Financial bill intelligent identification and archiving method of embedded hardware terminal

Abstract

The invention relates to the technical field of image processing and data acquisition, and discloses a financial bill intelligent identification and archiving method of an embedded hardware terminal, and the method comprises the steps: firstly obtaining a global low-resolution image, and generating an acquisition instruction queue containing heterogeneous ISP parameters; then controlling the image sensor to execute hardware-level asynchronous serialization acquisition by utilizing vertical blanking interval configuration to obtain heterogeneous local image slices; and eliminating view field displacement between the slice and the global image by using gradient domain registration, and executing logic consistency verification on the identified and extracted service data. And if the verification is passed, generating a layered archived file, otherwise, based on the image histogram, generating a correction parameter and triggering hardware feedback recollection. According to the invention, through a heterogeneous acquisition strategy and a business logic closed loop, the problems of high delay of high-resolution acquisition of a mobile terminal and difficult alignment of handheld shooting are effectively solved, and the bill identification accuracy and filing efficiency are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of image processing and data acquisition technology, specifically to a method for intelligent recognition and archiving of financial documents using embedded hardware terminals. Background Technology

[0002] With the development of mobile internet and financial digitization, the on-site collection and real-time entry of financial documents using handheld PDAs, high-speed scanners, or mobile smart terminals has become the industry norm. To meet the clarity requirements of fine text and anti-counterfeiting textures on documents, terminal devices are typically equipped with high-pixel image sensors. However, in environments with limited embedded hardware resources, the real-time throughput and processing of high-resolution images faces multiple technical challenges.

[0003] Existing technologies typically employ a full-frame, high-resolution acquisition mode, which involves a single high-pixel exposure and transmission of the entire document. Due to the massive amount of image data, this not only consumes significant bus bandwidth and memory resources, leading to a decrease in preview frame rate and increased processing latency, but also easily causes overheating and frequency throttling of the embedded processor. If the acquisition resolution is reduced to improve response speed, crucial and subtle features on the document, such as invoice codes and decimal points, often become blurred, thereby reducing the accuracy of optical character recognition.

[0004] Furthermore, financial documents possess complex physical and optical properties. For instance, red stamp ink typically has high reflectivity, while black dot-matrix or inkjet printing primarily exhibits light-absorbing characteristics. Under a single global automatic exposure strategy, it's often difficult to balance the image quality of both the stamp and the text, easily resulting in overexposed, washed-out stamps or underexposed, noisy text. Although existing high dynamic range (HDR) imaging technology addresses the light ratio issue by synthesizing multiple frames with different exposures, in handheld acquisition scenarios, even slight device jitter can cause pixel-level spatial displacement between frames, easily leading to ghosting or blurring during the synthesis process, affecting subsequent feature extraction.

[0005] On the other hand, most existing document recognition systems employ open-loop processing logic, meaning that image acquisition and subsequent business verification are unidirectionally isolated. When optical character recognition results show low confidence or business logic (such as amount calculation) imbalances, the system typically only reports an error at the software level or prompts the user to manually retake the image, failing to control the hardware acquisition parameters based on the specific error type. This lack of feedback mechanism leads to low acquisition success rates in complex lighting environments, frequent manual intervention, and difficulty in generating high-quality, hierarchical voucher files that meet the requirements for long-term electronic record storage. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides an intelligent recognition and archiving method for financial documents on embedded hardware terminals. This method solves the problems of high processing latency caused by high-resolution full-frame acquisition and image blurring and alignment difficulties caused by handheld shooting shakiness in the existing mobile financial document acquisition process.

[0007] The first aspect of this invention provides a method for intelligent recognition and archiving of financial documents in an embedded hardware terminal.

[0008] The method includes: acquiring a global low-resolution image and generating an acquisition command queue containing spatial coordinate parameters and heterogeneous ISP parameters based on the analysis results of the global low-resolution image; parsing the acquisition command queue, controlling the image sensor to perform hardware-level asynchronous serialized acquisition, and acquiring heterogeneous local image slices corresponding to the spatial coordinate parameters; performing cross-domain spatial registration between the heterogeneous local image slices and the global low-resolution image to eliminate the field-of-view displacement caused by asynchronous acquisition and updating the physical coordinates of the heterogeneous local image slices; performing optical character recognition on the registered heterogeneous local image slices and extracting business data, and performing logical consistency verification on the business data; if the verification passes, generating a hierarchical archive file containing the global low-resolution image, heterogeneous local image slices, and business data; if the verification fails, generating a re-acquisition command containing correction parameters and inserting it into the acquisition command queue to trigger feedback re-acquisition control.

[0009] In the step of generating the acquisition command queue, this invention employs a combination of physical image stabilization calculation and heterogeneous parameter configuration. The system calculates the theoretical physical bounding box of key fields on the sensor plane, reads the angular velocity data from the gyroscope to calculate the image stabilization expansion coefficient, and redundantly expands the theoretical physical bounding box to generate a safe acquisition bounding box as spatial coordinate parameters. Simultaneously, based on the physical attributes of the key fields, a preset lookup table is consulted to obtain the reference vector and attribute bias vector, and the exposure and color configuration for that key field is calculated as heterogeneous ISP parameters. The system encapsulates the spatial coordinate parameters and heterogeneous ISP parameters into command units and constructs the acquisition command queue according to the spatial scanning order. This method can configure corresponding imaging parameters for areas with different physical attributes in a document.

[0010] In controlling image sensor acquisition, this invention utilizes the vertical synchronization signal timing for register configuration. The system registers a hardware interrupt service routine to monitor the vertical synchronization signal of the image sensor. During the vertical blanking interval, it extracts instruction units from the acquisition instruction queue, writes spatial coordinate parameters into the windowing control register, and writes heterogeneous ISP parameters into the shadow register. Using a double-buffering mechanism, when the vertical blanking interval ends and the next frame cycle begins, the data in the shadow register is copied to the active register, making the heterogeneous ISP parameters effective in the current frame. The image sensor's row address decoder activates the pixel array between the start and end rows based on the spatial coordinate parameters in the windowing control register. The column-level analog-to-digital converter samples and quantizes the pixel charges from the start to the end columns, outputting heterogeneous local image slices in windowed readout mode.

