is relative to the first The importance ratio of each indicator: The weight vector is obtained by normalizing the corresponding feature vector. ; and The lower and upper quartiles of the statistical distribution were determined based on 10,000 randomly selected natural scene images from the open-source image dataset V6, after adding Gaussian white noise, and then normalized. ; In stacked QR code recognition, the integrity of structural information is crucial. Its contribution to decoding success rate is higher than that of signal-to-noise ratio alone. , and The lower and upper quartiles of the statistical distribution of OpenImageDatasetV6 are used as boundary constants for normalization. After traversal calculation, when hour The system aims to maximize the value of the signal-to-noise ratio, thereby preserving key texture details while effectively denoising. In this way, the system applies strong filtering parameters, such as Gaussian filtering, to areas with severe interference, and weak filtering or sharpening to areas with clear data texture, ultimately obtaining the original carrier image with the optimal signal-to-noise ratio. This embodiment can significantly reduce the noise level of input data at the source by using multiple exposure fusion and targeted region-weighted denoising; in particular, the identification and processing of screen scratches and light spots allows the subsequent decoding module to focus on effective data textures, avoiding wasting computing power on invalid noise analysis, thereby improving the initial decoding success rate.

[0019] In a preferred embodiment of the present invention, the method for constructing the expected decoding computational power consumption of the current original carrier image includes: performing a two-dimensional fast Fourier transform on the original carrier image to obtain an image spectrum. Calculate the proportion of high-frequency components and spectral entropy of the image spectrogram as frequency domain noise features; obtain a preset benchmark computing power consumption model; The frequency domain noise characteristics are input into the benchmark computing power consumption model to map the theoretical number of calculation cycles required in the full-level decoding mode; the chip temperature and memory fragmentation index of the current system are obtained; The chip temperature and memory fragmentation index are normalized and weighted and summed to obtain the hardware thermal exhaustion coefficient; the theoretical number of calculation cycles is multiplied by the hardware thermal exhaustion coefficient to obtain the expected decoding computing power consumption.

