A wide dynamic light ray self-adaptive license plate recognition method and system

Abstract

The present application relates to the field of traffic scene license plate recognition, and discloses a wide dynamic light adaptive license plate recognition method and system, which comprises the following steps: collecting a wide dynamic scene image, generating a license plate feature enhancement graph through dynamic contrast stretching and local brightness compensation, and encoding the graph into a half-edge structure graph; running a light scene adaptive switching algorithm based on the graph, constructing a five-state hidden Markov chain to calculate state posterior probability, and outputting a scene confidence vector; activating a scene-specific AI model group to generate multi-model recognition results and construct a recognition partial order set; under the constraint of the partial order set, extracting and physically checking through the upper limit to complete license plate binding and unmanned order generation; the present application solves the problems of unclear character extraction, low recognition rate and the need for manual correction in the wide dynamic light scene, improves accuracy, reduces manual cost, and ensures smooth unmanned business.

Description

Technical Field

[0001] This invention relates to the field of license plate recognition technology in traffic scenarios, and in particular to a wide dynamic range light-adaptive license plate recognition method and system. Background Technology

[0002] In traffic scenarios with uneven lighting conditions, such as direct sunlight, backlighting and shadows, low light at night in rain, and neon interference, traditional license plate recognition technology typically uses a wide dynamic range camera to capture images and then employs global contrast stretching, fixed parameter filtering, and a unified character recognition algorithm to extract license plate information for subsequent business processing in scenarios such as parking. Its technical characteristics are that it relies on global and fixed image processing and recognition strategies, without targeted optimization for lighting differences and environmental interference in different areas, and can only achieve a certain level of accuracy in license plate recognition under stable lighting conditions.

[0003] However, traditional license plate recognition technology suffers from low accuracy in wide dynamic range lighting conditions due to issues such as overexposure, underexposure, or environmental interference in localized areas of the license plate, leading to unclear character features. These problems arise because traditional technologies struggle to account for varying lighting conditions across different regions when processing images holistically, failing to specifically optimize image quality in overexposed or underexposed areas. Fixed filtering parameters also exhibit poor adaptability to different types of interference, such as salt-and-pepper noise and neon flare, resulting in limited interference suppression. Furthermore, single character recognition strategies do not consider the impact of strong light, rainy nights, or neon lighting on character features, making it impossible to capture clear character outlines. These issues directly lead to misjudgments and missed recognitions in license plate character recognition, causing delays or loss of parking order generation. Manual verification and correction of misjudged license plate information is also necessary, increasing operational costs and hindering the advancement of unmanned front-end management. Solving this problem would allow for clear display of license plate character features in wide dynamic range lighting conditions, improving recognition accuracy, preventing business problems caused by recognition errors, reducing manual intervention costs, and ensuring the stable operation of unmanned business processes in wide dynamic range scenarios. Summary of the Invention

[0004] To achieve accurate license plate recognition and unmanned order generation in wide dynamic range lighting conditions, this invention provides the following technical solution:

[0005] A wide dynamic range light-adaptive license plate recognition method includes:

[0006] Step S10: Acquire wide dynamic range scene images, generate license plate feature enhancement maps through dynamic contrast stretching and local brightness compensation, and encode them into half-side structure maps;

[0007] Step S20: Based on the half-side structure diagram, run the light scene adaptive switching algorithm to construct a five-state hidden Markov chain and calculate the posterior probability of each state under the current observation through the forward and backward algorithm, and output the scene confidence vector after integration;

[0008] Step S30: Based on the scene confidence vector, activate the scene-specific AI model group, generate multi-model recognition results, and construct a partial order set of recognition results;

[0009] Step S40: Under the constraints of the constructed partial order set, the final license plate binding and unmanned order generation are performed through supremacy extraction and physical verification.

[0010] Furthermore, the dynamic contrast stretching includes:

[0011] Divide the area of ​​interest of the license plate into sub-blocks evenly;

[0012] Calculate the grayscale histogram of each sub-block and count the range of grayscale values ​​of pixels within the sub-block;

[0013] Perform contrast stretching on each sub-block based on the grayscale value range;

[0014] By splicing all the stretched sub-blocks back to their original positions, the region of interest for the license plate with contrast-stretched sub-blocks is obtained.

[0015] Furthermore, the local brightness compensation includes:

[0016] Calculate the average brightness of each sub-block in the region of interest of the license plate after contrast stretching;

[0017] Set a standard brightness threshold;

[0018] Brightness compensation is performed on each sub-block according to a standard brightness threshold. The brightness compensation includes: if the average brightness of the sub-block is greater than the sum of the standard brightness threshold and the brightness deviation threshold, the sub-block is determined to be an overexposed sub-block and brightness attenuation is performed; if the average brightness of the sub-block is less than the difference between the standard brightness threshold and the brightness deviation threshold, the sub-block is determined to be an underexposed sub-block and brightness enhancement is performed.

[0019] All the sub-blocks after brightness compensation are stitched together in their original positions to obtain the license plate feature enhancement map.

[0020] Furthermore, the half-edge structure graph is a graph structure used to describe the spatial relationship of pixel blocks, illumination attributes, and character topology in the license plate feature enhancement graph, including a vertex set, edge set, half-edge set, local brightness compensation coefficient field, and character connectivity label; the local brightness compensation coefficient field is a set of coefficients defined in the vertex set, and the coefficient corresponding to each vertex is the compensation coefficient used by the pixel block to which the vertex belongs in local brightness compensation; the character connectivity label is a binary label defined in the half-edge set, used to identify whether the half-edge belongs to the edge path of the license plate character.

[0021] Furthermore, the adaptive switching algorithm for lighting scenes includes:

[0022] Three types of feature parameters are extracted from the half-side structure graph and encoded into observation vectors; the three types of feature parameters include vertex brightness distribution histogram, half-side gradient direction consistency, and texture entropy temporal fluctuation.

[0023] Construct a five-state hidden Markov chain based on observation vectors;

[0024] The forward-backward algorithm is used to calculate the posterior probability of each state under the current observation;

[0025] The integrated output scene confidence vector is then used.

[0026] Furthermore, the five-state hidden Markov chain includes a state set, a transition matrix, an observation probability, and an initial probability. The state set contains five lighting scene states: bright sunlight, low light in rainy night, neon interference, mixed backlight, and normal.

[0027] Furthermore, the scene confidence vector is obtained by arranging the posterior probabilities of the five states under the current observation in the order of the state set, and each element of the scene confidence vector corresponds to the posterior probability of a light scene state.