[0011] To address the spatial displacement that may be introduced by asynchronous acquisition and handheld shooting, this invention employs a gradient domain registration method. A reference image patch is extracted from the global low-resolution image. The heterogeneous local image patch is downsampled to generate downsampled patches with consistent pixel density. The Sobel operator is used to calculate the gradient magnitude maps of the reference image patch and the downsampled patch respectively, eliminating the influence of brightness and color differences caused by heterogeneous ISP parameters. Sliding window matching is performed within the gradient magnitude domain, and the normalized cross-correlation coefficient is calculated to determine the matching position of the downsampled patch relative to the reference image patch. Subsequently, the system calculates the relative offset vector at low resolution based on the matching position. Combining this with the physical resolution scaling ratio of the image sensor between preview mode and windowed readout mode, the relative offset vector is mapped to a physical pixel displacement at high resolution. This physical pixel displacement is then used to correct the physical coordinates of the heterogeneous local image patch.

[0012] This invention adjusts hardware acquisition parameters through business logic verification feedback. During the verification phase, the system parses the text output from optical character recognition into structured data and fills it into the business data vector. It loads a logical verification rule set containing monetary arithmetic balance rules and checksum format rules, and performs logical verification operations on the business data vector. If the verification result is false or the recognition confidence level is lower than a preset value, the system triggers feedback re-acquisition control: it statistically analyzes the brightness histogram of the current heterogeneous local image slice, calculates the proportion of bright pixels and the average brightness; if the proportion of bright pixels exceeds the overexposure threshold, it calculates a correction coefficient to reduce the exposure time; if the average brightness is lower than the underexposure threshold, it calculates a correction coefficient to increase the exposure time or gain; based on the correction coefficients, it generates new heterogeneous ISP parameters, encapsulates them into a high-priority acquisition instruction unit, and inserts it into the head of the acquisition instruction queue.

[0013] Furthermore, this method includes specific image preprocessing and layered archiving steps. Before optical character recognition, if the physical attribute label of the heterogeneous local image slice indicates a seal, the red channel data of the image is extracted, and adaptive binarization is performed using the maximum inter-class variance method to separate the seal characters from the background. The final generated archive file adopts a layout document container format, and its logical structure includes: a global background layer storing the global low-resolution image; an original voucher layer storing the registered and digitally signed heterogeneous local image slices; and a structured semantic layer storing business data and descriptions of the location coordinates of each field. During rendering, the original voucher layer is overlaid on the global background layer.

[0014] A second aspect of the present invention provides an embedded hardware terminal.

[0015] The terminal includes an image sensor, a memory, and a processor. The image sensor supports windowed readout mode and dynamic parameter configuration; the memory stores computer programs; and the processor executes the computer programs to implement the method described in the first aspect above.

[0016] This invention provides a method for intelligent recognition and archiving of financial documents in an embedded hardware terminal. It has the following beneficial effects:

[0017] 1. This invention solves the data transmission bottleneck and processing latency issues caused by high-pixel full-frame acquisition on mobile devices by combining a global low-resolution preview with a local high-resolution slice heterogeneous acquisition mode. Utilizing the windowing readout and vertical blanking interval parameter configuration mechanism at the image sensor's underlying layer, high resolution and specific ISP parameters can be independently configured for key fields in the document. This significantly reduces the amount of invalid background data processed while ensuring the imaging clarity and recognition accuracy of core business fields.

[0018] 2. This invention overcomes the field-of-view displacement caused by handheld shooting shake and the differences in image brightness and color caused by heterogeneous ISP parameters during asynchronous serialization acquisition by using a cross-domain spatial registration algorithm based on gradient magnitude. By calculating the normalized cross-correlation coefficient in the gradient domain and combining it with the physical resolution scaling ratio for coordinate mapping, spatial deviations between image layers of different resolutions are eliminated, ensuring accurate overlay and layout restoration of the original voucher and background image in the final layered archive file.

[0019] 3. This invention employs a hardware feedback re-acquisition mechanism driven by business data logical consistency verification, which changes the traditional one-way "acquisition and recognition" process. When business data verification fails or the confidence level is insufficient, the system can automatically calculate the corrected exposure or gain parameters based on the image histogram and trigger a high-priority local re-acquisition command. This utilizes the logical constraints of the semantic layer to reverse-calibrate the acquisition quality of the physical layer, significantly improving the availability and automated processing efficiency of financial data under complex lighting conditions. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the embedded hardware terminal hardware architecture of the present invention;

[0021] Figure 2 This is an overall flowchart of the financial document intelligent recognition and archiving method of the present invention;

[0022] Figure 3 The flowchart for generating the acquisition instruction queue of this invention is as follows;

[0023] Figure 4 This is a schematic diagram of the hardware acquisition timing control of the vertical blanking interval (VBI) of the present invention;

[0024] Figure 5 This is a flowchart of the cross-domain spatial registration and coordinate correction process of the present invention;

[0025] Figure 6 This is a flowchart of the logic verification and feedback re-sampling control of the present invention;

[0026] Figure 7 This is a schematic diagram illustrating the global uniform exposure effect of the prior art of this invention;

[0027] Figure 8 This is a schematic diagram illustrating the heterogeneous local acquisition effect of the present invention;

[0028] Figure 9 This is a schematic diagram illustrating the gradient domain spatial registration principle of the present invention;

[0029] Figure 10 This is a comparison chart of the accuracy rates of key field recognition in this invention;

[0030] Figure 11 This is a comparison chart of the processing time for the entire process of this invention;

[0031] Figure 12 A comparison chart of the average volume of archived documents for invention.

[0032] Among them, 100 is an embedded hardware terminal; 110 is an image acquisition module; 111 is an optical lens; 112 is a CMOS sensor; 120 is an image signal processing module; 130 is a central processing module; 131 is a general-purpose computing core; 132 is a neural network acceleration core; 140 is a storage module; and 150 is a communication bus. Detailed Implementation

[0033] 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.

[0034] See attached document Figure 1 , Figure 1 This is a schematic diagram of the hardware architecture of an embedded hardware terminal according to an embodiment of the present invention. The intelligent recognition and archiving system for financial documents provided by the present invention runs on the embedded hardware terminal 100.

