A mileage calibration system based on a multi-stage deep learning model and a multi-dimensional false detection prevention mechanism

Abstract

The application discloses a kind of based on multi-stage deep learning model and multi-dimensional false detection prevention mechanism's mileage verification system, including double-layer nested adaptive search module: the collected scene image is divided into several overlapping sub-regions and searches, and the mileage equipment of the sub-region is roughly positioned, and the current magnification value and sub-region coordinate parameter are persistently stored;Multi-stage deep learning model cascade inference architecture: the application is positioned by double-layer nested adaptive search module, improves the adaptive ability of system under continuous verification scene;Through double-layer geometric correction module, the fundamental restriction of image distortion on recognition rate is eliminated;Through multi-stage cascade inference, false detection rate is exponentially reduced with the identification stage progressive;Through the time domain filtering of multi-stage deep learning model cascade inference architecture, digital tube transient noise and isolated error frame are systematically removed from output stream;Through reverse historical verification module, the probability of signal false emission is reduced to very low.

Description

Technical Field

[0001] This invention belongs to the field of mileage measurement verification and relates to a mileage verification system based on a multi-stage deep learning model and a multi-dimensional anti-false detection mechanism. Background Technology

[0002] In automated mileage verification scenarios, mileage numerical recognition is the core step in the verification process. The recognition system needs to continuously and accurately read the accumulated mileage value displayed on the screen from the real-time video stream and send a collection signal to external devices when the value reaches a preset verification point. This scenario places extremely high demands on recognition accuracy and signal transmission reliability; any misreading or mistransmission will directly affect the legal validity of the verification conclusion.

[0003] However, existing technologies face the following specific technical problems in achieving the above goals: First, the odometer device is limited by its installation location, angle, and physical size, making it difficult to ensure it is in the optimal recognition area, resulting in initial positioning difficulties and a high failure rate for initial locking due to the lack of an adaptive search mechanism; Second, the screen image of the odometer device suffers from complex geometric distortions, reducing the input quality of the recognition model; Third, the screen types of odometer devices are diverse, and existing single-model solutions are insufficient to adapt to cross-type scenarios; Fourth, the output results of a single frame of the model cannot be directly trusted, lacking a quality screening mechanism across frame time dimensions; Fifth, the recognition result stream contains transient noise from digital tube switching and isolated misidentified frames, lacking multi-dimensional time-domain filtering capabilities; Sixth, there is a lack of historical state reverse verification before the odometer value update trigger signal is sent, posing a risk of false transmission.

[0004] In summary, existing technologies urgently need a precise mileage identification method that covers the entire link from adaptive search of camera parameters, image geometric correction, adaptive type recognition, cross-frame temporal stability judgment to reverse verification before transmission, in order to systematically solve the above-mentioned false detection and false transmission problems. Summary of the Invention

[0005] In order to overcome at least one of the shortcomings of the prior art, the present invention provides a mileage verification system based on a multi-stage deep learning model and a multi-dimensional anti-false detection mechanism.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a mileage verification system based on a multi-stage deep learning model and a multi-dimensional false detection prevention mechanism, comprising... Double-layer nested adaptive search module: Divides the acquired scene image into several overlapping sub-regions and searches them, coarsely locates the mileage of the sub-regions, and persistently stores the current magnification value and sub-region coordinate parameters; Multi-stage deep learning model cascaded inference architecture: sequentially performs screen type determination, mileage window detection, digit region purification and numerical recognition on the mileage device image, and outputs a numerical string; Multi-stack collaboration module: performs multi-dimensional time-domain filtering on the recognition result stream composed of numerical strings and outputs stable mileage values; Reverse history verification module: Performs reverse history verification on the mileage stability value, and allows the transmission of acquisition signals after successful verification.

[0007] Furthermore, the dual-layer nested adaptive search module consists of an outer layer controller and an inner layer searcher. The outer layer controller uses the optical magnification of the industrial camera as the outer layer control variable, while the inner layer searcher uses the region scanning within the image plane as the inner layer search variable. The outer layer controller adjusts the camera magnification step by step according to the magnification step control mechanism with a preset step increment. At each magnification level, the inner layer searcher divides the image into several overlapping sub-regions in sequence according to the automatic region scanning mechanism with a predefined region scanning path, and calls the device detection and recognition model to identify each sub-region.

[0008] Furthermore, the multi-stage deep learning model cascaded inference architecture includes a mileage device localization and screen type determination module, a mileage window detection and inverse perspective mapping module, a digital region purification module, and a numerical classification and recognition module connected in series. The mileage device localization and screen type determination module, the mileage window detection and inverse perspective mapping module, the digital region purification module, and the numerical classification and recognition module are set independently of each other, and the output of the previous module strictly constrains the search space of the next module.

[0009] Furthermore, the mileage verification system also includes a confidence and non-maximum suppression joint screening module. This module has a confidence and non-maximum suppression joint screening mechanism. The confidence and non-maximum suppression joint screening mechanism performs hard filtering based on the confidence threshold to eliminate low-quality candidates, and performs non-maximum suppression screening on the filtered candidate box set to retain the single candidate box with the highest confidence, while all others are eliminated. Finally, only one optimal localization result is retained, thus eliminating multi-box noise and overlapping misselection from the model output level.

[0010] Furthermore, the mileage verification system also includes a dual-layer image geometric correction module. The dual-layer image geometric correction module is equipped with a dual-layer image geometric correction model. The dual-layer image geometric correction model performs perspective transformation to straighten the tilted mileage device as a whole to a standard rectangular view and outputs a corrected mileage device image.

[0011] Furthermore, the mileage verification system also includes a continuous frame stable queue verification and front-and-back frame position comparison module. The continuous frame stable queue verification and front-and-back frame position comparison module is equipped with a continuous frame stable queue verification and front-and-back frame position comparison mechanism. Based on the continuous frame stable queue verification and front-and-back frame position comparison mechanism, the mileage window of the continuous frames is verified to eliminate jitter frame interference and obtain a stable mileage window.