[0020] This embodiment specifies the method for constructing the expected decoding computational power consumption of the current original carrier image and introduces a dynamic evaluation mechanism for environmental perception. The system performs a two-dimensional fast Fourier transform on the original carrier image; converts the spatial domain image into a frequency domain image, and obtains the image spectrum. The proportion of high-frequency components and spectral entropy of the image spectrogram are calculated as frequency domain noise features; to quantify the proportion of high-frequency components, a normalized frequency threshold is defined in the system. For example, if we take 0.6, the distance between the center of the spectrum is greater than... The sum of the energies of all frequency components and the ratio of the total energy are used as the proportion of high-frequency components. To quantify the spectral entropy, the system normalizes the spectral amplitude into a probability distribution matrix. , where subscript Represent the frequency domain coordinates and calculate: As the spectral entropy; a preset baseline computing power consumption model is obtained; to address the issue of missing model complexity parameters, this embodiment introduces decoding model complexity as the input base of the baseline computing power consumption model; the system obtains the preset decoding model complexity. This value is defined as the total number of standard instructions required for full-level decoding in a noise-free ideal environment, and is uniformly measured using MI as the unit of measurement. For example, setting... ; The specific method for obtaining and calibrating this preset value is as follows: During the calibration phase before the system leaves the factory, the standard test pattern ISO / IEC15415 Reference Test Card is selected, and 1000 consecutive decoding tests are performed under conditions without interference from other background tasks. The total number of instruction submissions in the CPU performance counter is recorded and the average value is taken as the value for this specific hardware platform. Solidified parameters; based on this, this embodiment uses an enhanced noise gain regression model to construct a benchmark computing power consumption model, the functional expression of which is modified as follows: This formula explicitly establishes the complexity of the decoding model. Compared with the final theoretical calculation of the number of cycles The mathematical mapping relationship; where, and The normalized dimensionless value; weighting parameter to is a dimensionless noise sensitivity coefficient, characterizing the amplification factor of noise on the basic complexity; This represents a fixed system scheduling overhead, measured in minutes (MI). In this embodiment, the parameters are set to the following values: The above noise sensitivity coefficient to The calibration method is as follows: Based on 10,000 stacked QR code image samples in typical scenarios collected during the system development phase, the frequency domain noise features of each image are extracted as independent variables, and the actual full-level decoding time recorded by a high-precision timer, i.e. converted into the number of standard instructions MI, is used as the dependent variable. The multivariate nonlinear least squares method is used for regression fitting to calculate the coefficient combination that minimizes the sum of squared residuals. 10,000 samples covered and The system ensures full coverage distribution and that the condition number of samples is less than 10 to guarantee that the ill-conditioned nature of the regression equation is within a controllable range. The system inputs frequency domain noise features and the preset decoding model complexity into the model, mapping the theoretical computational load required in the full-level decoding mode. ; Based on this, the current system chip temperature and memory fragmentation index are obtained; the system reads the core temperature of the decoding chip in real time through the underlying hardware driver. and the degree of fragmentation of the memory allocator To accurately assess the actual power consumption under the current hardware conditions, this embodiment introduces a hardware thermal exhaustion factor. The calculation formula is as follows: To prevent normalized values ​​from overflowing due to abnormal sensor data or extreme environments, the system defines a truncation normalization function: ; Then calculate the coefficients: Temperature weighting factor 0.7, the value is based on the following: In the main frequency adjustment strategy of the decoding chip, the impact of temperature-triggered frequency reduction on performance is about 2.3 times that of bus pauses caused by memory fragmentation. Current chip temperature; The lower and upper limits of the safe operating temperature range for the chip are set according to industrial-grade chip standards in this embodiment. ; Memory fragmentation weight factor: 0.3; The memory fragmentation index is explicitly specified in this embodiment. The value is a normalized value, and its range is... The specific calculation method is as follows: in, and The method for obtaining the number of free blocks is as follows: The system reads the memory allocator interface of the operating system kernel, such as reading the / proc / buddyinfo file in the Linux system to obtain the number of free blocks of each order, or directly accesses the free_area member variable under the zone structure in the kernel module and traverses the free block linked list of all orders. The sum of the sizes of all free blocks. The size of the memory block corresponding to the currently existing maximum order, i.e. ;in, This represents the maximum order of the currently non-free linked list in the memory allocator (such as the buddy system). This represents the physical memory page size of the system; if the underlying interface returns a non-normalized value, it represents the total number of fragmented free memory blocks in the current system. Then it needs to be divided by the preset baseline value. Normalization is performed, that is: Among them, the preset benchmark value Defined as the critical fragmentation count when frequent page swapping occurs in the system, it is determined by: performing a memory stress test on the target system and recording the first derivative of the system's page fault rate as greater than 1. The number of free fragments at that time, taken as 1.2 times that value. For example, measured in a 4GB memory system as A fragment; In addition, regarding The amplification effect of the hardware state on the expected consumption may exceed 100%, that is, double. Regarding the reasonableness of this, this embodiment points out that the amplification ratio is based on the measured results of the hardware physical characteristics. Experiments show that when the chip is under high temperature and frequency reduction and memory is highly fragmented, resulting in frequent page table replacements, the number of CPU instruction cycles per instruction will increase non-linearly and exponentially, and the actual processing time is indeed 1.8 to 2.0 times the theoretical value. The original heat depletion level in the physical model may exceed In this embodiment, it is clarified that the formula calculation... As a benchmark reference coefficient for the scheduling system, it is mapped to a normalization function in the algorithm implementation. Within a closed interval, to ensure the convergence of scheduling instructions; while physically exceeding... Performance degradation is achieved by setting Forced circuit breaker mechanism handling; In this embodiment, the hardware thermal depletion coefficient Restricted by normalization The interval is used to characterize the proportion of additional computing resources consumed by the current hardware state; the performance degradation in the physical model is linearly mapped to this interval to eliminate the risk of computational overflow; thus, the expected computing power consumption for decoding is calculated. : This embodiment introduces an explicit decoding model complexity variable. Including the hardware thermal depletion coefficient, a complete closed loop for predicting computing power consumption was constructed, ensuring an accurate mapping from the theoretical model to the physical environment.

[0021] In a preferred embodiment of the present invention, the method for generating hierarchical dimensionality reduction instructions includes: obtaining predefined hierarchical structure information in an optical stacked data carrier, the hierarchical structure information including multiple data levels and their corresponding business types; and assigning priority weights to each data level according to predefined business rule priorities. Calculate the remaining computing power minus the expected decoding computing power consumption. If the value is less than zero, the absolute value of the value is determined as the computing power gap. Based on the size of the computing power gap, select the data levels to be blocked sequentially from low priority weight to high priority weight until the expected decoding consumption of the remaining data levels is less than or equal to the computing power gap. Mark the determined data levels to be blocked as blocking targets and generate hierarchical dimensionality reduction instructions containing the blocking targets.