[0035] The embedded hardware terminal 100 includes an image acquisition module 110, an image signal processing module 120, a central processing module 130, and a storage module 140. Each module interacts with the other via a communication bus 150 and a control link for data exchange and command transmission. The image acquisition module 110 converts the optical signal from an external document into a digital image signal. The image acquisition module 110 integrates an optical lens 111 and a CMOS sensor 112. The CMOS sensor 112 supports pixel array windowing readout, responding to external register commands to activate only a specific rectangular area in the photosensitive array for photoelectric conversion and data reading, and outputting raw image data. The image signal processing module 120 is connected to the image acquisition module 110 and is used to preprocess the raw image data. The image signal processing module 120 includes an independent programmable register set, supporting dynamic reception and switching of image processing parameters within the inter-frame vertical blanking interval. These parameters include exposure time, analog gain, color correction matrix, and edge sharpening. The central processing module 130 is the core of the embedded hardware terminal 100's computation. The central processing module 130 integrates a general-purpose computing core 131 and a neural network acceleration core 132. The general-purpose computing core 131 performs system scheduling, logic verification, and archiving / packaging tasks. The neural network acceleration core 132 performs edge detection and document type classification operations. The storage module 140 is connected to the central processing module 130 and stores system firmware and a document semantic topology database. The document semantic topology database records key field definitions for different document types, including standard normalized coordinates and physical attribute descriptions. Regarding the data path, the image acquisition module 110 transmits image data to the image signal processing module 120 via a high-speed serial interface. The central processing module 130 connects to the image acquisition module 110 and the image signal processing module 120 via a serial control bus and sends control commands.

[0036] See attached document Figure 2 , Figure 2This is a flowchart of a method for intelligent recognition and archiving of financial documents according to an embodiment of the present invention. The present invention provides a method for intelligent recognition and archiving of financial documents using an embedded hardware terminal, comprising the following steps:

[0037] S100, control the image acquisition module 110 to output a video stream at a first resolution, and use the neural network acceleration core 132 to analyze the frame images in the video stream to determine the physical boundary and type of the ticket;

[0038] S200: Based on the ticket type, the corresponding semantic topology configuration is called from the storage module 140 to calculate the physical coordinates of each key field on the current sensor plane. Combined with the physical attributes of each key field and the anti-shake and wide-view coefficient, a collection instruction queue containing the coordinates of a specific photosensitive area and heterogeneous image processing parameters is generated.

[0039] S300, the central processing module 130 uses the inter-frame vertical blanking interval to write configuration parameters to the image acquisition module 110 and the image signal processing module 120 frame by frame according to the timing of the acquisition instruction queue, controls the CMOS sensor 112 to output only high-resolution raw data of a specific area, and generates a series of heterogeneous local image slices through the image signal processing module 120.

[0040] S400 uses a gradient domain registration algorithm to calculate the displacement vector of heterogeneous local image slices relative to the global background image, and corrects the spatial coordinates of heterogeneous local image slices based on the displacement vector to eliminate physical misalignment caused by hardware jitter.

[0041] S500 performs character recognition on heterogeneous local image slices after coordinate correction, executes business logic balance verification, and encapsulates the global low-resolution image, heterogeneous local image slices, and structured text data into a hierarchical archive file after the verification passes.

[0042] See attached document Figure 2 , Figure 2 A flowchart illustrating the intelligent identification and archiving method for financial documents is shown. Step S100 specifically executes the global low-resolution topology mapping and type anchoring process. This includes the following steps:

[0043] Step S101: Control the image acquisition module 110 to output a video stream at a first resolution. The central processing module 130 sends a control command to the image acquisition module 110 to configure the CMOS sensor 112 to enter a low-resolution preview mode. In this mode, the CMOS sensor 112 reads the photosensitive array data through pixel merging or skipping, and outputs video stream frame images with a resolution lower than the sensor's physical resolution limit. The specific value of the first resolution depends on the input layer dimension requirements of the neural network acceleration core 132 and the system bus bandwidth limit, and is preferably set to VGA (640×480) or 720p format. The selection of the first resolution is based on meeting the minimum sampling frequency requirement of Shannon's sampling theorem for large-scale features of the document (such as borders and title bars), while controlling the transmission bandwidth to less than 30% of the single-channel rate of the MIPICSI-2 interface, reserving bus bandwidth for subsequent high-concurrency command transmission.

[0044] Step S102: Perform edge detection and perspective transformation matrix calculation on the video stream frame images. Neural network acceleration core 132 or general computing core 131 processes the current frame image. Edge extraction is performed. For the edge extraction algorithm, this embodiment uses a lightweight edge segmentation model based on a fully convolutional neural network (FCN). This model employs an encoder-decoder structure, with the input being a downsampled grayscale image and the output being a binarized mask of the same size as the input. The encoder uses depthwise separable convolutional layers to extract features, while the decoder restores spatial resolution through upsampling. When training this model, an image set labeled with the four physical edges of the ticket is used as samples, and backpropagation training is performed using binary cross-entropy as the loss function. The system extracts the coordinates of the four vertices of the ticket in the image coordinate system based on the output mask and establishes a geometric projection model of the ticket plane. The physical meaning of the perspective transformation matrix lies in describing the projection mapping relationship between two planes. The system calculates the homography matrix, assuming the coordinates of a point on the standard normalized template plane are... ,in , representing the normalized proportional position relative to the width and height of the ticket itself; the corresponding physical coordinates on the current sensor imaging plane are . The two satisfy the following projection transformation relationship:

[0045] ;

[0046] in, This is the scaling factor in a homogeneous coordinate system, with no specific physical unit. The transformation matrix is ​​3×3. The system uses the least squares method or the RANSAC (Random Sample Consensus) algorithm to calculate the matrix elements based on the four detected vertex pairs. to The value of . When using the RANSAC algorithm, set the reprojection error threshold. (For example, 3 pixels), iteratively remove outliers in corner detection caused by uneven lighting or background interference until the proportion of inliers meets the preset confidence requirement (for example, 95%). This matrix H constitutes the geometric basis for mapping from the business semantic space to the physical acquisition space in subsequent steps.

[0047] Step S103: Determine the ticket type based on the frame image content. The system extracts the feature region of the ticket image after perspective correction, inputs it into a pre-set lightweight classification network model, and outputs a ticket type identifier. This classification network model can be built on the MobileNetV2 or ShuffleNet architecture and includes several depthwise separable convolutional layers, global average pooling layers, and fully connected layers. The model's input is a size-normalized (e.g., 224×224 pixels) feature region image, and the output is the corresponding... The system selects the category with the highest probability value that exceeds a preset classification threshold (e.g., 0.85) as the final classification result. The model is trained using supervised learning. The dataset contains various invoice samples collected under different lighting conditions, along with their corresponding category labels. The network weight parameters are updated using stochastic gradient descent or the Adam optimizer. Invoice types include, but are not limited to, VAT invoices, train tickets, taxi invoices, and bank receipts.