[0028] Furthermore, the scene-specific AI model group comprises five sub-model groups, each corresponding to one of five lighting scenarios. The five sub-model groups include: a dedicated AI model group for sunny, bright light scenarios; a dedicated AI model group for rainy, low-light scenarios; a dedicated AI model group for neon interference scenarios; a dedicated AI model group for mixed backlight scenarios; and a dedicated AI model group for normal scenarios.

[0029] Furthermore, the identification partial order set consists of a candidate license plate string set and a partial order relation, expressed as identification partial order set = (candidate license plate string set, partial order relation), represented in the form of a Hasse diagram; each node in the Hasse diagram corresponds to an element in the candidate license plate string set.

[0030] A wide dynamic range light-adaptive license plate recognition system is provided to implement the aforementioned wide dynamic range light-adaptive license plate recognition method. The system includes:

[0031] Image processing and encoding module: used to acquire wide dynamic range scene images, generate license plate feature enhancement maps through dynamic contrast stretching and local brightness compensation, and encode them into half-side structure maps;

[0032] Scene discrimination module: Based on a half-side structure graph, it runs a light scene adaptive switching algorithm, constructs a five-state hidden Markov chain, calculates the posterior probability of each state through a forward-backward algorithm, and integrates and outputs a scene confidence vector;

[0033] Model recognition and integration module: used to activate scene-specific AI model groups based on scene confidence vectors, generate multi-model recognition results and construct a partial order set of recognition;

[0034] License plate binding order module: Used to perform final license plate binding and unmanned order generation by extracting the supremum and performing physical verification under the constraint of the partial order set.

[0035] Compared to existing technologies, the advantages of this invention are as follows: This invention, through a combination of dynamic contrast stretching and local brightness compensation, specifically addresses the problem in wide dynamic range lighting scenarios where traditional technologies, due to their inability to handle local illumination differences in global processing, lead to grayscale confusion between overexposed license plate characters and background, and loss of detail in underexposed characters. This effectively improves the grayscale difference between license plate characters and the background, significantly enhancing character edge clarity. By combining a half-side structure diagram with the structured representation of the license plate feature enhancement map, the illumination-character symbiotic structure is solidified, providing accurate illumination and character feature association for lighting scene discrimination. Furthermore, a five-state Hidden Markov Chain is constructed through an adaptive lighting scene switching algorithm, enabling accurate discrimination of different wide dynamic range scenes such as strong light, rainy nights, and neon lights, overcoming the limitations of poor adaptability of traditional single recognition strategies. A scene-specific AI model group based on scene confidence vector activation can call adapted models for different scenes. Combined with the integration and physical verification of multi-model results using a partial order set, it further avoids character misjudgment and omissions, reducing the need for manual review and correction. Ultimately, this invention improves the accuracy of license plate recognition in wide dynamic range lighting scenarios, reduces operating costs, ensures the smooth operation of business processes such as unmanned order generation, and solves the problem that traditional technologies cannot meet the needs of unmanned management in wide dynamic range scenarios. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1This is a flowchart of a wide dynamic range light-adaptive license plate recognition method according to the present invention;

[0038] Figure 2 This is a flowchart illustrating the activation process of the scene-specific AI model group in an embodiment of the present invention;

[0039] Figure 3 This is a flowchart illustrating the final license plate determination and unmanned order generation process in an embodiment of the present invention;

[0040] Figure 4 This is a functional block diagram of a wide dynamic range light-adaptive license plate recognition system according to the present invention. Detailed Implementation

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] Example 1:

[0043] Please see Figure 1 As shown, this embodiment provides a wide dynamic range light-adaptive license plate recognition method, including:

[0044] Step S10: Acquire wide dynamic range scene images, generate license plate feature enhancement maps through dynamic contrast stretching and local brightness compensation, and encode them into half-side structure maps.

[0045] The wide dynamic range scene image refers to traffic scene images that include various uneven lighting conditions such as direct strong light, backlighting and shadows, low light at night in rain, and neon interference, and are acquired by a wide dynamic range industrial camera.

[0046] Based on the acquired wide dynamic range scene image, the first step is to extract the region of interest (ROI) of the license plate, thus obtaining the ROI region. Specifically, the license plate ROI extraction process includes the following steps:

[0047] The first step is to perform grayscale processing on the wide dynamic range scene image, converting the RGB three-channel image into a single-channel grayscale image. The conversion formula adopts the weighted average method, that is, grayscale value = 0.299 × R channel value + 0.587 × G channel value + 0.114 × B channel value. This formula conforms to the human eye's perception of color and can preserve the grayscale difference between the license plate area and the background.

[0048] The second step is to perform median filtering on the grayscale image to remove noise (adapting to the characteristics of salt and pepper noise). The size of the filter kernel is dynamically determined according to the noise intensity in the image. When the proportion of noise pixels in the image (statistically determined by the salt and pepper noise detection algorithm) is less than 5%, a 3×3 median filter kernel is used, and when it is greater than 5%, a 5×5 median filter kernel is used, so as to efficiently remove salt and pepper noise while avoiding blurring of the license plate edges.

[0049] The third step is to use the Canny edge detection algorithm to extract edge information from the grayscale image. The high threshold and low threshold of edge detection are calculated by the Otsu algorithm. The high threshold = 1.5 × the low threshold, and the low threshold is the optimal segmentation threshold obtained by the Otsu algorithm, so as to accurately identify the rectangular outline edge of the license plate.

[0050] The fourth step is to filter the contours in the edge detection results. The filtering conditions include: the contour has 4 sides, i.e., any quadrilateral, and the area of ​​the contour is between 1000 and 5000 pixels. After performing a four-point perspective transformation to correct the selected quadrilateral contours into standard rectangles, the aspect ratio of the corrected rectangles is further verified to be between 2.5 and 3.5. The area obtained by filtering through the above conditions is the license plate ROI area.

[0051] Furthermore, a dynamic contrast stretching operation is performed on the extracted license plate ROI region to obtain the license plate ROI region after sub-block contrast stretching. Specifically, the dynamic contrast stretching refers to the operation of enhancing the contrast of different sub-blocks within the license plate ROI region based on the brightness distribution differences, in order to solve the problem of local overexposure or underexposure caused by traditional global contrast stretching. The operation steps are as follows:

[0052] The first step is to evenly divide the license plate ROI region into M×N sub-blocks. The size of the sub-blocks is determined based on the pixel size of the license plate ROI region. If the pixel size of the license plate ROI region is 300×100 pixels, then it is divided into 6×4 sub-blocks (each sub-block is 50×25 pixels). This ensures that each sub-block contains enough pixels to statistically analyze the brightness distribution and to reflect local lighting differences. M and N are the number of rows and columns of the sub-blocks.