[0022] This embodiment specifies the method for generating hierarchical dimensionality reduction instructions, employing a dynamic programming strategy based on a greedy algorithm; the system acquires predefined hierarchical structure information in the optical stacked data carrier; the system reads the configuration table and identifies multiple data levels contained in the stacking code; Based on predefined business rule priorities, priority weights are assigned to each data level; the system maintains a business priority mapping table for each data level. Assign a specific numerical priority weight For example, integers from 1 to 10, the smaller the value, the lower the priority and the more easily it is sacrificed; for example, in the allocation of business points. Electronic ticketing layer allocation Basic identity layer allocation ; Calculate the computing power gap; The system calculates the difference between the remaining computing power and the expected decoding computing power consumption; In response to the difference being less than zero, its absolute value is determined as the computing power gap. ; Based on this, the data layers to be shielded are selected according to the size of the computing power gap; the system quantifies the expected decoding consumption of each data layer in the current environment. Considering the need to maintain the speed of the greedy algorithm's decision-making process without losing awareness of texture complexity during the layered scanning phase, a fast estimation model that preserves the main linear features is adopted; among which, the single-layer baseline complexity is... The method for determining is as follows: based on predefined hierarchical structure information, obtain the first... Data codeword capacity at each data level Total stacking capacity The proportion is calculated as follows: in, The typical value for the overhead coefficient of independent layer decoding is 1.1. This coefficient of 1.1 is chosen based on the following: Testing shows that when decoding a single layer independently, the fixed basic overhead, such as data structure initialization, context switching, and memory page allocation, accounts for approximately 10% of the pure computation time for that layer on average. Therefore, 1.1 is set as the redundancy coefficient. The specific formula for the estimated decoding cost per layer is revised as follows: in, This represents the baseline complexity for this level. The total number of data layers contained in the stacked QR codes, used as a normalization factor in the calculation. and The calibration parameters determined when building the baseline computing power consumption model in the previous section are ignored here because they are computationally expensive higher-order interaction terms. To achieve microsecond-level scheduling; here, to ensure the decision-making speed of the greedy algorithm, a first-order simplified model is adopted; this embodiment uses a model based on The greedy masking algorithm for sorting, with the following specific steps: Sorting: The system constructs a collection containing all data levels. And according to the assigned priority weight Sort the levels in the set in ascending order to generate an ordered list of levels. , making ; Traversal and accumulation: Initialization saves computing power and the set of shielded targets The system iterates through the ordered list sequentially. The hierarchy within; Decision: For the currently traversed level The expected decoding cost Accumulate to and will join in ;like If the computing power gap has been filled, the traversal will terminate immediately; otherwise, the next level will continue to be processed. Through the above logic, the system strictly follows the decision path from low-priority weights to high-priority weights, ensuring the priority weight variables... Its key control role in the core algorithm; generating hierarchical dimensionality reduction instructions containing the shielded targets; and determining the set of shielded targets. It is encapsulated as an instruction and sent to the dynamic code display module.

[0023] In a preferred embodiment of the present invention, the method for performing region masking processing on non-critical hierarchical information in the original carrier image includes: determining the spatial distribution region of the corresponding non-critical hierarchical information in the original carrier image based on the masking target in the hierarchical dimensionality reduction instruction; constructing a mask matrix corresponding to the spatial distribution region; applying the mask matrix to the original carrier image, setting the pixel values ​​of the non-critical hierarchical information to zero or replacing them with preset background noise values, and generating a dimensionality-reduced carrier image. Methods for targeted decoding of retained key hierarchical information include: inputting the dimensionality-reduced carrier image into a pre-trained adversarial generative network model, performing super-resolution reconstruction and deblurring on the region where the key hierarchical information is located to generate enhanced key layer features; and using an error correction decoding algorithm to analyze the enhanced key layer features, extract core business data, and use it as a local access pass.