[0048] Step S104: Load semantic topology configuration and field attribute descriptions. The central processing module 130 then... Retrieve the corresponding semantic topology set from the database of storage module 140. This collection contains several key field definitions. The formal description is as follows:

[0049] ;

[0050] Each key field It contains two parts of data: spatial distribution information With physical attribute tags Spatial distribution information This records the normalized rectangular range of this field in the standard invoice template. ,in The normalized coordinates of the top-left corner of the rectangular area relative to the origin of the document. This represents the normalized width and height of the rectangular region.

[0051] Physical attribute tags This describes the optical and material properties of this field region, used to guide the heterogeneous configuration of subsequent ISP parameters. Physical Property Label The specific values ​​include:

[0052] Spectral-sensitive properties: such as red ink or blue carbon paper marks, indicate that the main information in that area is carried in a specific color channel;

[0053] Reflective properties: such as anti-counterfeiting hot stamping and reflective coating, indicating that the area is prone to specular reflection under high exposure;

[0054] Contrast attributes: such as the faded text on thermal paper or the dot matrix of dot-matrix printing, indicate that the histogram distribution in that area is narrow or the signal-to-noise ratio is low;

[0055] Frequency domain attributes: such as QR codes on electronic screens, indicating that there is a specific frequency of refresh flickering in that area.

[0056] Through the above steps, the system completes the parsing of the physical form and business semantics of the invoice at a low resolution level, providing the necessary geometric parameters and attribute basis for the subsequent generation of high-resolution heterogeneous acquisition instructions.

[0057] See attached document Figure 3 , Figure 3 A detailed flowchart illustrating the heterogeneous acquisition strategy construction and anti-shake view expansion process is shown. Step S200 specifically executes the process of generating hardware execution instructions based on the document semantic topology. This includes the following steps:

[0058] Step S201: Calculate the theoretical physical coordinates of the key fields on the current sensor plane. General computing core 131 reads the perspective transformation matrix generated in step S100. and semantic topology set For each key field in the collection The system extracts its standard normalized coordinates. ,in The coordinates of the center point, To normalize the width and height, the system uses the forward projection principle of the homography matrix to map the normalized coordinates to the pixel coordinate system of the current CMOS sensor 112, thus obtaining the theoretical physical bounding box. Here Represents the center coordinates in pixels. These represent the width and height after projection. These theoretical coordinates characterize the geometric projection position of the target semantic region onto the imaging plane under ideal static conditions.

[0059] Step S202: Generate a safe acquisition bounding box with image stabilization margin. Since embedded hardware terminals are mostly handheld, high-frequency micro-vibrations can cause instantaneous field-of-view shifts during imaging; simultaneously, the CMOS sensor 112 uses a line-by-line exposure mechanism, and rapid movement can lead to geometric skew. To compensate for these errors at the physical level, the system introduces a field-of-view expansion factor. Redundant expansion of the theoretical physics bounding box.

[0060] The expansion coefficient λ is determined using a dynamic mapping mechanism. The system reads the three-axis angular velocity output from the microelectromechanical system gyroscope in real time. . The calculation formula is as follows:

[0061] ;

[0062] in, This is the basic safety margin, with a value ranging from 1.10 to 1.15. The magnitude of the angular velocity vector; This is the jitter sensitivity coefficient, with a value ranging from 0.01 to 0.05 s / rad; This represents the angular velocity saturation threshold (e.g., 2.0 rad / s). This formula indicates that the more severe the device jitter, the greater the margin of the system's reserved reading field.

[0063] Secure data acquisition bounding box The calculation logic is as follows:

[0064] ;

[0065] in, The starting coordinates are the top left corner after expansion. This refers to the expanded read width and height. and This represents the total width and height of the physical pixel array of the CMOS sensor 112. (Function) Perform boundary clamping operations to prevent calculated coordinates from going out of bounds and causing sensor driver errors.

[0066] Step S203: Calculate heterogeneous ISP parameter configuration based on field attributes. The core principle of this step is to maximize the signal-to-noise ratio in specific spectral bands or brightness ranges by adjusting the gain distribution of the analog front-end and digital processing pipeline. The system establishes physical attribute labels. With the underlying ISP register control vector The mapping relationship between them. Control vector It includes exposure time, analog gain, red / blue / green channel gain, and color correction matrix coefficients.

[0067] The general-purpose computing core 131 first calculates the environmental baseline vector. The system selects a global low-resolution frame. Calculate the weighted average brightness of the central region (e.g., the central 50% area of ​​the image). and color temperature estimate The environmental baseline vector is represented as... ,in The baseline exposure compensation value calculated by the automatic exposure (AE) algorithm.

[0068] Subsequently, the system based on field attributes Query the pre-defined multidimensional parameter lookup table to obtain the parameter bias vector. And calculate the final hardware configuration parameters.

[0069] ;

[0070] The bias vector in the lookup table Obtained through offline calibration experiments. The calibration method is as follows: Under standard light source, for typical ticket samples with different attributes (such as red seals and holographic anti-counterfeiting films), traverse the ISP parameter space and select the parameter combination that maximizes the contrast or edge gradient amplitude of the target area as the baseline value for storage. The specific mapping logic includes:

[0071] when When dealing with the red stamp attribute, the red channel gain bias output by the lookup table is positive (e.g., +3.0dB), while the green / blue channels are negative, giving the image a red-pass filter effect, thereby enhancing the stamp texture while removing background interference.

[0072] when When dealing with the properties of reflective anti-counterfeiting film, the exposure time offset output by the lookup table is negative (e.g., -1.5EV) to prevent highlight overflow.

[0073] when When dealing with low-contrast bitmap text attributes, the lookup table specifies the loading of a high-intensity sharpening kernel to enhance high-frequency information.

[0074] Step S204: Construct a hardware-level asynchronous acquisition instruction queue. The general-purpose computing core 131 encapsulates the calculated spatial parameters and image processing parameters into standardized instruction units. The system follows the spatial scanning order of each field on the sensor plane (i.e., The instruction units are sorted (from smallest to largest) to generate an ordered queue. The purpose of sorting is to reduce the switching overhead of the CMOS sensor's row 112 addressing circuit. The final generated queue... It is written to the direct memory access (DMA) buffer and awaits transfer to the configuration interface between the image acquisition module 110 and the image signal processing module 120.