[0053] The second step is to calculate the grayscale histogram of each sub-block and count the grayscale value range of pixels within the sub-block [G_min,i,j, G_max,i,j], where i and j are the row number and column number of the sub-block, respectively, G_min,i,j is the minimum grayscale value of the sub-block in the i-th row and j-th column, and G_max,i,j is the maximum grayscale value of the sub-block.

[0054] The third step is to perform contrast stretching on each sub-block. The stretching formula adopts an adaptive block-based linear stretching formula, that is, the formula for calculating the stretched grayscale value G' is: G'=clip[(G-G_min,i,j)×(255 / (G_max,i,j-G_min,i,j+1e-6))×C,0,255], where 1e-6 is used to avoid the denominator being 0 when G_max,i,j=G_min,i,j, and clip(˙,0,255) is used to prevent the grayscale value from overflowing the range of 0-255; G is the original grayscale value of a single pixel in the sub-block; C is the contrast adjustment coefficient, and the value of C is based on the brightness variance σ of the sub-block. 2 Calculate σ 2 It is obtained by summing the squared deviations of the gray values ​​of all pixels within a sub-block from the average gray value of the sub-block, specifically C = 1 + (σ0) 2 -σ 2 ) / σ0 2 , σ0 2 The preset standard luminance variance is obtained by statistically analyzing the luminance variance of 1000 license plate ROI sub-blocks under normal lighting conditions. When the sub-block luminance variance σ... 2 Less than σ0 2 When C > 1, to enhance contrast, when σ 2 Greater than σ0 2 When C < 1, over-enhancement is suppressed. This formula can be used to achieve adaptive adjustment of the brightness of sub-blocks.

[0055] The fourth step is to stitch all the stretched sub-blocks together in their original positions to obtain the license plate ROI area after contrast stretching.

[0056] Furthermore, based on the license plate ROI area after contrast stretching of the sub-blocks, a local brightness compensation operation is performed to obtain a license plate feature enhancement map. Specifically, the local brightness compensation refers to the operation of further adjusting the brightness of overexposed or underexposed sub-blocks that still exist after contrast stretching to achieve brightness balance across the entire license plate ROI area. The operation steps are as follows:

[0057] The first step is to calculate the average brightness L_avg,i,j of each sub-block in the license plate ROI region after contrast stretching. L_avg,i,j is the arithmetic mean of the grayscale values ​​of all pixels in the sub-block after stretching.

[0058] The second step is to set a standard brightness threshold L_std. L_std is obtained by statistically analyzing the average brightness of 500 clear license plate images under normal lighting conditions, with a value range of 120-150 (grayscale value range of 0-255).

[0059] The third step is to perform brightness compensation on each sub-block. If the average brightness L_avg,i,j of the sub-block is greater than L_std + ΔL, then the sub-block is determined to be an overexposed sub-block, and brightness attenuation is performed. ΔL is the brightness deviation threshold, preferably ΔL = 30, which is verified through experiments that can effectively distinguish overexposed areas. The average brightness of the sub-block after attenuation is L'_avg,i,j = L_std + (L_avg,i,j - (L_std + ΔL)) × k, where k is the attenuation coefficient. The value of k is calculated based on the dynamic range DR of the wide dynamic range scene image. DR is the dynamic range of the wide dynamic range scene image. The dynamic range of the industrial camera is considered. The higher the dynamic range of the camera, the larger the DR and the smaller the attenuation. If the average brightness of a sub-block is L_avg,i,j < L_std - ΔL, then the sub-block is determined to be an underexposed sub-block, and brightness enhancement is performed. The enhanced average brightness of the sub-block is L'_avg,i,j = L_std - ((L_std - ΔL) - L_avg,i,j) × m, where m is the enhancement coefficient. The value of m is related to the minimum gray value G_min,i,j of the sub-block. The smaller G_min,i,j is, the more severe the underexposure, and the larger m is.

[0060] The fourth step is to stitch all the brightness-compensated sub-blocks together in their original positions to obtain a license plate feature enhancement image. In this image, the license plate characters have clear edges, the overall brightness is balanced, and there are no obvious overexposed or underexposed areas.

[0061] Furthermore, an encoding operation is performed on the license plate feature enhancement map to generate a half-side structure map. This half-side structure map is a graph structure used to describe the spatial relationships of pixel blocks, illumination attributes, and character topology in the license plate feature enhancement map. Its expression is: Half-side structure map = (vertex set, edge set, half-side set, local brightness compensation coefficient field, character connectivity label). The definitions and generation methods of each component are as follows:

[0062] (1) Vertex set: The vertex set consists of the center coordinates and average brightness of the pixel blocks divided by the license plate feature enhancement image. First, the license plate feature enhancement image is uniformly divided into M×N sub-blocks (i.e., pixel blocks) as in the dynamic contrast stretching step. Each pixel block corresponds to a vertex. The center coordinates of the vertex are the pixel coordinates (x, y) of the geometric center of the pixel block. The average brightness of the vertex is the arithmetic mean of the gray values ​​of all pixels in the pixel block after compensation. The expression of the vertex set is vertex set = {(x_11, y_11, L_11), (x_12, y_12, L_12), ..., (x_MN, y_MN, L_MN)}; where L_ij is the average brightness of the pixel block in the i-th row and j-th column, x_MN and y_MN are the pixel coordinates of the pixel block in the M-th row and N-th column, and L_MN is the average brightness of the pixel block in the M-th row and N-th column.