[0012] Furthermore, the dual-layer image geometric correction model is set as dual-layer distortion correction. Dual-layer distortion correction performs two levels of distortion correction on the mileage numerical image. The first layer performs perspective transformation based on the four vertex coordinates of the mileage device rotation frame to straighten the tilted mileage device as a whole to a standard rectangular view. The second layer performs affine rotation correction based on the rotation angle of the mileage window in the original video frame to eliminate local angle deviations.

[0013] Furthermore, the mileage window detection and inverse perspective mapping module includes a mileage window detection module and an inverse perspective mapping module. The mileage window detection module is equipped with a mileage window detection model. The mileage window detection model precisely locates and corrects the position of the mileage window in the mileage device image. The mileage window position is filtered by a confidence and non-maximum suppression joint screening module, inversely mapped by the inverse perspective mapping module, and corrected and cropped by a dual-layer image geometric correction module to output a clean mileage window image.

[0014] Furthermore, the mileage window detection and inverse perspective mapping module pre-acquires and persistently stores the position parameters of the mileage window in the original video frame coordinate system based on the mileage position pre-acquisition mechanism.

[0015] Furthermore, the three data stacks of the multi-stack collaboration module perform string classification processing on the numerical recognition output of each frame. The three data stacks include... The correct data stack includes the identification results whose differences from the current stable value are within the tolerance range; The difference in the normal jump stack is approximately equal to the recognition result of the standard step size; The exception jump stack includes other unexpected results.

[0016] In summary, the advantages of this invention are: This invention addresses the initial positioning failure caused by differences in the location of mileage measuring devices through a double-layer nested adaptive search module, improving the system's adaptability in continuous verification scenarios. A double-layer geometric correction module eliminates the fundamental limitation of image distortion on recognition rate. Mileage location pre-acquisition enables precise identification of both LED and LCD screens while maintaining real-time processing efficiency. Multi-stage cascaded inference exponentially reduces the false detection rate as the recognition stage progresses. A multi-stage deep learning model cascaded inference architecture with temporal filtering systematically removes transient noise and isolated erroneous frames from the output stream. A reverse history verification module reduces the probability of signal mistransmission to an extremely low level, providing a highly reliable recognition algorithm solution for automated mileage verification scenarios. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the mileage verification system of the present invention.

[0018] Figure 2 This is a schematic diagram of the double-layer nested adaptive search module of the present invention.

[0019] Figure 3 This is a schematic diagram of the inverse perspective mapping in stage two of the present invention.

[0020] Figure 4 This is a schematic diagram of the dual-layer image geometric correction of the present invention.

[0021] Figure 5 This is a schematic diagram illustrating the confidence level and cross-frame stability joint screening of the present invention.

[0022] Figure 6 This is a flowchart of the multi-stack collaboration module of the present invention.

[0023] Figure 7 This is a flowchart of the reverse history verification mechanism before transmission according to the present invention. Detailed Implementation

[0024] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0025] Existing technologies face the following specific technical problems in achieving the above objectives: First, initial positioning of odometers is difficult, and the lack of an adaptive search mechanism leads to a high initial positioning failure rate. In actual verification scenarios, the installation position, angle, and physical dimensions of odometers, such as taxi meters, vary significantly depending on the vehicle model and installation location. The initial field of view and focal length settings of optical acquisition equipment (industrial cameras) often cannot guarantee that the meter is precisely in the optimal recognition area. Existing technologies generally rely on preset fixed parameters or manual adjustment of camera parameters to position the meter, requiring repeated manual adjustments for different vehicle models, resulting in low automation and poor adaptability. When the meter is too small in the image or off-center from the field of view, the recognition success rate of the equipment detection and recognition model drops sharply. Especially in continuous assembly line scenarios, the positional differences of the meter for each vehicle cause frequent positioning failures, severely impacting verification efficiency.

[0026] Secondly, the images on the odometer screen suffer from complex geometric distortions, directly impacting the input quality of the recognition model. Because the optical acquisition device's viewing axis is rarely perfectly perpendicular to the meter screen, and the limited space inside the vehicle leads to varying shooting angles, the acquired images generally exhibit two distinct types of geometric distortion: one is overall perspective trapezoidal distortion, manifested as an irregular quadrilateral shape of the meter's rectangular outline in the image; the other is local rotational distortion, characterized by a slight rotational angle deviation of the mileage display area within the meter. Existing solutions typically correct only one type of distortion or simply apply a uniform transformation to the entire image, resulting in morphological distortion in the character images fed into the recognition model, thus lowering the recognition rate.

[0027] Third, mileage measurement devices come in a variety of screen types, and existing single-model solutions are insufficient to adapt to cross-type scenarios. Mileage measurement devices, such as taxi meters, are mainly divided into two display types: LED (light-emitting diode) and LCD (liquid crystal display). The two differ fundamentally in pixel distribution characteristics, stroke contrast, and backlight mode. If a single recognition model is used to make mixed inferences for the two types, the model will inevitably compromise between the two types of scenarios, making it difficult to achieve optimal recognition results for either, and it is more likely to produce misrecognitions at the type boundaries.

[0028] Fourth, the output results of a single frame cannot be directly trusted, lacking a quality screening mechanism across frames and temporal dimensions. Deep learning target device detection and recognition models output multiple candidate boxes for each frame. In practice, it has been found that due to slight image jitter caused by engine vibration and changes in lighting during transition frames, the positions of candidate boxes output in a single frame fluctuate, and the confidence level also fluctuates. If the candidate box with the highest confidence level in a single frame is directly used as the mileage window localization result, the noise boxes generated by the jitter will contaminate subsequent character recognition input, leading to misidentification.

[0029] Fifth, the recognition result stream contains transient noise from digital tube switching and isolated misidentified frames, lacking multi-dimensional temporal filtering capabilities. In mileage devices, such as taximeter digital tubes, there is an aliasing transition state during digit switching where old strokes are not completely off while new strokes are already lit. High-speed cameras easily capture this intermediate state, causing the OCR model to output incorrect values. Furthermore, occasional changes in lighting and specular reflections can also generate isolated misidentified frames. If existing systems lack temporal hierarchical filtering of the recognition result stream, these erroneous values ​​will directly pass through the defense layer, leading to subsequent logical judgment errors or even false signals.