[0075] See attached document Figure 4 , Figure 4 The diagram illustrates the timing and data flow of the hardware-level asynchronous serialization acquisition process. Step S300 specifically executes the process of using the underlying hardware registers to control the time-division multiplexing acquisition of the image sensor, including the following steps:

[0076] Step S301 involves performing register atomic configuration based on the vertical blanking interval. This step utilizes the inter-frame idle period in the image transmission protocol for hardware state switching. The general-purpose computing core 131 in the central processing module 130 does not directly call the camera capture interface of the upper-layer operating system, but instead listens for the vertical synchronization signal (VSYNC) output by the image acquisition module 110 by registering a hardware interrupt service routine (ISR). When a frame of image transmission ends and the data bus is in a low-level state, the system enters the vertical blanking interval (VBI). For a common 60fps frame rate setting, the duration of VBI is typically 1 to 4 milliseconds.

[0077] During the VBI period, the general-purpose computing core 131 triggers an interrupt response from the acquisition command queue generated in step S200. Extract the head command unit from the queue. The general-purpose computing core 131 writes the spatial coordinate parameters in the instructions into the windowing control register of the CMOS sensor 112 via the I2C or SP1 serial bus, and writes the image processing parameters into the shadow register of the ISP.

[0078] A double-buffering mechanism is used here to ensure atomicity: the parameters currently being used by the hardware are stored in the active register set, while the general-purpose computing core 131 writes to the shadow register set. After the write operation is complete, the general-purpose computing core 131 writes an activation instruction to the update trigger bit. The moment the VBI ends and the next frame begins, the hardware logic automatically atomically copies the data from the shadow register to the active register, thereby ensuring that the exposure parameters and spatial parameters within a single frame are strictly synchronized, preventing tearing phenomena where the first half of the frame uses parameter A and the second half uses parameter B.

[0079] In step S302, the CMOS sensor 112 performs physical-level windowing exposure and readout. When VBI ends and a new frame period begins, the timing generator inside the CMOS sensor 112 reads the updated windowing control register. The line address decoder is only activated. The starting line defined in Until the end of the line Pixel arrays between; column-level ADCs (analog-to-digital converters) only for The starting column defined in To the end column The pixel charge is sampled and quantized.

[0080] This step utilizes the hardware feature of the CMOS sensor 112 that supports random pixel addressing. Compared to the traditional full-frame readout mode, the frame readout time in the windowed readout mode is... It is significantly shortened, and its theoretical calculation formula is as follows:

[0081] ;

[0082] in, The height of the safety bounding box (unit: number of rows); The physical time required for the sensor to read a single row of pixels, including row gating, reset, charge transfer and quantization time, typically ranging from 10 microseconds to 30 microseconds; The inherent line overhead for frame headers and trailers is typically 10 to 50 lines.

[0083] For example, when the full frame height For 3000 rows, and the local slice height For 300 lines, the frame readout time It can be reduced to about 1 / 10 of the full-frame mode. This means that, under the same data bus bandwidth constraints, this method allows the system to perform multiple exposures on local areas at a higher instantaneous frame rate (e.g., above 120fps).

[0084] In step S303, the image signal processing module 120 performs heterogeneous parameter pipeline processing. The windowed raw data stream output by the CMOS sensor 112 enters the image signal processing module 120.

[0085] At this time, the automatic exposure (AE) and automatic white balance (AWB) algorithms within the image signal processing module 120 are locked. The heterogeneous parameters written in the hardware pipeline forced loading step S301 are... .

[0086] The ISP pipeline executes the following processing logic sequentially:

[0087] Digital gain applications: based on parameters The gain component in the data performs digital multiplication on the original Raw data to increase the signal amplitude of a specific channel;

[0088] De-mosaic: The Bayer array is restored to a full-color RGB image using an edge-sensitive interpolation algorithm. Since the input data is a local slice, the interpolation algorithm only performs neighborhood convolution within the current window, and a mirror filling strategy is used at the edges to avoid black border effects.

[0089] Color space transformation: Apply a specific color correction matrix (CCM) to map the RGB space to the target color space, thereby enhancing the color separation of specific physical properties such as the red of a stamp.

[0090] Finally, the image signal processing module 120 outputs the heterogeneous local image slices corresponding to the instruction. The slice is directly stored in a reserved buffer in memory and marked with the corresponding instruction number. The system repeats steps S301 to S303 until the acquisition instruction queue is reached. The result is an empty sequence of high-resolution image slices that are spatially discrete and heterogeneous in attributes.

[0091] See attached document Figure 5 , Figure 5 This diagram illustrates the principle of cross-domain spatial registration based on gradient domain invariance. Step S400 specifically performs a calibration process to eliminate field-of-view displacement caused by asynchronous hardware acquisition and handheld shaking. The core of this step lies in solving the registration problem between heterogeneous images and multi-scale images. Because local image slices have high resolution and have undergone specific ISP enhancement (such as red text enhancement), while the global background image is low resolution and has standard colors, there are significant differences between the two in pixel grayscale values ​​and spatial scale. Therefore, this step uses gradient-based feature space for registration, rather than directly utilizing pixel grayscale, and includes the following steps:

[0092] Step S401: Perform multi-scale spatial alignment and gradient magnitude extraction. First, resolve the resolution mismatch issue. The system acquires the physical resolution scaling ratio of the CMOS sensor 112 between preview mode (step S100) and windowed mode (step S300). ( ).

[0093] General-purpose computing core 131 from global low-resolution images In the middle, based on the theoretical physical boundary box calculated in step S201 Extract the corresponding reference image patch .at this time This is a low-resolution template.

[0094] Simultaneously, heterogeneous local image slices are acquired. In order to perform matching at the same frequency domain scale, the system... Perform downsampling processing to generate downsampling slices. Downsampling algorithms employ bilinear interpolation or Gaussian pyramid downsampling, making... pixel density and Consistent.

[0095] Subsequently, the system utilizes the invariance of the gradient domain to eliminate the differences introduced by the ISP parameters. The principle is that the gradient features of an image reflect the rate of change of pixel intensity, rather than the absolute intensity. For additive illumination changes (… Its gradient remains unchanged. For multiplicative gain changes ( Although its gradient magnitude changes proportionally, it is canceled out by the denominator in the normalized correlation calculation.

[0096] The system will and Convert to a single-channel grayscale image and calculate the horizontal gradient using the Sobel operator. and vertical gradient .

[0097] The convolution kernel of the Sobel operator is defined as follows:

[0098] ;

[0099] For the input image (represent or Its gradient magnitude map The calculation formula is:

[0100] ;

[0101] Here, * denotes a two-dimensional convolution operation. Through the above calculations, the system generates a low-resolution reference gradient map. and low-resolution search gradient graph In the gradient magnitude domain, the text edges and table borders of the document are preserved as high-frequency features, while the overall brightness differences caused by different exposure parameters are effectively filtered out.