[0063] (2) Edge set: The edge set is the set of line segments that connect the corresponding vertices of adjacent pixel blocks. The adjacency relationship is defined as 4-neighbor adjacency, that is, each pixel block is only adjacent to pixel blocks in the four directions of up, down, left and right. If the vertex of the pixel block in the i-th row and j-th column is adjacent to the vertex of the pixel block in the (i+1)-th row and j-th column, then an edge is established between the two vertices. The expression of the edge set is edge set = {e_1, e_2, ..., e_K}; where K is the total number of edges, K = M × (N-1) + N × (M-1);

[0064] (3) Half-edge set: The half-edge set is a subset of edges with directional attributes. Each edge corresponds to two half-edges with opposite directions. For example, the edge e connecting vertex V_ij (the vertex of the pixel block in row i and column j) and V_(i+1)j corresponds to two half-edges: the half-edge he_1 pointing from V_ij to V_(i+1)j and the half-edge he_2 pointing from V_(i+1)j to V_ij. Each half-edge carries three attributes: direction (represented by angle, such as the direction of the half-edge from left to right is 0° and the direction of the half-edge from top to bottom is 90°), gradient vector (calculated by the Sobel operator, performing Sobel edge detection on the pixels within the pixel block). The gradient values ​​in the horizontal and vertical directions are obtained, and the gradient vector is (G_x, G_y), where G_x is the horizontal gradient value and G_y is the vertical gradient value. The texture entropy (calculated through the gray-level co-occurrence matrix, with the distance of the gray-level co-occurrence matrix set to 1 and the angle set to 0°, and the texture entropy being the sum of the negative values ​​of the product of the probability and the logarithm of each element in the gray-level co-occurrence matrix, reflecting the complexity of the texture within the pixel block) is expressed as half-side set = {he_1, he_2, ..., he_2K}; where he_2K is the 2Kth half-side, 2K is the total number of half-sides in the half-side set, and K = M × (N-1) + N × (M-1).

[0065] (4) Local brightness compensation coefficient field: The local brightness compensation coefficient field is a set of coefficients defined on the vertex set. The coefficient corresponding to each vertex is the compensation coefficient used by the pixel block to which the vertex belongs in the local brightness compensation step (the attenuation coefficient k of the overexposed sub-block or the boost coefficient m of the underexposed sub-block). The expression of the local brightness compensation coefficient field is local brightness compensation coefficient field = {λ_11,λ_12,...,λ_MN}; where λ_MN is the compensation coefficient of the vertex in the Mth row and Nth column; (5) Character connectivity label: The character connectivity label is a binary label defined on the half-edge set (with a value of 0 or 1) Used to identify whether a half-edge belongs to the license plate character edge path. The license plate character edge path is determined by an edge tracking algorithm: First, perform edge detection on the license plate feature enhancement map to obtain character edge pixels; then determine whether the line segment corresponding to each half-edge intersects with the character edge pixels. If they intersect, the character connectivity label of the half-edge is 1, otherwise it is 0. The expression of the character connectivity label is character connectivity label = {τ_1, τ_2, ..., τ_2K}, where τ_p is the label of the p-th half-edge, 2K is the total number of half-edges in the half-edge set, and τ_2K is the label of the 2K-th half-edge.

[0066] The core function of step S10 is to complete the preprocessing and structured representation of license plate images in wide dynamic range scenarios, providing a high-quality feature foundation for subsequent lighting scene discrimination and license plate recognition. In terms of results, the combination of dynamic contrast stretching and local brightness compensation effectively solves the problem of local overexposure or underexposure in the license plate area in wide dynamic range scenes, significantly improving the grayscale difference between the license plate characters and the background, and enhancing the clarity of character edges compared to traditional preprocessing methods. The coded half-edge structure map breaks through the limitations of using matrices or scalar fields to represent images in traditional preprocessing, and solidifies the symbiotic structure of "lighting-characters" in the form of geometric topology. The vertex set and local brightness compensation coefficient field retain the lighting information of each pixel block, the gradient vector and texture entropy attribute of the edge set and half-edge set reflect the spatial correlation and character features between pixel blocks, and the character connectivity label directly associates the character edge path. This structure enables the overall continuity of the character area to be maintained even in scenes with strong reflection (such as the reflection caused by the midday sun shining directly on the license plate in summer) or neon interference (such as the colored light spots caused by neon lights shining on the license plate at night intersections), avoiding the global distortion caused by the loss of spatial lighting correlation in traditional preprocessing. From the perspective of subsequent application value, the half-side structure graph, as the input variable of step S20, carries the illumination continuity constraint and character connectivity topology, which provides a key physical basis for the construction of a five-state hidden Markov chain in the light scene adaptive switching algorithm in step S20. This enables the calculation of the state transition probability of the hidden Markov chain to be based on the correlation between real image illumination and character features, rather than simple statistical model assumptions. This is something that ordinary matrices or scalar fields cannot replace, and it is also the primary guarantee of the robustness of the entire wide dynamic range light adaptive license plate recognition method.

[0067] For example, in a strong midday sunlight scene during summer, in a wide dynamic range (WDR) scene image captured by a WDR industrial camera, the upper half of the license plate is overexposed due to direct sunlight (grayscale value close to 255), while the lower half is underexposed due to being in shadow (grayscale value below 50). This step involves first extracting the license plate's Region of Interest (ROI), dividing it into 6×4 sub-blocks, then performing dynamic contrast stretching (C-value reduced to 0.8) and brightness attenuation (k-value set to 0.6) on the overexposed upper sub-blocks, and performing dynamic contrast stretching (C-value increased) on the underexposed lower sub-blocks. (Up to 1.3) and brightness enhancement, the overall brightness of the license plate in the obtained license plate feature enhancement image is balanced (average brightness of about 130), and the character edges are clear; in the half-side structure image generated by further encoding, the average brightness of the vertices corresponding to the overexposed sub-blocks and the compensation coefficient record its lighting processing process, the gradient vector of the half-side set accurately captures the horizontal and vertical edges of the characters, and the half-side with the character connectivity label of 1 forms a complete character outline path. After this half-side structure image is input into step S20, it can accurately support the discrimination of the "sunny day strong light" scene, laying the foundation for subsequent model calling and recognition.

[0068] Step S20: Based on the half-side structure graph, run the light scene adaptive switching algorithm to construct a five-state hidden Markov chain and calculate the posterior probability of each state under the current observation using the forward and backward algorithm. After integration, output the scene confidence vector.

[0069] The aforementioned adaptive lighting scene switching algorithm is an adaptive decision-making algorithm based on the image structural features (i.e., half-side structure graph) under wide dynamic range scenes. Through three core steps—feature extraction, model construction, and probability calculation—it dynamically determines the current lighting scene type and outputs a scene confidence vector to guide subsequent model selection. The core logic of this algorithm is to utilize the topological correlation of the half-side structure graph to constrain the scene discrimination process, avoiding scene misjudgments caused by fragmented image region analysis in traditional methods, and achieving stable scene switching under complex lighting interference.