[0024] This embodiment specifies the methods for semantic masking and targeted decoding, and adopts generative AI enhancement technology; performs semantic masking on non-critical hierarchical information in the original carrier image; based on the masking target in the hierarchical dimensionality reduction instruction, the system uses a macroblock index mapping mechanism to determine the spatial distribution area; constructs a mask matrix corresponding to the spatial distribution area; and generates a binary matrix with the same size as the original image, where the value corresponding to the masked area is 0 and the value of the remaining area is 1. Generate a dimension-reduced carrier image; apply the mask matrix to the original carrier image; replace the pixel values ​​of non-critical regions with preset background noise values, such as Gaussian noise with a mean of 128, to maintain stable image statistical characteristics; Based on this, the key hierarchical information is decoded in a targeted manner; the dimensionality-reduced carrier image is input into a pre-trained adversarial generative network model; the model uses the SRResNet architecture as the generator and the discriminator adopts a classification structure based on the VGG network. Specifically, the generator network contains 16 residual blocks, each consisting of two... The model consists of convolutional layers and batch normalization layers; the discriminator comprises 8 convolutional layers and 2 fully connected layers; the model is trained using the Adam optimizer, with an initial learning rate of [value missing]. and in The number of iterations is halved after the second iteration, for a total of 100 iterations. The batch size is set to 16. To meet the requirement of full disclosure, this embodiment discloses in detail the hybrid loss function used in the model training process. The specific mathematical expression is: in, For example, the balance coefficient. Perceived loss Euclidean distance is calculated based on feature maps extracted from a pre-trained VGG-19 network to constrain content consistency. in, This represents the feature map output by the ReLU5_4 layer of the VGG-19 network, which is the fourth convolutional layer before the fifth max pooling layer. For high-resolution ground truth images, The enhanced image output by the generator, where, Specifically refers to the dimension-reduced carrier image generated after processing by the dynamic code-unwrapping module. This corresponds to the original, uncompressed, high-resolution image; Specifically refers to the first The first pooling layer The width and height of the feature map output by each convolutional layer are in pixels; Represents the pixel spatial coordinates in the feature map, with values ​​ranging from 1 to 2. and ; combating losses The standard generator loss from generative adversarial networks is used to recover high-frequency texture details. in, The normalized probability output by the discriminator represents the probability that the input image is judged as a real high-quality image. The model is iteratively trained by minimizing the above-mentioned hybrid loss function, and performs super-resolution reconstruction and deblurring on the regions where key hierarchical information is located to generate enhanced key layer features. Core business data is extracted. Error correction decoding algorithm is used to analyze the enhanced features and extract local access credentials. The local access credential uses a lightweight type-length-value format for encoding. The specific structure includes: Tag = 0x01 User Core ID, Tag = 0x02 Validity Token, and Tag = 0x03 Gate Control Instruction. The verification tag is stored in JSON format, with the specific fields defined as {Shielded Layer: [LayerID_List], Context Hash Value: HashValue, Timestamp: UnixTime}, which are used for subsequent asynchronous indexing.

[0025] In a preferred embodiment of the present invention, the method for probabilistic reconstruction and consistency compensation of undecoded non-critical level information in the background system includes: receiving a local access pass and a verification mark, parsing the verification mark to identify the type of non-critical level information that has been blocked; and retrieving historical transaction records and user historical behavior feature data associated with the local access pass. Based on historical transaction records and user historical behavior data, the posterior probability of non-critical level information meeting compliance requirements is calculated using a Bayesian inference model or a Markov chain model. If the posterior probability is greater than the preset compliance threshold, the non-critical level information is automatically supplemented, a temporary credit certificate is generated, and the ledger to be verified is updated to complete consistency compensation. If the posterior probability is less than or equal to the compliance threshold, an anomaly alarm is generated and the relevant rights and interests are frozen, pending manual review.