[0030] Sixth, the lack of historical state verification before sending mileage update trigger signals makes it difficult to eliminate the risk of false transmissions. Even with relatively sophisticated front-end filtering, triggering signal transmission solely based on "the current identified value being greater than the previous value" cannot eliminate the risk of transmissions triggered by pseudo-stable values ​​formed by the convergence of abnormally fluctuating data. The lack of reverse confirmation logic—whether the preceding historical data stably matches the previous scale value before the current value jumps—is the root cause of false signal transmissions in the existing system.

[0031] To solve the above problems, such as Figures 1-7 As shown, this application provides a mileage verification system based on a multi-stage deep learning model and a multi-dimensional false positive prevention mechanism, including: Double-layer nested adaptive search module: Divides the acquired scene image into several overlapping sub-regions and searches them, coarsely locates the mileage of the sub-regions, and persistently stores the current magnification value and sub-region coordinate parameters; Scene images are captured by a camera within a defined area. The industrial camera captures scene images within this area, including the vehicle, dashboard, meter, and background environment.

[0032] Multi-stage deep learning model cascaded inference architecture: sequentially performs screen type determination, mileage window detection, digit region purification and numerical recognition on the mileage device image, and outputs a numerical string; Multi-stack collaboration module: performs multi-dimensional time-domain filtering on the recognition result stream composed of numerical strings and outputs stable mileage values; Reverse history verification module: Performs reverse history verification on the mileage stability value, and allows the transmission of acquisition signals after successful verification.

[0033] This invention can be applied to various scenarios, such as mileage display on car dashboards, mileage display on treadmills, and mileage recognition on taxi meters. To better explain the specific mileage assumption method of this invention, the mileage device is set as a taximeter, and the application scenario of displaying mileage on a taximeter is used as an example for illustration. In the scenario of reading mileage on a taximeter, the vehicle is detected, the location of the taximeter inside the vehicle is located sequentially, the location of the mileage window on the taximeter is located, the digital sub-region of the mileage window is located, and finally, the mileage characters displayed in the digital sub-region are recognized.

[0034] In real-world verification scenarios, the installation position, size, and angle of taximeters vary significantly across different vehicle models, causing industrial cameras to fail to guarantee that the taximeter is within the optimal recognition area under default parameters. A dual-layer nested adaptive search module automatically activates upon initial or consecutive detection failures, enabling rapid adaptive locking of the taximeter.

[0035] The dual-layer nested adaptive search module consists of an outer controller and an inner searcher. The outer controller uses the optical magnification of the industrial camera as the outer control variable, while the inner searcher uses the region scan within the image plane as the inner search variable. The outer controller adjusts the camera magnification stepwise according to a preset step increment based on a magnification step control mechanism. At each magnification level, the inner searcher divides the image into several overlapping sub-regions according to an automatic region scanning mechanism using a predefined region scanning path, and calls the device detection and recognition model to identify each sub-region. Once the odometer is successfully detected in a sub-region, the search immediately terminates and locks and persistently stores the current magnification and region coordinate parameters. If no sub-regions are detected at the current magnification level, the outer controller adjusts the magnification to the next level, and the inner searcher restarts the region traversal. This dual-layer search strategy ensures that regardless of the initial size and position of the meter in the image, the system can complete automatic positioning within a limited number of steps without manual intervention, significantly improving the initial positioning success rate for different vehicle models in continuous assembly line scenarios.

[0036] Magnification step control mechanism: A set of discrete magnification levels Z is preset, Z = {z1, z2, …, z_n}, typically covering the entire range from wide-angle (low magnification) to telephoto (high magnification). The outer controller starts at the initial magnification z1. After each round of inner-layer region search, if no meter is detected, the magnification is adjusted to the next level z_{k+1}. The magnification increment Δz can be configured according to the actual operating conditions of the calibration line, typically increasing by a fixed percentage of magnification per step. The stepping process has a clear termination condition: if no meter is detected after traversing all preset levels, the system reports a positioning failure and waits for manual intervention or a new vehicle to enter. Automatic region scanning mechanism: At each magnification level, the current screen is divided into several sub-regions with a certain overlap ratio according to a predefined scanning path. The overlap ratio setting ensures that even if the meter crosses the boundary of an adjacent sub-region, at least one sub-region will completely contain its main body. The scanning path supports multiple strategy configurations: the center-priority spiral scanning strategy expands outward from the center of the screen, suitable for scenarios where the meter is likely located in the center of the field of view; the row and column grid scanning strategy traverses sequentially by row and column, suitable for scenarios where the meter's location is completely unknown.

[0037] The double-nested adaptive search module also has a preset confidence threshold. It determines whether a mileage device is identified by detecting the confidence level. When each sub-region is sent to the device detection and identification model for identification, once the detection confidence level output by the model exceeds the preset threshold, it is determined that there is a mileage device in that sub-region, and the search is terminated immediately.

[0038] Upon successful detection, the current magnification value and sub-region coordinate parameters are persistently stored. Subsequent real-time frame processing within the same verification process directly reuses these parameters, eliminating the need for repeated searches. When the system detects several consecutive frames of positioning failure (e.g., the meter completely disappears due to vehicle replacement), it automatically resets the persistent parameters and restarts the two-layer search process. This mechanism is particularly crucial in batch continuous verification scenarios: upon the entry of each new vehicle, the system can automatically switch positioning within seconds, eliminating the need for manual adjustment of camera parameters and significantly improving verification efficiency and automation.

[0039] The multi-stage deep learning model cascaded inference architecture includes a cascaded mileage device localization and screen type determination module, a mileage window detection and inverse perspective mapping module, a digital region purification module, and a numerical classification and recognition module. The mileage device localization and screen type determination module, the mileage window detection and inverse perspective mapping module, the digital region purification module, and the numerical classification and recognition module are set up independently. Each module is undertaken by a special model or algorithm module, and the output of the previous module strictly constrains the search space of the next module.