[0070] Specifically, the adaptive lighting scene switching algorithm includes the following three core steps:

[0071] The first step, based on the half-edge structure graph output in step S10, is to extract three types of feature parameters and jointly encode them into an observation vector. The half-edge structure graph includes a vertex set, an edge set, a half-edge set, a local brightness compensation coefficient field, and character connectivity labels. The first type of feature parameter is the vertex brightness distribution histogram, obtained by statistically analyzing the distribution of the average brightness values ​​of all vertices in the vertex set. Specifically, the brightness value range (0-255) is divided into several intervals, and the proportion of vertices in each interval to the total number of vertices is calculated to form histogram data. The second type of feature parameter is the half-edge gradient direction consistency, obtained by calculating the consistency of the direction angles of the gradient vectors of all half-edges in the half-edge set. Specifically, for each half-edge gradient vector (Gx, Gy), the direction angle θ = atan2(Gy, Gx) is calculated, where atan2(*, *) represents the two-parameter arctangent function; then each θ is converted into a unit vector (cosθ, sinθ), and the sum of all unit vectors is calculated. The magnitude of this sum is taken as the consistency index; the closer the magnitude is to the total number of half-edges, the higher the gradient direction consistency. The third type of feature parameter is the temporal fluctuation of texture entropy, which is calculated by the difference in texture entropy between the half-structure graphs corresponding to three consecutive frames. Specifically, the texture entropy of the half-set of the current frame and the previous two frames is taken, the absolute value of the difference in texture entropy between adjacent frames is calculated, and then the average of these two absolute values ​​is taken as the temporal fluctuation value. The larger this value is, the higher the probability of flickering or dynamic light and shadow interference in the scene. The above three types of feature parameters are concatenated in sequence to form an observation vector. The dimension of the observation vector is the sum of the dimensions of the three types of feature parameters. The dimension of the vertex brightness distribution histogram is determined by the number of interval divisions, while the half-gradient direction consistency and the temporal fluctuation of texture entropy are each 1-dimensional.

[0072] The second step is to construct a five-state Hidden Markov Chain (HMC) based on the observation vectors. The expression for the five-state HMC is: Five-state HMC = (State Set, Transition Matrix, Observation Probability, Initial Probability). The state set contains five elements: bright sunlight, low light during rainy nights, neon interference, mixed backlighting, and normal lighting, each corresponding to a typical lighting scenario. The initial probability is the probability distribution of each state at the initial moment, determined by statistically analyzing the frequency of each scenario in historical data; that is, the initial probability of a state is equal to the ratio of the number of times that state appears in historical data to the total number of occurrences.

[0073] Specifically, the transition matrix is ​​constructed as follows: It is a 5×5 matrix, where each element represents the probability of transitioning from one state to another. The values ​​of the transition matrix are calculated based on the local brightness compensation coefficient field of the half-sided structure graph. For any two adjacent vertices, the difference in their local brightness compensation coefficients, Δλ = |λ|, is calculated. f -λ g |;λ f , λ gLet represent any two adjacent vertices; then calculate the illumination continuity weight ω=exp(-Δλ) for the adjacent regions using Δλ. 2 / η 2 ), where η is the scale parameter, determined by statistically analyzing the standard deviation of the λ difference between adjacent vertices in a normal scene. The global illumination continuity weight ω_avg is obtained by averaging the illumination continuity weights ω of all adjacent vertices. The values ​​of the elements in the transition matrix are positively correlated with ω_avg; that is, the larger ω_avg is, the more stable the current scene illumination, the lower the state transition probability, and the higher the probability of maintaining the current state. The element A[s][t] in the transition matrix represents the probability of transitioning from state s to state t, where s and t are state-specific identifiers, each ranging from 1 to 5, corresponding to the five lighting scene states of the five-state Hidden Markov Chain: bright sunlight, low illumination at night, and neon. For interference, mixed backlight, and normal conditions, the formula for calculating A[s][t] is: A[s][t]=(1-α×(1-ω_avg))×I+α×(1-ω_avg)×P0[s][t], where α is the adjustment coefficient, obtained by calibrating the frequency of state transitions in historical data, P0[s][t] is the initial transition probability distribution based on historical data, representing the initial probability of transitioning from state s to state t; I is the indicator function, where I=1 when s=t, i.e., when transitioning from a certain state to itself; and I=0 when s≠t, i.e., when transitioning from a certain state to another state.

[0074] The observation probability is a 5×L matrix, representing the probability of observing a specific feature parameter in a given state, where L is the dimension of the observation vector. The observation probability is calculated based on the matching degree between each feature dimension of the observation vector and the corresponding feature template of each state. The feature template of each state is obtained by splitting and statistically analyzing the observation vectors of historical data according to their feature dimensions. Specifically, the one-dimensional Euclidean distance d between the value of the v-th feature dimension of the observation vector and the mean of the v-th feature dimension of the feature template corresponding to state s is calculated, and then the observation probability B[s][v] is calculated using d. Where B[s][v] represents the probability of observing the feature value of the v-th feature dimension in state s. The variance of the v-th feature dimension observed in state s is calculated by the dispersion of the v-th feature dimension values ​​belonging to that state in historical data.

[0075] The third step involves using the constructed five-state Hidden Markov Chain and observation vectors to calculate the posterior probability of each state under the current observation using a forward-backward algorithm. Specifically, the forward-backward algorithm includes a forward process and a backward process: the forward process calculates the probability of being in a certain state and observing the corresponding feature up to the current time step; the backward process calculates the probability of being in a certain state and observing the corresponding feature in the future starting from the current time step; the forward and backward probabilities are multiplied and normalized to obtain the posterior probability of each state. The posterior probabilities of the five states are arranged in the order of the state set to obtain the scene confidence vector. Each element of the scene confidence vector corresponds to the posterior probability of a state, with a value ranging from 0 to 1, and the sum of all elements is 1.

[0076] Step S20 aims to accurately identify lighting scenes, providing a basis for subsequent model selection. Its advantage lies in utilizing the topological continuity constraint of a half-sided graph to constrain the state transitions of a five-state Hidden Markov Chain, solving the scene misjudgment problem caused by the fragmented analysis of the license plate ROI region in traditional methods. By introducing the local brightness compensation coefficient field into the calculation of the transition matrix, the state transition probability can reflect the true trend of illumination changes, maintaining the stability of state judgment even in scenarios with combined interference from flickering and raindrop refraction. Compared to traditional scene discrimination methods based on pixel statistics, this step improves the accuracy of scene discrimination through a graph-driven spatiotemporal smoothing scene discrimination mechanism. The scene confidence vector, as the input to step S30, directly ensures the correct activation of the subsequent scene-specific AI model group, serving as the core support for the robustness of the entire algorithm.