[0026] This embodiment specifies the probabilistic reconstruction and consistency compensation method of the backend system and constructs a trust inference mechanism based on big data. The backend system receives the local access pass and the verification mark uploaded by the frontend, parses the verification mark to identify which levels, such as the commercial points level, are blocked on the frontend; retrieves related data; the system retrieves historical transaction records and user historical behavior feature data associated with the local access pass. The posterior probability is calculated using a Bayesian inference model. This embodiment constructs a probabilistic model to calculate the probability that, given a user's core identity and historical behavior, the missing non-critical information meets compliance requirements. The calculation formula is as follows: in, Assume the event, i.e., the user's non-critical level information, is compliant; Evidential events refer to the context in which the current transaction takes place, such as the time, place, and validity of the associated local access pass. Background information refers to user historical behavioral characteristics data, such as the fulfillment rate of a certain number of past transactions and credit score; The posterior probability to be calculated; To ensure the computability of the above formula in practical applications, this embodiment further discloses the prior probability. Likelihood probability Specific calculation method: Regarding prior probability The system starts from background information. Extract the total number of historical transactions of users Compared with historical violations ;like For users who are cold-starting, a preset initial trust value is assigned. Otherwise, calculate the basic performance rate. Simultaneously, through the system's reserved third-party credit scoring API interfaces, such as Sesame Credit or the weighted evaluation interfaces of various internal business subsystems, the user's current credit score S is obtained, and mapped to a maximum-minimum normalization method. Interval The calculation formula is: in, The original credit score obtained from the query. and To define the theoretical lower and upper limits of this credit system, in this embodiment, taking the integration with Sesame Credit as an example, we set... If integrated with other credit scoring systems or internal scoring systems, these two parameters are dynamically configured based on the actual score range of that system; the prior probability calculation formula is then: Constructing feature vectors : in, The timestamp of the transaction. and These are the longitude and latitude coordinates of the transaction location, respectively. and These are the mean and standard deviation of the corresponding feature in the compliant historical sample set, respectively; the system is based on a set of N transaction samples marked as compliant within the most recent 6 months of the user's historical records, where, If the number is less than 30, then the average characteristic data of the same user group should be added. Construct a multidimensional Gaussian distribution model; calculate the sample mean vector. With the sample covariance matrix: The current feature vector Input the model and calculate the likelihood probability using the multidimensional Gaussian probability density function: in, For feature dimension, Let be the determinant of the covariance matrix. To improve the robustness of the algorithm under sparse samples and prevent the covariance matrix from becoming non-invertible (i.e., a singular matrix) due to insufficient samples or eigenvalue collinearity, thus causing computational interruption, the system applies Tikhonov regularization, i.e., constructs... and use Replace the original covariance matrix Substituting into the above Gaussian density function formula to perform determinant calculation and inverse matrix The calculation; where, This is a preset small perturbation constant, selected based on characteristic data. The effective precision lower limit, i.e., 4 decimal places, corresponds to the smallest unit of resolution. To ensure the regularization term The least significant bit that eliminates singularities without interfering with valid data is selected as a value one order of magnitude lower than the precision lower limit. ; Choosing this value ensures that numerical instability is eliminated during matrix inversion without introducing noise exceeding the sensor's measurement error. It is the identity matrix; this probability value represents the likelihood of the current spatiotemporal context occurring under the compliance assumption; for the denominator The law of total probability is used to calculate: in, For hypothetical events Evidence events under conditions that are not valid The probability of occurrence is calculated based on the probability density of historical violation sample distributions; to achieve code-level closure, the system maintains a set of violation samples. Using the same feature extraction method as the compliant samples, a multidimensional Gaussian distribution model is constructed to calculate the mean of the non-compliant samples. With covariance matrix Substituting into the Gaussian density function, we can calculate... ; If the initial sample of violations in the system is insufficient, for example Then Set to the preset uniform distribution constant The constant is determined based on a one-in-a-million probability benchmark defined by the system, which is used as a lower bound for the non-zero probability when samples are missing, so as to avoid the denominator being zero. Perform consistency compensation or anomaly handling; in response to the calculated posterior probability being greater than a preset compliance threshold, such as 0.99, the system determines that the risk of this dimensionality reduction pass is controllable and triggers an automatic completion process for non-critical level information; this completion process is not arbitrary, but executes the following strict data reconstruction logic: the system uses the user's unique identifier ID in the local pass credential as an index to attempt to retrieve a complete ledger snapshot of the previous period from the cold data storage area. ; In this step, the system performs a null value check: if no null value is found... Or the number of users' historical transactions is insufficient to calculate the average growth rate, i.e. The system will then call the preset new user benefits initialization template and... Set the base equity value and skip extrapolation calculations; If historical data is complete, identify the type of the masked layer, such as the integration layer, from... Extract the corresponding field value from the system to obtain the user's most recent... The set of values ​​corresponding to the equity tier in the next historical transaction record. and its corresponding timestamp The first derivative of the numerical value over time is calculated using the least squares method as the growth rate. The formula is: in, The arithmetic mean of the timestamps. It is the arithmetic mean of the equity value; combined with the user's historical average growth rate. Perform linear extrapolation to calculate the reconstructed value: in, The Unix timestamp of the current QR code scanning request. The Unix timestamp used when the previous ledger snapshot was generated. This refers to the value of the corresponding equity level extracted from the previous period's ledger snapshot; Fill in the missing fields of the current transaction record, generate a temporary credit certificate and update the ledger to be verified to complete the consistency compensation; in response to the posterior probability being less than or equal to the compliance threshold, the system considers there to be a risk, generates an abnormal alarm, temporarily freezes the relevant rights and interests, and pushes the record to the manual review queue. This embodiment is key to achieving a closed loop of "pass first, then verify". By introducing Bayesian inference and clarifying the specific probability calculation logic and cold start exception handling, the system transforms deterministic data reading into probabilistic risk assessment. This allows non-core verification tasks to be safely offloaded to the backend when front-end physical computing power is limited, and the predictive capabilities of big data to fill the information gaps that the front-end has not read, thereby greatly improving passage efficiency without sacrificing security.

[0027] In a preferred embodiment of the present invention, the method for real-time monitoring of the current system's computing power reserve includes: real-time acquisition of the current operating frequency, current temperature, and current task queue length of the decoding chip; based on the current temperature, querying a preset temperature-frequency derating curve to determine the current maximum allowed operating frequency; calculating the occupied computing power according to the current task queue length; converting the current maximum allowed operating frequency into total available computing power, and subtracting the occupied computing power from the total available computing power to obtain the computing power reserve.