[0102] Step S402: Perform sliding window matching based on normalized cross-correlation (NCC). The system will... As a template, in Slide matching is performed on top.

[0103] At each sliding position Calculate the normalized cross-correlation coefficient. The formula for calculating this coefficient is as follows:

[0104] ;

[0105] in, Traversal Reference Template All pixel coordinates; It is the pixel mean of the reference gradient map; It is the search gradient map at the current window position. The local pixel mean.

[0106] The system traverses the search region, constructs a correlation coefficient matrix, and searches for the global maximum value. and their corresponding positions To prevent false matches, the system sets a match confidence threshold. (For example, 0.6). If This indicates that the currently acquired slice is unusable due to severe blurring or complete deviation from the field of view. The system will mark this instruction as an acquisition failure and trigger a re-acquisition process in subsequent logic.

[0107] Step S403: Calculate the high-resolution displacement correction vector and update the coordinates. Since step S402 is performed at a low-resolution scale, the calculated displacement... It needs to be mapped back to a high-resolution coordinate system. The system first calculates the relative offset vector at low resolution. :

[0108] ;

[0109] ;

[0110] in, For downsampling slices Width and height, The width and height of the reference image patch are then used. Subsequently, scaling is applied. Restore the offset vector to physical pixel displacement. :

[0111] ;

[0112] The vector This vector represents the true physical misalignment of the high-resolution slice relative to its ideal position. The system uses this vector to correct the slice's metadata coordinates, updating the slice's absolute coordinates in the global coordinate system to its true physical coordinates. Through this step, the system utilizes a low-computing-power downsampling matching strategy to accurately compensate for the spatial errors introduced by high-resolution asynchronous acquisition.

[0113] See attached document Figure 6 , Figure 6 A flowchart illustrating the business logic closed-loop verification and hierarchical archiving process is shown. Step S500 specifically performs logical consistency verification of the identified data, hardware feedback control, and structured encapsulation of the final file. This step establishes a closed-loop correction mechanism from semantic understanding to physical acquisition by introducing business rule constraints. It includes the following steps:

[0114] In step S501, optical character recognition (OCR) is performed on the heterogeneous local image slices after coordinate correction. The central processing module 130 calls the built-in OCR inference engine and inputs the heterogeneous local image slice sequence after registration correction in step S400.

[0115] For the OCR inference engine, this embodiment adopts a convolutional recurrent neural network (CRNN) architecture. This model mainly consists of three parts:

[0116] Feature extraction layer: Based on a ResNet-34 or MobileNetV3 backbone network, used to extract high-dimensional feature maps from the input image. The input image is normalized to a fixed height (e.g., 32 pixels), and the width is scaled proportionally.

[0117] Sequence modeling layer: A bidirectional long short-term memory network is used, with each layer containing 256 hidden units to capture contextual information in the feature sequence and predict the character probability distribution at each time step.

[0118] Transcription layer: The connection-time classification loss function is used to decode the frame-level probability sequence into the final text string without the need for character-by-character segmentation and annotation.

[0119] The training data for the model consists of two parts: synthetic text images generated based on a font library (accounting for approximately 70%) and labeled images collected from real invoices (accounting for approximately 30%). During training, the Adadelta optimizer is used, with backpropagation aimed at minimizing CTCLoss.

[0120] Before inference, the system uses physical attribute labels. Perform specific image preprocessing. For example, when When a red stamp is indicated, the system discards the blue and green channels, extracting only the R channel data, and applies the maximum inter-class variance method for adaptive binarization. The physical principle behind this operation is to utilize the high reflectivity of the stamp ink in the red light band to eliminate interference from the black carbon ink in the background, thus separating the stamp characters into the foreground. Finally, the OCR engine outputs the recognized text string. And character-level confidence scores .

[0121] Step S502: Construct the business data vector and perform a logical balance check. The system will process all identified field text. Parsed into structured data, and populated into the business data vector. .

[0122] Subsequently, the system loads the ticket type. Corresponding logical verification rule set Logical verification function The definition is as follows:

[0123] ;

[0124] Here, Result is a Boolean value. Taking a VAT invoice as an example, the arithmetic balance rule requires:

[0125] ;

[0126] in, This is the floating-point calculation error tolerance, with a value of 0.01. The format verification rule uses a weighted modular arithmetic algorithm to verify whether the mathematical relationship between the invoice code and the check digit holds true.

[0127] Step S503: Perform hardware feedback resampling control based on the verification result. If the Result in step S502 is "false", or the confidence level of any key field... Below the preset safety threshold (For example, if set to 0.90, corresponding to a 90% probability confidence level), the system triggers a feedback control loop.

[0128] The system first locates the target field that caused the validation failure or low confidence. This feedback mechanism does not involve blind retries, but rather quantitative correction based on histogram statistics. The system calculates the brightness histogram of the current slice. ,in [0.255].

[0129] Overexposure detection and correction: Statistical analysis of the proportion of pixels with a brightness greater than 250. .like (For example, 5%) is considered overexposed. New exposure time. The calculation formula is:

[0130] ;

[0131] in This is the current exposure time. This is an adjustment factor (e.g., 0.5).

[0132] Underexposure detection and correction: Calculating the average brightness of the image .like (For example, if the target brightness is 128, it is determined to be underexposed. The correction formula is:)

[0133] ;

[0134] Simultaneously restrict Do not exceed the maximum exposure threshold to prevent motion blur.

[0135] Contrast enhancement correction:

[0136] If the brightness is normal but recognition fails, the system adjusts the contrast parameter in the ISP. Or switch the sharpening kernel coefficient .

[0137] Corrected parameters It is encapsulated as a high-priority acquisition instruction unit and inserted into the acquisition instruction queue. The header. This mechanism constitutes a negative feedback control system between the physical layer and the semantic layer, until the verification passes or the maximum number of retries (3 times) is reached.

[0138] Step S504: Encapsulate the hierarchical archive file. After data verification passes, the system generates the final archive file. This file uses OFD or a PDF container format with custom metadata and contains the following three logical layers:

[0139] Global background layer (Layer1): Stores the global low-resolution image obtained in step S100, using JPEG compression.

[0140] Original Evidence Layer (Layer 2): Stores heterogeneous local image slices obtained in step S300 and registered in step S400. This layer's data is stored in lossless PNG format and embedded with a digital signature, serving as a legally valid electronic original.