[0040] The mileage calibration system also includes a confidence and non-maximum suppression joint screening module, a two-layer image geometric correction module, and a continuous frame stable queue verification and frame-to-frame position comparison module. The confidence and non-maximum suppression joint screening module has a confidence and non-maximum suppression joint screening mechanism. This mechanism performs hard filtering based on the confidence threshold to eliminate low-quality candidates, and then performs non-maximum suppression screening on the filtered candidate box set to retain the single candidate box with the highest confidence, while all others are eliminated. Finally, only one optimal positioning result is retained, eliminating multi-box noise and overlapping misselection from the model output level. The two-layer image geometric correction module has a two-layer image geometric correction model. This model performs perspective transformation to straighten the tilted mileage device to a standard rectangular view and outputs a calibrated mileage device image. The continuous frame stable queue verification and frame-to-frame position comparison module has a continuous frame stable queue verification and frame-to-frame position comparison mechanism. Based on this mechanism, the mileage window of continuous frames is verified to eliminate jitter frame interference and obtain a stable mileage window.

[0041] The dual-layer image geometric correction model employs a two-layer distortion correction approach, performing distortion correction on the mileage value image at two levels: the first layer performs perspective transformation based on the four vertices coordinates of the mileage device's rotation frame, straightening the tilted mileage device to a standard rectangular view; the second layer performs affine rotation correction based on the rotation angle of the mileage window in the original video frame, eliminating local angular deviations. This dual-layer model performs two independent levels of correction, complementing each other to ensure maximum restoration before character shape recognition, while fundamentally eliminating the interference of character deformation on subsequent mileage value recognition. The distortion correction algorithm of the dual-layer image geometric correction model directly affects the imaging quality of the mileage values, thereby significantly improving the accuracy of mileage value recognition.

[0042] The mileage device positioning and screen type determination module has a device detection and recognition model: based on the device detection and recognition model, the position of the mileage device in the image is output in the form of a rotating box. This position is filtered by the confidence and non-maximum suppression joint filtering module and corrected by the dual-layer image geometric correction module. At the same time, the device detection and recognition model identifies the screen type of the mileage device and assigns a classification label, which is persisted as currentScreenType for all subsequent stages to share and call. Compared to traditional horizontal frame detection, rotating frame detection can more accurately detect the tilted boundary of the pricing device caused by the installation angle deviation, reducing the introduction of background areas.

[0043] The joint screening mechanism of confidence level and nonmaximum suppression is as follows: All candidate bounding boxes output by the device detection and recognition model are first subjected to hard filtering based on a confidence threshold τ_conf (default 0.70) to remove low-quality candidates. Then, Non-Maximum Suppression (NMS) is performed on the filtered candidate box set to calculate the Intersection over Union (IoU) of any two boxes. For overlapping box groups where the IoU exceeds a preset threshold τ_iou, the single candidate box with the highest confidence is retained, and all others are discarded. Finally, only one optimal localization result is retained, eliminating multi-box noise and overlapping misselections at the model output level.

[0044] The dual-layer image geometric correction model performs perspective transformation to straighten the tilted odometer device to a standard rectangular view. The method for outputting the corrected odometer device image is as follows: The coordinates of the four vertices of the rotating frame, sorted by top left, top right, bottom right, and bottom left, are set as {P1, P2, P3, P4}. The four vertices of the target rectangle are {Q1, Q2, Q3, Q4}. A 3×3 perspective transformation matrix H is solved using Direct Linear Transformation (DLT) to map the tilted vertices of the odometer device in the image to the vertices of a standard rectangle. Affine spatial resampling (WarpPerspective) is then performed on the original video frames using H, outputting a geometrically corrected standard rectangular image of the odometer device. This eliminates overall perspective trapezoidal distortion, restoring the screen boundary of the odometer device to an orthogonal rectangular shape.

[0045] This application simultaneously classifies and determines the screen type during the overall mileage device detection process. This determination result serves as a global type marker, `currentScreenType`, which is shared and referenced in all subsequent recognition stages. Screen types include LED and LCD classifications. The LED classification has a dedicated LED model, and the LCD classification has a dedicated LCD model. The LED model is optimized for the high-brightness pixel matrix features of LED screens, and its feature extractor focuses on capturing the spatial distribution patterns of isolated luminous pixels. The LCD model is optimized for the continuous transmission stroke features of liquid crystal display screens, and its feature extractor focuses on capturing edge sharpness and contrast distribution. Through type-based classification, the model feature extraction strategy at each stage is strictly matched to the current screen pixel characteristics, eliminating misidentification caused by cross-type mixing.

[0046] The mileage window detection and inverse perspective mapping module includes a mileage window detection module and an inverse perspective mapping module. The mileage window detection module has a mileage window detection model. The mileage window detection model precisely locates and corrects the position of the mileage window in the mileage device image. The mileage window position is filtered by the confidence and non-maximum suppression joint screening module, inverse perspective mapping by the inverse perspective mapping module, and correction and cropping by the double-layer image geometric correction module, outputting a clean mileage window image.

[0047] The mileage window detection module calls the mileage window detection model on the mileage device rectangle that has undergone perspective correction, and outputs a mileage window rotation box representing the mileage value sub-region. The mileage window detection model selects either the LED-specific model or the LCD-specific model based on currentScreenType. The mileage window rotation frame is uniquely determined by the four vertices P_(taximeter).

[0048] The coordinates of the four vertices of the mileage window rotation frame are extracted by a joint filtering module of confidence and non-maximum suppression. The four vertices of the odometry window bounding box output by the odometry window detection model are rotated and first subjected to hard filtering based on a confidence threshold τ_conf (default 0.70) to remove low-quality candidates. Then, non-maximum suppression (NMS) is performed on the filtered candidate box set to calculate the intersection-over-union (IoU) ratio of any two boxes. For overlapping box groups where the IoU exceeds a preset threshold τ_iou, the single candidate box with the highest confidence is retained, and all others are discarded. Ultimately, only one optimal localization result is retained at each stage, eliminating multi-box noise and overlapping misselections at the model output level.