[0028] This embodiment specifies the method for real-time monitoring of the current system's computing power margin, aiming to obtain the true hardware performance boundaries; the system collects the decoding chip status in real time; the system periodically, for example, reads the current operating frequency of the decoding chip every 5ms. Current temperature and task queue length ; The system determines the maximum allowed operating frequency; based on the current temperature, it queries a preset temperature-frequency derating curve; in this embodiment, this curve is hard-coded as a piecewise continuous function. : in, To ensure consistency of units, MHz is used here, which is the nominal base frequency of the chip. The temperature control threshold at which frequency reduction begins. The preset thermal decay coefficient is used to simulate the nonlinear decrease in heat dissipation efficiency. To force the protection of temperature, For the lowest sustaining frequency, the above parameters All calibrations are based on the thermal characteristic curves in the RISC-V system-on-chip hardware specifications used in this system. Ten key frequency-temperature discrete points in the [65°C, 95°C] range were extracted from the chip datasheet and fitted using second-order polynomial regression. ,in Represents operating frequency. Represents chip temperature. The constant coefficients are used to fit the second-order polynomial regression; the coefficients of the quadratic term are extracted. The absolute value is The system will collect data in real time. Input the above function to calculate the precise current maximum allowed operating frequency; Calculate the computing power margin; to ensure that the computing power margin and the expected decoding consumption are consistent in terms of physical dimensions, this embodiment introduces a system-preset maximum allowable response time limit for a single passage. For example, 0.2 seconds; computing power reserve The calculation formula is revised as follows: in, The maximum permissible operating frequency at the current temperature is derived from the piecewise function calculation above, and the unit is MHz; The conversion factor from frequency to computing power is derived from chip performance benchmark tests and is measured in millions of instructions per second (MIPS) / MHz. In the RISC-V architecture processor used in this embodiment, this factor is measured as follows: The determination method is as follows: Run the Dhrystone benchmark program 2.1 standard test program, measure the DMIPS score, and divide it by the operating frequency. System response time limit: converts the computing power rate (MIPS) into the total number of available instructions within the time limit; : Current task queue length, sourced from the system task scheduler, unit: tasks; The estimated average computing power consumption per task is derived from historical statistical averages and is expressed in millions of instructions (MIs). This value is based on operational data from the past week and is set as follows: ; The CPU instruction counter differences for all decoding tasks over the past 7 days were collected, and the arithmetic mean was calculated after removing the highest and lowest 5% outliers. The test was conducted in a multi-task scheduling environment with a single-core clock speed locked and the operating system disabled. The 5% outlier was removed based on Chebyshev's inequality under an unknown distribution. Extreme outliers are detected to eliminate measurement noise caused by system interruptions; the calculated computing power margin is output by the system and used by the strategy scheduling module. This embodiment no longer relies on the nominal performance of the chip, but dynamically evaluates the actual available computing power based on the real-time thermal state and queue backlog. By deducting the occupancy of queued tasks and considering the frequency capping caused by temperature, the calculation result can truly reflect the system's ability to accept new tasks within the specified response time limit, preventing system overload caused by overestimating the computing power.

[0029] In a preferred embodiment of the present invention, the system further includes an evolution analysis module, which is used to: collect frequency domain noise features extracted in different time periods and the corresponding decoding failure rates; perform cluster analysis on the frequency domain noise features to identify specific noise patterns that cause the actual decoding consumption to deviate from the expected decoding computing power consumption by more than a preset threshold or cause the decoding failure rate to increase; mark the specific noise patterns as adversarial wear features; and iteratively update the weight parameters of the benchmark computing power consumption model based on the adversarial wear features to improve the accuracy of subsequent decoding computing power consumption expectations.