[0141] Structured Semantic Layer (Layer 3): Stores validated business data vectors XML description of the position coordinates of each field in the global coordinate system.

[0142] In the rendering engine, Layer 1 serves as the base image; Layer 2 is dynamically overlaid on Layer 1 based on its coordinate information, and is only loaded and displayed when the user's zoom level exceeds a preset value (such as 2.0x). This layered rendering technology resolves the contradiction between high-resolution evidence preservation and low-bandwidth browsing transmission.

[0143] To provide a more intuitive understanding of the present invention, the above solution will be specifically explained below in the context of a mobile application for collecting "value-added tax invoices".

[0144] Application scenario description:

[0145] This embodiment operates on a handheld industrial-grade PDA terminal, which is equipped with a 12-megapixel (4000×3000) CMOS image sensor (model: Sony IMX series), has a MIPICSI-2 interface, and is connected to a SoC processor with a built-in NPU. The user is in a warehouse environment with complex lighting conditions, and uses the handheld PDA to scan and archive a VAT invoice laid flat on a table.

[0146] Execution process details:

[0147] Global perception and planning:

[0148] The PDA starts the camera preview and outputs a VGA (640×480) resolution video stream. The system recognizes a VAT invoice in the center of the image. .

[0149] The system loads semantic topology and identifies three key heterogeneous regions:

[0150] Area A (Invoice Code / Number): Located in the upper right corner, with the following attributes: (Black dot-matrix printed text), susceptible to motion blur. System configuration strategy: high gain, short exposure (1 / 1000s), enhanced sharpening.

[0151] Area B (Buyer's Name): Located in the top left corner, with the following attributes: (Standard printed text), standard configuration.

[0152] Area C (Invoice Stamp): Located in the lower right corner, with the following attributes: (Red ink) is easily affected by background text. System configuration strategy: Only enable red channel gain (R_Gain=2.5x), and reduce exposure to prevent red channel overflow.

[0153] The gyroscope detected a slight tremor in the user's hand. The system calculates the field of view factor. This generates an expanded security capture frame.

[0154] Asynchronous serialization acquisition:

[0155] Using the inter-frame VBI time slot, the system writes the parameters of region A to the ISP shadow register. At the start of the next frame, the sensor reads only the 300 rows of pixels corresponding to region A, which takes about 4ms.

[0156] Following the VBI, the system writes the parameters for region C (in red enhancement mode). In the next frame, the sensor reads the 400 rows of pixels corresponding to region C.

[0157] At this point, the entire acquisition process took only about 20ms, much faster than the 33ms (30fps) required for full-frame high-resolution acquisition, effectively freezing hand tremors.

[0158] Spatial registration and loop closure verification:

[0159] Although the system-acquired region C slice clearly displays the seal, it is displaced by 15 pixels relative to the global preview image due to hand movement. Step S400 calculates the correction vector through gradient magnitude matching. The seal slices are precisely aligned back to the document coordinate system.

[0160] The OCR engine recognized the code in area A with a confidence level of 0.98; however, due to glare, the confidence level for recognizing the stamp content in area C was only 0.65.

[0161] Closed-loop feedback: The verification logic detects low stamp confidence and triggers S503. The system analyzes the histogram and finds highlight overflow, automatically calculates negative exposure compensation (-1.0EV), and immediately inserts a resampling instruction for region C. The hardware immediately performs resampling in the next VBI, obtaining a clear slice, and the verification passes.

[0162] Hierarchical archiving:

[0163] The final OFD file contains: a 50KB JPEG preview image (Layer 1), three high-resolution, lossless PNG slices (Layer 2, occupying only 300KB), and XML structured data (Layer 3). Compared to traditional lossless full-image saving (approximately 8MB), storage space is reduced by more than 95%.

[0164] To verify the advantages of the proposed "method for acquisition based on asynchronous windowing and heterogeneous parameters" (hereinafter referred to as "the method of this application") compared with the prior art, the applicant conducted a comparative experiment in a standard testing environment.

[0165] Experimental environment setup:

[0166] Hardware platform: Development board equipped with Qualcomm Snapdragon 8 series processor and IMX586 sensor.

[0167] Test dataset: Contains 1000 different types of financial documents (invoices, checks, receipts), scene coverage: direct strong light, shadow occlusion, hand-held shaking (simulated by a robotic arm simulating sinusoidal vibration), and large-angle tilt.

[0168] Comparison objects:

[0169] Option A (existing technology): Traditional single-shot full-frame automatic exposure acquisition. This involves capturing a high-resolution full-frame image after focusing, using uniform ISP parameters, and then performing OCR.

[0170] Option B (method of this application): adopts the asynchronous multi-window heterogeneous acquisition and closed-loop feedback process described in steps S100 to S500.

[0171] Evaluation index definition:

[0172] Field character recognition accuracy: The percentage of key fields (amount, date, code) that are completely and correctly recognized.

[0173] Average processing time: The total time from when the user presses the capture button to when structured data is output.

[0174] Archive file size: The average size of the final file containing evidence images.

[0175] Robustness in bright / blurred scenes: Success rate of recognition under extreme lighting or jitter.

[0176] Analysis of experimental results:

[0177] The test results are shown in Table 1, and the corresponding comparison charts are shown below. Figures 7 to 12 As shown.

[0178] Table 1 shows the performance comparison data between the method of this application and the prior art in different dimensions.

[0179] Test metrics Option A (Primary Technology) Option B (the method described in this application) Performance improvements / changes Overall OCR accuracy 89.40% 98.20% 8.80% Seal / Red Letter Recognition Rate 72.10% 96.50% 24.40% Average processing time 450ms 320ms -28.90% Average file size 4.5MB 0.4MB Save 91% Success rate in highly reflective scenes 45.00% 92.00% 47.00%

[0180] The results are discussed in the appendix. Figures 7 to 12 Analysis of Improved Recognition Rate: Existing technologies, when processing text covered by stamps or reflective anti-counterfeiting films, often result in overexposed, washed-out highlights or confusion between the stamp and background text due to the use of globally uniform exposure. This application, through heterogeneous ISP parameters, separately enables red channel enhancement and underexposure for the stamp area and performs histogram equalization for the anti-counterfeiting film area, significantly improving image quality in complex areas, thereby increasing the stamp recognition rate by over 24%.

[0181] Timeliness Analysis: Although this application involves multiple data acquisitions, the use of windowed readout means that only about 1 / 10 of the sensor's data is read each time, resulting in extremely short physical readout time. Furthermore, it eliminates the computational overhead of full-image high-resolution ISP processing and full-image OCR, thus reducing overall link time by approximately 30%.