[0049] The inverse perspective mapping module performs inverse perspective mapping on the selected four vertex coordinates to obtain the coordinates of the mileage window in the original frame; To enable subsequent real-time frame processing to directly manipulate the original video coordinates, inverse perspective coordinate mapping is performed on the four vertices of the rotating frame. Let H be the perspective transformation matrix of the odometer positioning and screen type determination module; then its inverse matrix H... -1 Map the four vertices P_(taximeter) of the mileage window within the odometer image back to the original video frame coordinates: P_(original) = H -1 · P_(taximeter) The mapping result P_(original) is persistently stored as originalRect. Subsequent frames are directly cropped around this region, avoiding the computational overhead of repeatedly performing the overall search in each frame. At the same time, the processing range of real-time frames is precisely limited to the mileage window, eliminating the possibility of mistaking other numerical fields such as amount and time as mileage areas from an architectural perspective.

[0050] The dual-layer image geometric correction model performs affine rotation correction and cropping on the dual-layer image input to the mileage window, eliminating local angular deviations and obtaining a clean mileage window image.

[0051] After mapping `originalRect` back to the original video frame, the rotated bounding box typically still retains a slight rotation angle θ in the real-time frame. A two-dimensional rotation affine matrix `M_rot` is constructed based on the center of the rotated bounding box, C = (c_x, c_y). Rotation correction (WarpAffine) is performed on the odometer window region using `M_rot` to eliminate local rotation deviations and restore the horizontal alignment of the digits. After correction, the corrected image is finely cropped according to an expanded scale to remove the black borders introduced by rotation compensation, obtaining a clean odometer window image for subsequent recognition.

[0052] The continuous frame stability queue verification and frame position comparison mechanism is to maintain a continuous frame stability detection queue in the system. By comparing the positions of the preceding and following frames and making a joint judgment based on confidence, the subsequent character recognition thread is only started after a window that is valid and has a stable position is detected in multiple consecutive frames. Interference from jittery frames is eliminated through time-domain cumulative verification.

[0053] The continuous frame stability queue verification and frame position comparison mechanism verifies the mileage window of continuous frames as follows: The system maintains a historical detection queue Q_pos for the center point of the mileage window in memory, with a capacity of N_stable (default N_stable = 5). After each frame successfully passes the first layer of filtering, the system not only pushes the center point of the mileage window into Q_pos, but also performs a position comparison between the previous and next frames: calculating the Euclidean distance between the detection position of the current frame and the detection position of the previous frame. If the distance exceeds the preset position jump tolerance threshold δ_pos, the detection result of the current frame is marked as a suspected jitter frame and is not included in the stable accumulation. Only when Q_pos continuously accumulates N_stable valid and positionally consistent detection results does the system consider the mileage window position to be stable and allow the subsequent numerical and character recognition threads to start. If no valid window is detected for a consecutive frame loss threshold N_missing (default N_missing = 8), the current positioning is determined to be invalid, the stable state is cleared, the numerical recognition thread stops, and waits for re-accumulation of stable confirmation.

[0054] This application employs a dual verification mechanism of "comparison of positions between consecutive frames + continuous accumulation confirmation". Based on confidence screening, it further introduces spatial position continuity constraints to ensure that character recognition is only performed when the physical position of the mileage window is truly stable, thus completely eliminating the situation where recognition is triggered by transient positioning results caused by engine vibration or rapid translation.

[0055] After successfully completing the entire detection process of the mileage device localization and screen type determination module, mileage window detection and inverse perspective mapping module, the position parameters of the mileage window in the original video frame coordinate system (including the coordinates of the four vertices of the rotating box, rotation angle, and rectangle width and height) are pre-acquired and persistently stored in `originalRect`. Subsequent real-time processing of each frame no longer repeats the global mileage device search and mileage window detection. Instead, it directly performs local region cropping and correction based on `originalRect` in the original frame, only executing the digital region purification module's digital region purification and the numerical recognition module's numerical classification. This compresses the real-time frame processing latency from the complete inference of four modules to the local processing of the latter two modules, achieving a balance between recognition accuracy and computational efficiency.

[0056] The mileage location pre-acquisition mechanism also combines continuous frame location drift monitoring and recalibration mechanism. When the deviation between the pre-acquisition location and the actual intra-frame mileage window is detected to exceed the allowable threshold, the location recalibration process is triggered to ensure the continuous validity of the location parameters under long-term operation.

[0057] Specifically, the continuous frame position drift monitoring and recalibration mechanism includes: The pre-acquired position parameters may accumulate deviations during long-term operation due to factors such as vehicle vibration and camera micro-displacement. During real-time processing, the confidence and stability indicators of the recognition results are continuously monitored. Position recalibration is automatically triggered when any of the following abnormal states are detected: the OCR recognition confidence of several consecutive frames is lower than a preset threshold; the recognition results of consecutive frames show abnormally frequent jumps; or the mileage window purification stage fails to extract effective digital connected components. The recalibration process re-executes the complete detection process of the mileage device positioning and screen type determination module, the mileage window detection and inverse perspective mapping module, and updates the `originalRect` parameter to ensure the continued reliability of subsequent recognition.