[0030] This embodiment specifically defines the function of the evolution analysis module, giving the system the ability to adaptively evolve; the system continuously collects frequency domain noise features and decoding failure rate; the evolution analysis module records the frequency domain noise feature vector of each decoding attempt and the final decoding result in the background, including the success or failure status; The module identifies specific noise patterns that cause the actual decoding cost to deviate from the expected decoding computing power cost by more than a preset threshold; the module performs cluster analysis on the collected noise feature dataset, such as using the K-Means algorithm; the system focuses on clusters with decoding failure rates significantly higher than the average level. In response to the detection of a certain type of noise feature, such as stripe noise at a specific frequency, which causes the actual decoding consumption to be much greater than the expected decoding computing power consumption, the system marks the pattern as an anti-wear feature; this usually corresponds to a new type of screen protector scratch or LED interference at a specific frequency. The system iteratively updates the baseline computing power consumption model; based on the identified adversarial wear characteristics, the system updates the weight parameters of the baseline computing power consumption model online; in specific implementation, to address the mismatch between the feature vector dimension and the model parameter dimension, the system adopts a primary term-first correction strategy; the system targets the cluster center feature vectors corresponding to the identified specific noise patterns. Calculate the corresponding regression model weight adjustment: in, The preset correction step size, for example, 500, is expressed in units of microseconds (MI). This value is determined based on the ratio of the feature vector magnitude to the target correction amount: the feature vector is calculated by statistically analyzing 50,000 historically accumulated noise feature data points from this system. The average L2 norm is approximately 0.8. To ensure that the expected correction to computing power in a single iteration reaches a significant level, approximately 400 MI is set as follows: in, The unit is Ensure with The units are consistent; the four standard deviations are set based on the following: system logs show that the fluctuations in computing power caused by sudden anti-wear and tear usually follow a normal distribution, with a 99.9% confidence interval approximately equal to the mean. The range is defined, therefore the correction amount must cover this range to ensure response speed; This ensures that the correction amount is dimensionally consistent with the weighting parameters. Maintain consistency while possessing sufficient adjustment capabilities; This represents the current decoding failure rate in this mode. The baseline failure rate threshold allowed by the system is set to 0.05 in this embodiment; in, The calculation method is as follows: the system maintains a list containing the most recent... The sliding window of decoding attempts under this specific noise mode is used to count the number of failures. ,but ; The system only applies weights to linear terms. correspond and correspond Perform targeted updates; considering It has dimensions MI, while Since these are dimensionless coefficients, normalization is necessary; the updated formula is corrected as follows: in, For feature vectors of The square of the norm, and For the updated linear term weights, and The weights of the current linear terms before the update are introduced in the formula. As the denominator, it ensures that the incremental term is reduced to a dimensionless value in terms of physical dimensions, thus relating to the weighting coefficients. Maintain consistency; weight for interaction items and higher-order items. In this embodiment, this remains unchanged to prevent model oscillations caused by overfitting of higher-order terms; At the same time, a weight saturation threshold is introduced. If the absolute value of the updated weights exceeds Then cut it into or ,in, This is a sign function used to maintain the positive and negative directions of the updated weights and prevent parameter explosion; in this embodiment, the weight saturation threshold is used. Set to 200% of the initial weight value, that is ; The threshold was selected based on the model's stability boundary test. The system constructed a mixed test set including Gaussian noise, salt-and-pepper noise, and motion blur, and conducted 100 rounds of generalization tests under different weight drift upper limit constraints. Experimental data showed that when the drift of a single weight parameter exceeds 100% of its initial value, the model's generalization error in non-specific noise scenarios will increase exponentially. Therefore, 200% was set as a safe cutoff point to prevent overfitting. In this way, the system will output a higher expected computational power consumption when it encounters similar features again, avoiding scheduling errors caused by model aging. This embodiment endows the system with the ability to self-evolve; as the device is used for longer periods and the environment changes, new interference patterns, namely adversarial wear, will continue to emerge; through evolution analysis, the system can automatically learn the impact of these new patterns on computing power, continuously correct the prediction model, and ensure that the accuracy of the expected computing power consumption is always maintained at a high level.

[0031] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A stacked two-dimensional code display, reading and checking management system based on reinforcement learning, characterized by, include: The image acquisition module is configured to perform real-time scanning of the optically stacked data carrier on the passage medium to acquire the original carrier image; The consumption prediction module is configured to perform spectral feature analysis on the original carrier image, extract frequency domain noise features, and combine them with the preset decoding model complexity to construct the expected decoding computing power consumption of the current original carrier image. The strategy scheduling module is configured to monitor the current system's computing power reserve in real time, compare the expected decoding computing power consumption with the computing power reserve, and if the expected decoding computing power consumption is greater than the computing power reserve, generate a hierarchical dimensionality reduction instruction based on the predefined business rule priority. The dynamic code display module is configured to respond to hierarchical dimensionality reduction instructions. According to the masking target indicated in the instructions, it performs pixel replacement or masking processing on the corresponding non-critical hierarchical information areas in the original carrier image, and performs targeted decoding on the retained critical hierarchical information to generate local access credentials and verification marks. The asynchronous verification module is configured to perform inference reconstruction and consistency compensation based on statistical probability models on the undecoded non-critical level information in the background system, based on the flag to be verified, to complete the final ledger verification.

2. The reinforcement learning based stacked two-dimensional code unrolling, reading and checking management system according to claim 1, characterized in that, The image acquisition module is configured to perform the following operations to obtain the raw carrier image: The optical scanning device is controlled to perform multiple exposure acquisitions within a preset time window to obtain multiple initial images with different exposure levels. Pixel-level fusion of multiple initial images is performed to generate a high dynamic range image; Edge detection and texture analysis are performed on high dynamic range images to identify areas of physical wear and areas of light interference. Based on the distribution of physical wear areas and light interference areas, weighted denoising processing is performed on the high dynamic range image to obtain the original carrier image.