[0182] Storage optimization: The tiered archiving strategy avoids storing useless high-resolution background data and retains only high-resolution slices of key evidence, greatly reducing server storage costs for large-scale enterprise archiving.

[0183] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent recognition and archiving of financial documents in an embedded hardware terminal, characterized in that, Includes the following steps: Acquire global low-resolution images and generate an acquisition instruction queue containing spatial coordinate parameters and heterogeneous ISP parameters based on semantic analysis results; The acquisition instruction queue is parsed, and the image sensor is controlled to perform hardware-level asynchronous serialized acquisition to obtain heterogeneous local image slices corresponding to the spatial coordinate parameters; Cross-domain spatial registration is performed between the heterogeneous local image slice and the global low-resolution image to eliminate the field-of-view displacement caused by asynchronous acquisition and update the physical coordinates of the heterogeneous local image slice. Optical character recognition is performed on the registered heterogeneous local image slices to extract business data, and logical consistency verification is performed on the business data. If the verification passes, a hierarchical archive file containing a global low-resolution image, heterogeneous local image slices, and business data will be generated. If the verification fails, a re-sampling instruction containing correction parameters is generated and inserted into the acquisition instruction queue, triggering feedback re-sampling control.

2. The method for intelligent recognition and archiving of financial documents in an embedded hardware terminal according to claim 1, characterized in that, The generation of the acquisition instruction queue, which includes spatial coordinate parameters and heterogeneous ISP parameters, specifically includes: Calculate the theoretical physical bounding box of the key fields on the sensor plane, and read the angular velocity data of the gyroscope to calculate the anti-shake expansion coefficient. Redundantly expand the theoretical physical bounding box to generate a safe acquisition bounding box, which serves as the spatial coordinate parameter. Based on the physical attributes of the key field, a preset lookup table is queried to obtain the reference vector and attribute bias vector, and a specific exposure and color configuration for the key field is calculated and generated as the heterogeneous ISP parameters. The spatial coordinate parameters and the heterogeneous ISP parameters are encapsulated into instruction units, and the acquisition instruction queue is constructed by sorting them according to the spatial scanning order.

3. The method for intelligent recognition and archiving of financial documents in an embedded hardware terminal according to claim 1, characterized in that, The control of the image sensor to perform hardware-level asynchronous serialization acquisition specifically includes: Register a hardware interrupt service routine to listen to the vertical synchronization signal of the image sensor. During the vertical blanking interval, extract the instruction unit from the acquisition instruction queue, write the spatial coordinate parameters in the instruction unit into the windowing control register, and write the heterogeneous ISP parameters into the shadow register. By using a double buffering mechanism, when the vertical blanking interval ends and the next frame period begins, the data in the shadow register is automatically atomically copied to the active register so that the heterogeneous ISP parameters take effect in the current frame.

4. The method for intelligent recognition and archiving of financial documents in an embedded hardware terminal according to claim 3, characterized in that, The control of the image sensor to perform hardware-level asynchronous serialization acquisition also includes: The image sensor's row address decoder activates only the pixel array between the start and end rows based on the spatial coordinate parameters in the windowing control register. The column-level analog-to-digital converter of the image sensor samples and quantizes the pixel charge from the start column to the end column, and outputs the heterogeneous local image slice in a windowed readout mode.

5. The method for intelligent recognition and archiving of financial documents in an embedded hardware terminal according to claim 1, characterized in that, The cross-domain spatial registration of the heterogeneous local image slices with the global low-resolution image specifically includes: A reference image block is extracted from the global low-resolution image, and the heterogeneous local image slice is downsampled to generate a downsampled slice, so that the pixel density of the two is consistent. The gradient magnitude maps of the reference image block and the downsampled slice are calculated using the Sobel operator to filter out the brightness and color differences caused by the heterogeneous ISP parameters. Perform sliding window matching within the gradient magnitude domain, calculate the normalized cross-correlation coefficient, and determine the optimal matching position of the downsampled slice relative to the reference image block.

6. The method for intelligent recognition and archiving of financial documents in an embedded hardware terminal according to claim 5, characterized in that, Updating the physical coordinates of the heterogeneous local image slice specifically includes: Calculate the relative offset vector at low resolution based on the optimal matching position; Obtain the physical resolution scaling ratio of the image sensor between preview mode and windowed readout mode; The relative offset vector is mapped to a physical pixel displacement at high resolution using the physical resolution scaling factor, and the physical pixel displacement is used to correct the physical coordinates of the heterogeneous local image slice.

7. The method for intelligent recognition and archiving of financial documents in an embedded hardware terminal according to claim 1, characterized in that, The logical consistency verification of the business data specifically includes: The text output by optical character recognition is parsed into structured data and populated into the business data vector; Load the logical verification rule set corresponding to the current bill type. The logical verification rule set includes at least the amount arithmetic balance rule and the check code format rule. Perform logical verification operations on the business data vector to generate a Boolean verification result.

8. The method for intelligent recognition and archiving of financial documents in an embedded hardware terminal according to claim 1, characterized in that, The trigger feedback resampling control specifically includes: Statistically analyze the brightness histogram of the current heterogeneous local image slices, and calculate the proportion of bright pixels and the average brightness; If the proportion of bright pixels exceeds the overexposure threshold, calculate the correction factor for reducing the exposure time; if the average brightness is lower than the underexposure threshold, calculate the correction factor for increasing the exposure time or gain. New heterogeneous ISP parameters are generated based on the correction coefficient, encapsulated as high-priority acquisition instruction units, and inserted at the head of the acquisition instruction queue.

9. The method for intelligent recognition and archiving of financial documents in an embedded hardware terminal according to claim 1, characterized in that, The hierarchical archive files adopt a layout document container format, and their logical structure includes: A global background layer stores the global low-resolution image; The original credential layer stores the registered heterogeneous local image slices and embeds digital signatures; The structured semantic layer stores the business data and the position coordinates of each field in the global coordinate system. The original credential layer is configured to be overlaid on the global background layer.

10. The method for intelligent recognition and archiving of financial documents in an embedded hardware terminal according to claim 1, characterized in that, Before performing optical character recognition on the registered heterogeneous local image slices, the method further includes: Preprocessing is performed based on the physical attribute labels of the heterogeneous local image slices; If the physical attribute label indicates a seal, only the red channel data of the image is extracted, and adaptive binarization is performed using the maximum inter-class variance method to separate the seal characters from background interference.

Images