[0058] The digital region purification module is equipped with a digital region purification model. The mileage window image is input into the digital region purification model, and digital region morphological purification is performed to extract the pure digital display area. The digital region purification model incorporates morphological operations to extract clean digital display areas from mileage window images. The specific method is as follows: Calculate the mean grayscale value μ and standard deviation σ of the region, and adaptively set the highlight area extraction threshold range according to the screen type: Pixel_(extract) = {p | p > μ + kσ} Where p is the grayscale value of any pixel in the image, μ is the mean of the grayscale values ​​of all pixels in the region, σ is the standard deviation of the grayscale values ​​of all pixels in the region, {p |condition} is a set notation, representing "the set of all p that satisfy a certain condition", Pixel_(extract) is the set of extracted "highlight pixels", and k is a coefficient adjusted according to the LED / LCD type. Due to the self-emissive nature of LED screens, the brightness of digital pixels is much higher than that of the background, so k is taken as a larger value to accurately extract the luminous strokes; due to the backlight transmission characteristics of LCD screens, the contrast between the numbers and the background is relatively low, so k is taken as a smaller value to avoid over-filtering out effective pixels. After extracting the binary image of the highlight pixels, morphological closing operations are performed with structuring elements of appropriate size to forcibly bridge the isolated stroke segments of broken digits, forming complete connected components. Then, through contour extraction and minimum bounding rectangle fitting, a clean digit display area is obtained, removing border backlight and non-character area noise, which is used as the input of the final recognition model.

[0059] The numerical classification and recognition module is equipped with a numerical recognition model, which performs character classification on the pure number display area based on the numerical recognition model.

[0060] After refining the digital area, the system further divides the data based on currentScreenType, calls the final numerical classification model of the LED or LCD special project, outputs the character category ID, and maps it to a mileage numerical string with a minimum precision unit of 0.1, thus completing single-frame recognition.

[0061] The advantages of the four-module cascaded inference architecture of the multi-stage deep learning model are as follows: with each stage advancement, the search space shrinks by at least an order of magnitude, and the probability of false detection decreases exponentially with each stage. Simultaneously, the design of class labels being determined only once and shared throughout ensures that both types of screens can obtain the optimal recognition conditions for their respective categories. The multi-model, staged invocation architecture also allows for independent optimization of model structure and training data for different stages, avoiding performance compromises between multiple tasks for a single end-to-end model and improving the overall recognition efficiency and accuracy of the system. The output of each stage is strictly used as the restricted input for the next stage, eliminating false detections caused by large-scale blind searches.

[0062] The numerical recognition model outputs a numerical string per frame, and these strings constitute a recognition result stream in the time dimension. The multi-stack collaborative module performs multi-dimensional temporal filtering on this recognition result stream. The multi-stack collaborative module has a three-queue collaborative delayed confirmation algorithm to filter isolated noise and transient misidentifications in the recognition result stream, ensuring that the finally confirmed stable mileage value has statistical reliability.

[0063] The string output from each frame's numerical recognition is processed by the three data stacks (queues) of the multi-stack collaboration module: The correct data stack includes the identification results whose differences from the current stable value are within the tolerance range; The difference in the normal jump stack is approximately equal to the recognition result of the standard step size; The exception jump stack includes other unexpected results.

[0064] This application does not rely on single-frame conclusions, but performs delayed confirmation through the cooperation of multiple stacks: only when a sufficient number of samples have accumulated in the normal transition stack is the current mileage value update confirmed, thus blocking aliasing noise frames and isolated misidentified frames generated by the transient switching of the digital tube from the confirmation gate.

[0065] The multi-stack collaborative module performs multi-dimensional temporal domain filtering on the recognition result stream using the following method: Initialization Phase: After system startup, the recognition results of the first N_init frames are sent to the initialization sampling stack Q_init, and do not participate in subsequent logic. Once Q_init is full, the system counts the most frequently occurring values. If the frequency of this value reaches a set proportion of the Q_init capacity, it is set as the initial stable benchmark value V_stable, and the system enters normal operation. The existence of the initialization phase ensures that the system will not make any misjudgments before establishing a reliable benchmark. This process is equivalent to the system performing a "cold start learning" on the current taximeter display state to obtain a reliable initial reference point.

[0066] The multi-stack real-time classification logic of the multi-stack collaboration module: The system maintains three sliding time window data stacks with limited capacity. Each frame of new identification value v_new is assigned to the corresponding stack based on its difference from the current stable value V_stable. The correct data stack Q_correct: records the identification results that satisfy |v_new - V_stable| ≤ ε_stab, which physically means that the current frame confirms that the current stable value is still valid.

[0067] Normal jump stack Q_jump: Records the recognition results that satisfy |v_new - V_stable - Δ_step| ≤ ε_jump, which physically means that the mileage value displayed in the current frame has progressed normally by one step.

[0068] Abnormal jump stack Q_abnormal: Includes identification results that do not meet the above two conditions. The physical meaning is that it cannot be classified into the normal mode and is suspected to be noise or transient misidentification.

[0069] Where Δ_step is the preset normal jump step size (usually 0.1 km), and ε_stab and ε_jump are the stability tolerance and jump tolerance, respectively. All three stacks adopt a first-in-first-out (FIFO) strategy, automatically discarding the oldest data when the capacity is exceeded, ensuring that the stack contents always reflect the recognition status within the most recent time window.

[0070] Multi-stack collaboration decision-making mechanism of multi-stack collaboration module: This application does not trust the conclusion of a single frame, but performs delayed confirmation through multi-stack collaboration: Prioritize checking Q_jump: If the number of samples accumulated in Q_jump within the most recent time window exceeds the confirmation threshold N_confirm, the system confirms the mileage value by jumping from V_stable to V_stable + Δ_step, overwriting V_stable, clearing the three stacks, and proceeding to the subsequent signal transmission verification stage. The existence of the confirmation threshold N_confirm ensures that the conclusion of the mileage value update is supported by multi-frame statistical consensus, rather than a single-frame judgment.

[0071] If Q_jump does not reach the acknowledgment threshold, Q_correct is checked. If it continues to accumulate, indicating that the current frame is still V_stable, the system remains stable and does not update. This mechanism ensures that the system will not generate false updates due to occasional noisy frames during periods when no actual odometer jump occurs.

[0072] When anomalous samples accumulate in Q_abnormal and form a statistically dominant pattern, and this pattern itself has internal consistency, the system allows adaptive recovery to be performed to prevent systematic and persistent misclassification caused by device parameter drift or abnormal taximeter display. Adaptive recovery is essentially a dynamic correction of the three-stack classification benchmark, enabling the system to recover its working capability under abnormal operating conditions.