[0077] For example, in a neon interference scenario, the vertex brightness distribution histogram of the half-structure graph exhibits multi-peak characteristics (corresponding to the colored spots of neon lights), the gradient direction consistency of the half-side is low (character edges are interfered with by light spots), and the texture entropy fluctuates significantly over time (neon lights flicker). After the observation vector constructed based on these features is input into a five-state Hidden Markov Chain, the posterior probability calculated using the forward-backward algorithm shows the highest probability for the neon interference state. This results in the corresponding dimension of the scene confidence vector having a significantly higher value than other dimensions, providing a reliable basis for activating the anti-flicker model group in step S30.

[0078] Step S30: Based on the scene confidence vector, activate the scene-specific AI model group, generate multi-model recognition results, and construct a partial order set of recognition.

[0079] Based on the scene confidence vector output in step S20, the main scene is first determined. The scene confidence vector contains five elements, corresponding to the confidence levels of five lighting scenarios: bright sunshine, low light in rainy night, neon interference, mixed backlight, and normal lighting. By comparing the values ​​of the five elements in the scene confidence vector, the scene corresponding to the element with the largest value is determined as the main scene. If two or more elements have the same value and are all maximum values, the vertex brightness distribution histogram of the half-side structure graph is used for further judgment, and the scene that better matches the vertex brightness distribution characteristics is selected as the main scene.

[0080] Based on the determined main scene, activate the corresponding scene-specific AI model group. The scene-specific AI model group comprises five sub-model groups, each corresponding to one of five lighting scenarios. Each sub-model group consists of 3-5 deep learning models. Specifically: the sunny, high-light scene-specific AI model group includes a specular suppression convolutional neural network, a character edge enhancement network, and a bidirectional long short-term memory recognition model; the rainy night, low-light scene-specific AI model group includes a low-light image enhancement generative adversarial network, a raindrop removal network, and an attention-based character recognition model; the neon interference scene-specific AI model group includes an anti-flicker convolutional neural network, a frequency-domain Transformer optical character recognition model, and a color channel separation recognition model; the mixed backlight scene-specific AI model group includes a backlight region segmentation network, a multi-exposure fusion network, and a context-aware character recognition model; and the normal scene-specific AI model group includes a basic convolutional neural network, a character feature extraction network, and a conventional optical character recognition model. Please refer to [reference needed]. Figure 2 As shown.

[0081] Specifically, the input to each model is the license plate feature enhancement map generated in step S10, and the output is a candidate license plate string. Taking the anti-flicker convolutional neural network in the dedicated AI model group for neon interference scenes as an example, its network structure includes an input layer, 5 convolutional blocks, 2 pooling layers, 3 fully connected layers, and an output layer. The input layer receives pixel data from the license plate feature enhancement map; the convolutional blocks consist of convolutional layers, batch normalization layers, and activation functions. The convolutional layers use 3×3 convolutional kernels, the number of which doubles with the network depth, to extract anti-flicker features from the image; the pooling layers use 2×2 max pooling to reduce the dimension of the feature map; the fully connected layers are used to map the extracted features to a character probability distribution; the output layer outputs the candidate license plate string through the softmax function. This model is trained on a dataset containing a large number of license plate images from neon interference scenes. During training, the cross-entropy loss function is used, the Adam algorithm is selected as the optimizer, and the learning rate is dynamically adjusted through a cosine annealing strategy to improve the model's character recognition accuracy under flicker interference.

[0082] The license plate feature enhancement map is input into each model in the activated scene-specific AI model group to obtain multiple candidate license plate strings, which form a candidate license plate string set. Each element in the candidate license plate string set is a string of 7 to 8 characters, containing Chinese characters, letters, and numbers. The first character is the abbreviation of the province in Chinese characters, the second character is a letter, and the last five characters (for ordinary license plates) or six characters (for new energy vehicle license plates) are a combination of letters and numbers.

[0083] Based on the candidate license plate string set, a partial order set for identification is constructed. The expression for the partial order set is: Partial order set = (Candidate license plate string set, Partial order relation). Here, the partial order relation is a binary relation defined on the candidate license plate string set, used to describe the inclusion, conflict, and consensus relationships among elements in the set.

[0084] Specifically, the partial order relation is defined as follows: For any two elements, candidate license plate string a and candidate license plate string b, in the candidate license plate string set, if the edit distance between candidate license plate string a and candidate license plate string b is ≤ 1, both candidate license plate string a and candidate license plate string b are output by at least two models, and the number of model outputs for candidate license plate string a is ≤ the number of model outputs for candidate license plate string b (when the number of outputs is equal, the character validity score of candidate license plate string a is ≤ the character validity score of candidate license plate string b, and the score is calculated based on the compliance of the province abbreviation, letter I / O prohibition rules, and other physical prior knowledge of the license plate), then candidate license plate string a and candidate license plate string b are said to satisfy a partial order relation, denoted as candidate license plate string a ≤ b. Here, the edit distance refers to the minimum number of single-character edit operations required to convert candidate license plate string a into candidate license plate string b. Single-character edit operations include insertion, deletion, and replacement. The edit distance is calculated using a dynamic programming algorithm, specifically: Define a (u+1)×(w+1) matrix, where u is the length of candidate license plate string a, w is the length of candidate license plate string b, and the element d[u][w] in the matrix represents the edit distance between the first u characters of candidate license plate string a and the first w characters of candidate license plate string b. Initialize the first row and first column of the matrix to 0 to u and 0 to w, respectively; for other elements in the matrix, if the u-th character of candidate license plate string a is the same as the w-th character of candidate license plate string b, then d[u][w] = d[u-1][w-1]; otherwise, d[u][w] = min(d[u-1][w]+1, d[u][w-1]+1, d[u-1][w-1]+1), where d[u-1][w]+1 represents a deletion operation, d[u][w-1]+1 represents an insertion operation, and d[u-1][w-1]+1 represents a replacement operation. Finally, d[u][w] is the edit distance between candidate license plate string a and candidate license plate string b.

[0085] The recognized poset is represented in the form of a Hasse diagram, also known as a Hasse graph in Chinese. Each node in the Hasse diagram corresponds to an element in the set of candidate license plate strings, that is, the candidate license plate strings. If the candidate license plate strings corresponding to two nodes satisfy a partial order relationship and there is no other node through which these two nodes indirectly satisfy the partial order relationship, then a directed edge is established between these two nodes, with the direction pointing from the candidate license plate string in the front of the partial order relationship to the candidate license plate string in the back. The construction steps of the Hasse diagram are as follows: First, calculate the partial order relationship between all elements in the set of candidate license plate strings to obtain a partial order relationship matrix. Second, according to the partial order relationship matrix, remove the transitive relationship, that is, if there exists a candidate license plate string a ≤ candidate license plate string c and candidate license plate string c ≤ candidate license plate string b, then delete the direct partial order relationship between candidate license plate string a and candidate license plate string b. Third, represent the candidate license plate strings with nodes and the partial order relationship after removing the transitive relationship with directed edges, and draw to obtain the Hasse diagram.