3. The stacked QR code display, reading, and verification management system based on reinforcement learning according to claim 2, characterized in that, Methods for constructing the expected decoding computational cost of the current original carrier image include: A two-dimensional fast Fourier transform is performed on the original carrier image to obtain the image spectrogram; Calculate the proportion of high-frequency components and spectral entropy of the image spectrogram as frequency domain noise features; Obtain the preset baseline computing power consumption model; The frequency domain noise characteristics are input into the benchmark computing power consumption model to map the theoretical number of calculation cycles required in the full-level decoding mode; Obtain the current system's chip temperature and memory fragmentation index; The chip temperature and memory fragmentation index are normalized and weighted summed to obtain the hardware thermal exhaustion coefficient. Multiply the theoretical number of calculation cycles by the hardware heat exhaustion factor to obtain the expected decoding computing power consumption.

4. The reinforcement learning-based stacked QR code display, reading, and verification management system according to claim 3, characterized in that, Methods for generating hierarchical dimensionality reduction instructions include: Obtain predefined hierarchical structure information in the optical stacked data carrier, wherein the hierarchical structure information includes multiple data levels and their corresponding service types; Assign priority weights to each data level based on predefined business rule priorities; Calculate the remaining computing power minus the expected decoding computing power consumption. If the value is less than zero, then the absolute value of the value is determined as the computing power gap. Based on the size of the computing power gap, the data levels to be masked are selected sequentially from low priority weight to high priority weight until the expected decoding consumption of the remaining data levels is less than or equal to the computing power margin. The identified data levels to be masked are marked as masking targets, and hierarchical dimensionality reduction instructions containing the masking targets are generated.

5. The reinforcement learning-based stacked QR code display, reading, and verification management system according to claim 4, characterized in that, Methods for region masking of non-critical hierarchical information in the original carrier image include: Based on the masking target in the hierarchical dimensionality reduction instruction, the spatial distribution region of the corresponding non-critical layer information in the original carrier image is determined. Construct a mask matrix corresponding to the spatial distribution area; The mask matrix is ​​applied to the original carrier image, and the pixel values ​​of non-critical layer information are set to zero or replaced with preset background noise values ​​to generate a dimension-reduced carrier image. Methods for targeted decoding of key hierarchical information include: The reduced-dimensional carrier image is input into a pre-trained adversarial generative network model, and super-resolution reconstruction and deblurring are performed on the region where the key layer information is located to generate enhanced key layer features. An error correction decoding algorithm is used to analyze the enhanced key layer features and extract core business data as local access credentials.

6. The reinforcement learning-based stacked QR code display, reading, and verification management system according to claim 5, characterized in that, Methods for probabilistic reconstruction and consistency compensation of undecoded non-critical level information in the backend system include: Receive local access credentials and verification tags, and parse the verification tags to identify the types of non-critical information that have been masked. Retrieve historical transaction records and user historical behavior data associated with local access passes; Based on historical transaction records and user historical behavior characteristics data, the posterior probability of non-critical level information meeting compliance requirements is calculated using Bayesian inference models or Markov chain models. If the posterior probability is greater than the preset compliance threshold, non-critical level information will be automatically completed, a temporary credit certificate will be generated, and the ledger to be verified will be updated to complete the consistency compensation. If the posterior probability is less than or equal to the compliance threshold, an anomaly alert will be generated and the relevant rights and interests will be frozen, pending manual review.

7. The stacked QR code display, reading, and verification management system based on reinforcement learning according to claim 3, characterized in that, The strategy scheduling module is configured to perform the following operations to monitor the current system's computing power reserve in real time: Real-time acquisition of the current operating frequency, current temperature, and current task queue length of the decoding chip; Based on the current temperature, query the preset temperature-frequency derating curve to determine the maximum allowable operating frequency. Calculate the occupied computing power based on the current task queue length; Convert the current maximum allowed operating frequency into total available computing power, and subtract the occupied computing power from the total available computing power to obtain the computing power reserve.

8. The stacked QR code display, reading, and verification management system based on reinforcement learning according to claim 3, characterized in that, The system also includes an evolution analysis module, which is used for: Collect frequency domain noise features extracted at different time periods and the corresponding decoding failure rates; Cluster analysis of frequency domain noise characteristics identifies specific noise patterns that cause the actual decoding consumption to deviate from the expected decoding computing power consumption by more than a preset threshold or lead to an increase in the decoding failure rate. Specific noise patterns are labeled as anti-wear characteristics; Based on the adversarial wear characteristics, the weight parameters of the benchmark computing power consumption model are iteratively updated to improve the accuracy of subsequent decoding computing power consumption prediction.

Images