[0073] The multi-stack collaborative module of this application performs multiple statistical verifications on the single-frame recognition conclusion in the time domain. The isolated digital tube switching transient noise, because it only occurs in one or two frames, can never accumulate to the confirmation threshold in Q_jump, and is therefore systematically intercepted, never triggering a stable value update, thus achieving the prevention of false detection.

[0074] Before the serial port acquisition signal is triggered by the update of the stable mileage value, the reverse history verification mechanism based on the reverse history verification module performs reverse history verification on the stable mileage value: it retrieves several recent historical records from the correct data stack, requiring that their values ​​are all equal to "the current new value minus a standard step size", that is, it verifies whether the previous history of the current value is consistent with the logic of this jump. Only when the reverse verification passes and the target mileage point has not sent a signal is the serial port signal allowed to be sent, completely eliminating the possibility of false transmissions caused by the unexpected convergence of abnormal data.

[0075] Specifically, the reverse historical verification mechanism is used to verify the stability value of mileage. Even if the stable value is updated after being confirmed by multiple stacks, the transmission of the serial port acquisition signal still needs to pass a final independent verification to eliminate the risk of false transmission caused by the unexpected convergence of abnormal data and the falsification of stable value updates.

[0076] Reverse history verification logic: Assume the current stable value confirmed by the multi-stack is V_new. The system retrieves the number of most recent reverse history verification samples / the number of valid historical records M_valid (M_valid indicates how many recent historical identification records need to be retrieved from the correct data queue Q_correct for verification before sending the signal) (default M_valid= 3, configurable) historical records {h1, h2, …, h_M_valid} from Q_correct and performs reverse verification: ∀i ∈ {1, …, M_valid} : h_i = V_new - Δ_step i represents a specific historical record within the M_valid historical records; This verifies that prior to this jump, the most recent frames of the recognition history were consistently equal to "the new value minus one step size". In physical semantics, this condition describes that the odometer indeed remained stably at the previous scale before reaching the current scale, without any abnormal jumps that skipped a step size in the middle.

[0077] Reverse history verification is an orthogonal complement to the forward confirmation of the three-stack stack: the three-stack stack verifies "whether there are enough frames to support the new value" in the forward time direction, while the reverse verification verifies "whether the state before the jump is reasonable" in the reverse time direction. Together, they constitute a complete temporal causal chain verification, logically blocking the extreme case where abnormal data happens to accidentally reach the confirmation threshold in the three stacks.

[0078] The acquisition signal will only be allowed to be sent if the reverse history verification is successful and all of the following conditions are met simultaneously: (1) The new stable value V_new is completely matched with the preset mileage target point of the verification task.

[0079] (2) The target point has not yet been recorded as sent (to prevent duplicate sending).

[0080] (3) The serial port connection has been established and the automatic sending function has been enabled.

[0081] After all the above conditions are met, the system will repeatedly send N_send (the number of times the acquisition signal is repeatedly sent, default 2) acquisition signals according to the configuration parameters, with an interval of T_interval (the acquisition signal sending interval represents the time interval between two consecutive acquisition signals, default 10 milliseconds) to improve the reliability of external device reception.

[0082] Additional termination condition: When the system detects that the current stable value has remained continuously within V_stop ± ε_stop (V_stop: the target value for termination of verification / the target mileage value for stopping judgment, representing the target mileage value used by the system to determine when the current verification process should end; ε_stop stopping judgment tolerance - representing the allowable error range when judging whether the current stable value is close to V_stop, the formula represents the allowable error range when judging whether the current stable value is close to V_stop, usually near the verification endpoint mileage) for more than T_stop (the stopping judgment duration threshold, representing how long the current stable value needs to remain continuously within the range of V_stop ± ε_stop before the system considers it safe to send a stop signal) milliseconds, the system actively sends a stop signal to terminate the current verification process. This design ensures that the verification end judgment is also constrained by the time-domain stability condition, rather than being determined by a single frame value.

[0083] The implementation process of the multi-stage deep learning model cascaded inference architecture is as follows: The device detection and recognition model, based on a rotating target detection network architecture, outputs the precise location of the mileage device in the image in the form of a rotating bounding box, and simultaneously outputs a screen type classification label (LED or LCD); the location of the mileage device in the image is filtered by a confidence and non-maximum suppression joint filtering module, and the coordinates of the four vertices of the rotating bounding box are extracted; the coordinates of the four vertices of the rotating bounding box are input into a two-layer image geometric correction module, which performs perspective transformation through the two-layer image geometric correction model to straighten the tilted mileage device to a standard rectangular view, and outputs a corrected mileage device image; the mileage window detection module calls the mileage window detection model on the corrected mileage device image and outputs a mileage window rotating bounding box representing the mileage value sub-region. The four vertices of the mileage window rotating bounding box are filtered by a confidence and non-maximum suppression joint filtering module. The large value suppression joint screening module filters and extracts the coordinates of the four vertices of the mileage window rotation box. The inverse perspective mapping module performs inverse perspective mapping on the selected four vertices to obtain the coordinates of the mileage window in the original frame. The dual-layer image geometric correction model performs affine rotation correction and cropping on the dual-layer image input to the mileage window to eliminate local angle deviations and obtain a clean mileage window image. The clean mileage window image is input into the digital region purification model to perform digital region morphological purification and extract the clean digital display area. Based on the numerical recognition model, character classification is performed on the clean digital display area. The multi-stack collaboration module performs multi-dimensional temporal domain filtering on the recognition result stream composed of numerical strings and outputs a stable mileage value. The reverse history verification module performs reverse history verification on the stable mileage value. After the verification is passed, the acquisition signal is allowed to be sent.