[0086] The function of this step S30 is to effectively integrate the recognition results of multiple models to form a structured recognized poset, providing a reliable intermediate result for the final determination of the license plate. Its advantage is that, compared with the traditional voting mechanism, the recognized poset can retain the semantic association between candidate license plate strings. In the case of character mutilation or mirroring, etc., the maximum consensus substring can still be traced through the partial order path. By introducing the edit distance and the number of model outputs as the determination basis of the partial order relationship, the partial order relationship can accurately reflect the similarity and credibility between different candidate license plate strings. The activation mechanism of the scene-specific AI model group ensures that the most suitable model can be called for recognition in different lighting scenarios, improving the overall quality of the recognition results of multiple models. The recognized poset, as the input of step S40, provides sufficient basis for the supremum extraction and physical verification with its rich structural information, which is the key link to achieve high-accuracy license plate recognition under complex lighting conditions.

[0087] Exemplarily, in the neon interference scene, activate the neon interference scene-specific AI model group. The three models respectively output candidate license plate strings Beijing A12345, Beijing A12346, and Beijing A12345. The set of candidate license plate strings is {Beijing A12345, Beijing A12346}. Calculate the edit distance. The edit distance between Beijing A12345 and Beijing A12346 is 1, and both are output by at least 2 models. Therefore, there are partial order relationships Beijing A12345 ≤ Beijing A12346 and Beijing A12346 ≤ Beijing A12345 (because the edit distance is bidirectional). The constructed Hasse diagram contains two nodes, corresponding to the two candidate license plate strings respectively. There is a bidirectional directed edge between the two nodes, indicating that there is a consensus relationship between them, providing a clear structural basis for the supremum extraction in step S40.

[0088] Step S40: Under the constraints of the constructed partial order set, the final license plate binding and unmanned order generation are performed through supremacy extraction and physical verification.

[0089] Based on the partial order set output in step S30, the Hasse graph it contains is first traversed to find the global maximum element. The partial order set consists of a set of candidate license plate strings and partial order relations. Each node in the Hasse graph corresponds to an element in the set of candidate license plate strings, and directed edges represent partial order relations. The global maximum element is the candidate license plate string corresponding to the unique node in the Hasse graph that is pointed to by all elements with which it has a partial order relation.

[0090] Specifically, the method for extracting the global maximum element is as follows:

[0091] The first step is to traverse all nodes in the Hasse graph, record the out-degree of each node (out-degree represents the number of edges originating from that node), and analyze the reachability of each node through path traversal (i.e., determine whether other nodes can reach that node through direct or indirect edges in the Hasse graph).

[0092] The second step is to filter out nodes that are "reachable by all other nodes through the Hasse graph and have an out-degree of 0". If such a node exists and is unique, then the candidate license plate string corresponding to that node is the global maximum element.

[0093] The third step is to determine if there are multiple nodes that meet the conditions or no nodes that meet the conditions, and then it is determined that there is no unique maximum element.

[0094] When there is no unique maximum element, search for the minimum edit distance completion path in the partial order set. The minimum edit distance completion path is defined as a path starting from any node in the Hasse graph, where the minimum number of characters added, deleted, or replaced results in the string corresponding to the endpoint satisfying the partial order constraints of all nodes. Specifically, the search method is as follows:

[0095] The first step is to calculate the edit distance matrix among all elements in the candidate license plate string set. The edit distance is calculated using the dynamic programming algorithm defined in step S30.

[0096] The second step is to construct a path graph based on the edit distance matrix, with each node as the starting point. The weight of the edge in the path graph is the edit distance between the strings corresponding to the two nodes.

[0097] The third step is to use Dijkstra's algorithm to find the path with the minimum total weight in the path graph. This path is the minimum edit distance completion path, and the string corresponding to the end point of the path is used as the candidate license plate string after completion.

[0098] The validity of the completed candidate license plate string or the global maximum element is pruned based on the physical prior of the license plate. The physical prior includes rules for ordinary license plates and rules for new energy vehicle license plates. The rules for ordinary license plates are: the first character is the abbreviation of the province in Chinese characters, the second character is a letter, and the last five characters are a combination of letters and numbers, with a total length of 7 characters. The rules for new energy vehicle license plates are:

[0099] The first character is the abbreviation of the province in Chinese characters, the second character is the letter D or F, and the last six characters are a combination of letters and numbers, for a total length of 8 characters. Specifically, the legality pruning steps are as follows:

[0100] The first step is to check if the string length is 7 or 8 characters.

[0101] The second step is to check if the length is 7 characters, whether the first character is the abbreviation of the province in Chinese characters, whether the second character is a letter, and whether the last five characters are letters or numbers.

[0102] The third step is to check if the length is 8 characters, and if the first character is the abbreviation of the province in Chinese characters, the second character is D or F, and the last six characters are letters or numbers.

[0103] The fourth step is to retain the strings that conform to the above rules and remove the strings that do not conform to the rules. If only one string is retained, it is used as the final license plate string. If multiple strings are retained, the average edit distance between these strings and all elements in the partial order set is recalculated, and the string with the smallest average edit distance is selected as the final license plate string.

[0104] The final license plate string is matched against the vehicle information database to obtain the corresponding vehicle ID. The vehicle information database stores the correspondence between license plate strings and vehicle IDs; the vehicle ID is a unique identifier for the vehicle and is obtained through a precise query using the license plate string. The vehicle ID is then associated with the unmanned order system to automatically generate an order record containing the vehicle ID, license plate string, recognition time, and scene type, completing the unmanned billing process. Please refer to the following procedure: Figure 3 As shown.

[0105] The purpose of step S40 is to determine the unique and correct license plate from the multi-model recognition results under the constraint of the partially ordered set, through a combination of mathematical methods and physical rules, and to automatically associate it with the order system. Its advantage lies in the fact that, compared to traditional post-processing methods that rely on manual intervention or simple rules, this step utilizes the mathematical completeness of the partially ordered set, preserving recognition diversity while forcing convergence to a physically valid solution, thus avoiding decision failures when multiple models conflict. By combining minimum edit distance path completion and physical prior knowledge of the license plate, the accuracy of the final license plate is guaranteed even in cases of incomplete or blurred characters. This autonomous closed-loop algorithm completely replaces manual correction, improving the operational efficiency and reliability of unattended systems in wide dynamic scenarios.

[0106] Example 2:

[0107] This embodiment, based on Embodiment 1, provides a wide dynamic range light-adaptive license plate recognition system, such as... Figure 4 As shown, it includes:

[0108] A wide dynamic range light-adaptive license plate recognition system is provided to implement the aforementioned wide dynamic range light-adaptive license plate recognition method. The system includes:

[0109] Image processing and encoding module: used to acquire wide dynamic range scene images, generate license plate feature enhancement maps through dynamic contrast stretching and local brightness compensation, and encode them into half-side structure maps;

[0110] Scene discrimination module: Based on a half-side structure graph, it runs a light scene adaptive switching algorithm, constructs a five-state hidden Markov chain, calculates the posterior probability of each state through a forward-backward algorithm, and integrates and outputs a scene confidence vector;

[0111] Model recognition and integration module: used to activate scene-specific AI model groups based on scene confidence vectors, generate multi-model recognition results and construct a partial order set of recognition;

[0112] License plate binding order module: Used to perform final license plate binding and unmanned order generation by extracting the supremum and performing physical verification under the constraint of the partial order set.

Claims

1. A wide dynamic range light-adaptive license plate recognition method, characterized in that, The method includes: Step S10: Acquire a wide dynamic range scene image, generate a license plate feature enhancement map through dynamic contrast stretching and local brightness compensation, and encode it as a half-edge structure map; the half-edge structure map is a graph structure used to describe the spatial relationship of pixel blocks, illumination attributes, and character topology in the license plate feature enhancement map, including a vertex set, edge set, half-edge set, local brightness compensation coefficient field, and character connectivity label; the local brightness compensation coefficient field is a set of coefficients defined in the vertex set, and the coefficient corresponding to each vertex is the compensation coefficient used by the pixel block to which the vertex belongs in local brightness compensation; the character connectivity label is a binary label defined in the half-edge set, used to identify whether the half-edge belongs to the edge path of the license plate character; Step S20: Based on the half-side structure diagram, run the light scene adaptive switching algorithm to construct a five-state hidden Markov chain and calculate the posterior probability of each state under the current observation through the forward and backward algorithm, and output the scene confidence vector after integration; Step S30: Based on the scene confidence vector, activate the scene-specific AI model group, generate multi-model recognition results, and construct a partial order set of recognition results; Step S40: Under the constraints of the constructed partial order set, the final license plate binding and unmanned order generation are performed through supremacy extraction and physical verification.

2. The wide dynamic range light-adaptive license plate recognition method according to claim 1, characterized in that, The dynamic contrast stretching is: Divide the area of ​​interest of the license plate into sub-blocks evenly; Calculate the grayscale histogram of each sub-block and count the range of grayscale values ​​of pixels within the sub-block; Perform contrast stretching on each sub-block based on the grayscale value range; By splicing all the stretched sub-blocks back to their original positions, the region of interest for the license plate with contrast-stretched sub-blocks is obtained.

3. The wide dynamic range light-adaptive license plate recognition method according to claim 1, characterized in that, The local brightness compensation is as follows: Calculate the average brightness of each sub-block in the region of interest of the license plate after contrast stretching; Set a standard brightness threshold; Brightness compensation is performed on each sub-block according to a standard brightness threshold. The brightness compensation includes: if the average brightness of the sub-block is greater than the sum of the standard brightness threshold and the brightness deviation threshold, the sub-block is determined to be an overexposed sub-block and brightness attenuation is performed; if the average brightness of the sub-block is less than the difference between the standard brightness threshold and the brightness deviation threshold, the sub-block is determined to be an underexposed sub-block and brightness enhancement is performed. All the sub-blocks after brightness compensation are stitched together in their original positions to obtain the license plate feature enhancement image.

4. The wide dynamic range adaptive license plate recognition method according to claim 1, characterized in that, The adaptive switching algorithm for lighting scenes is as follows: Three types of feature parameters are extracted from the half-side structure graph and encoded into observation vectors; the three types of feature parameters include vertex brightness distribution histogram, half-side gradient direction consistency, and texture entropy temporal fluctuation. Construct a five-state hidden Markov chain based on observation vectors; The forward-backward algorithm is used to calculate the posterior probabilities of the five states under the current observation; The integrated output scene confidence vector is then used.

5. The wide dynamic range light-adaptive license plate recognition method according to claim 1, characterized in that, The five-state hidden Markov chain consists of a state set, a transition matrix, observation probabilities, and initial probabilities. The state set includes five lighting scene states: bright sunlight, low light in rainy nights, neon interference, mixed backlighting, and normal.

6. The wide dynamic range light-adaptive license plate recognition method according to claim 1, characterized in that, The scene confidence vector is obtained by arranging the posterior probabilities of five states under the current observation in the order of the state set. Each element of the scene confidence vector corresponds to the posterior probability of a light scene state.

7. The wide dynamic range adaptive license plate recognition method according to claim 1, characterized in that, The scene-specific AI model group comprises five sub-model groups, each corresponding to one of five lighting scenarios. The five sub-model groups include: a dedicated AI model group for sunny, bright light scenarios; a dedicated AI model group for rainy, low-light scenarios; a dedicated AI model group for neon interference scenarios; a dedicated AI model group for mixed backlight scenarios; and a dedicated AI model group for normal scenarios.

8. The wide dynamic range light-adaptive license plate recognition method according to claim 1, characterized in that, The identification partial order set is composed of a set of candidate license plate strings and a partial order relation, expressed as identification partial order set = (candidate license plate string set, partial order relation), and represented in the form of a Hasse diagram; each node in the Hasse diagram corresponds to an element in the set of candidate license plate strings.

9. A wide dynamic range adaptive license plate recognition system, used to implement the wide dynamic range adaptive license plate recognition method according to any one of claims 1-8, characterized in that, The system includes: Image processing and encoding module: used to acquire wide dynamic range scene images, generate license plate feature enhancement maps through dynamic contrast stretching and local brightness compensation, and encode them into half-side structure maps; Scene discrimination module: Based on a half-side structure graph, it runs a light scene adaptive switching algorithm, constructs a five-state hidden Markov chain, calculates the posterior probability of each state through a forward-backward algorithm, and integrates and outputs a scene confidence vector; Model recognition and integration module: used to activate scene-specific AI model groups based on scene confidence vectors, generate multi-model recognition results and construct a partial order set of recognition; License plate binding order module: Used to perform final license plate binding and unmanned order generation by extracting the supremum and performing physical verification under the constraint of the partial order set.

Images