[0084] Through the synergistic effect of the above steps, the magnification step control and automatic region search of the double-nested adaptive search module achieve adaptive coverage of diverse installation scenarios from the positioning level. The multi-stage deep learning model cascaded inference architecture limits the false detection space from the architecture level. The double-layer geometric correction module eliminates recognition distortion from the perspective of image quality. The mileage position pre-acquisition achieves accurate and efficient recognition from the perspective of model matching and efficiency. The continuous frame stable queue verification and the position comparison mechanism between previous and subsequent frames, along with the multi-stack collaborative module filtering, remove noise penetration from the time domain. The reverse history verification mechanism verifies and plugs the last loophole of signal mistransmission from the perspective of logical closure. Thus, a full-link, multi-level, highly reliable mileage verification accurate recognition and false detection prevention system is constructed.

[0085] Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort should fall within the scope of protection of the present invention.

Claims

1. A mileage verification system based on a multi-stage deep learning model and a multi-dimensional false positive prevention mechanism, characterized in that: include Double-layer nested adaptive search module: Divides the acquired scene image into several overlapping sub-regions and searches them, coarsely locates the mileage of the sub-regions, and persistently stores the current magnification value and sub-region coordinate parameters; Multi-stage deep learning model cascaded inference architecture: sequentially performs screen type determination, mileage window detection, digit region purification and numerical recognition on the mileage device image, and outputs a numerical string; Multi-stack collaboration module: performs multi-dimensional time-domain filtering on the recognition result stream composed of numerical strings and outputs stable mileage values; Reverse history verification module: Performs reverse history verification on the mileage stability value, and allows the transmission of acquisition signals after successful verification.

2. The mileage verification system based on a multi-stage deep learning model and a multi-dimensional false detection prevention mechanism according to claim 1, characterized in that: The dual-layer nested adaptive search module consists of an outer layer controller and an inner layer searcher. The outer layer controller uses the optical magnification of the industrial camera as the outer layer control variable, while the inner layer searcher uses the region scanning within the image plane as the inner layer search variable. The outer layer controller adjusts the camera magnification step by step according to the magnification step control mechanism with a preset step increment. At each magnification level, the inner layer searcher divides the image into several overlapping sub-regions according to the automatic region scanning mechanism with a predefined region scanning path, and calls the device detection and recognition model to identify each sub-region.

3. The mileage verification system based on a multi-stage deep learning model and a multi-dimensional false detection prevention mechanism according to claim 1, characterized in that: The multi-stage deep learning model cascaded inference architecture includes a mileage device localization and screen type determination module, a mileage window detection and inverse perspective mapping module, a digital region purification module, and a numerical classification and recognition module connected in series. The mileage device localization and screen type determination module, the mileage window detection and inverse perspective mapping module, the digital region purification module, and the numerical classification and recognition module are set independently of each other, and the output of the previous module strictly constrains the search space of the next module.

4. The mileage verification system based on a multi-stage deep learning model and a multi-dimensional false detection prevention mechanism according to claim 1, characterized in that: The mileage verification system also includes a confidence and non-maximum suppression joint screening module. The confidence and non-maximum suppression joint screening module has a confidence and non-maximum suppression joint screening mechanism. The confidence and non-maximum suppression joint screening mechanism performs hard filtering according to the confidence threshold to eliminate low-quality candidates, and performs non-maximum suppression screening on the filtered candidate box set to retain the single candidate box with the highest confidence and eliminate all others. Finally, only one optimal localization result is retained, eliminating multi-box noise and overlapping misselection from the model output level.

5. The mileage verification system based on a multi-stage deep learning model and a multi-dimensional false detection prevention mechanism according to claim 1, characterized in that: The mileage verification system also includes a dual-layer image geometric correction module. The dual-layer image geometric correction module is equipped with a dual-layer image geometric correction model. The dual-layer image geometric correction model performs perspective transformation to straighten the tilted mileage device as a whole to a standard rectangular view and outputs a corrected mileage device image.

6. The mileage verification system based on a multi-stage deep learning model and a multi-dimensional false detection prevention mechanism according to claim 1, characterized in that: The mileage verification system also includes a continuous frame stable queue verification and front-and-back frame position comparison module. The continuous frame stable queue verification and front-and-back frame position comparison module is equipped with a continuous frame stable queue verification and front-and-back frame position comparison mechanism. Based on the continuous frame stable queue verification and front-and-back frame position comparison mechanism, the mileage window of the continuous frame is verified to eliminate jitter frame interference and obtain a stable mileage window.

7. The mileage verification system based on a multi-stage deep learning model and a multi-dimensional false detection prevention mechanism according to claim 5, characterized in that: The dual-layer image geometric correction model is set as a dual-layer distortion correction. The dual-layer distortion correction performs two levels of distortion correction on the mileage numerical image. The first layer performs perspective transformation based on the four vertex coordinates of the mileage device rotation frame to straighten the tilted mileage device as a whole to a standard rectangular view. The second layer performs affine rotation correction based on the rotation angle of the mileage window in the original video frame to eliminate local angle deviations.

8. The mileage verification system based on a multi-stage deep learning model and a multi-dimensional false detection prevention mechanism according to claim 3, characterized in that: The mileage window detection and inverse perspective mapping module includes a mileage window detection module and an inverse perspective mapping module. The mileage window detection module is equipped with a mileage window detection model. The mileage window detection model precisely locates and corrects the position of the mileage window in the mileage device image. The mileage window position is filtered by a confidence and non-maximum suppression joint screening module, inversely mapped by the inverse perspective mapping module, and corrected and cropped by a dual-layer image geometric correction module to output a clean mileage window image.

9. The mileage verification system based on a multi-stage deep learning model and a multi-dimensional false detection prevention mechanism according to claim 3, characterized in that: The mileage window detection and inverse perspective mapping module pre-acquires and persistently stores the position parameters of the mileage window in the original video frame coordinate system based on the mileage position pre-acquisition mechanism.

10. The mileage verification system based on a multi-stage deep learning model and a multi-dimensional false detection prevention mechanism according to claim 1, characterized in that: The multi-stack collaboration module's three data stacks perform string classification processing on the numerical recognition output of each frame. The three data stacks include... The correct data stack includes the identification results whose differences from the current stable value are within the tolerance range; The difference in the normal jump stack is approximately equal to the recognition result of the standard step size; The exception jump stack includes other unexpected results.

Images