Display screen alignment detection system and detection method thereof

Abstract

The invention relates to the technical field of machine vision, in particular to a display screen alignment detection system and a detection method thereof, and the system comprises an image acquisition and analysis module, a gradient distribution operation module, an intelligent vision detection module, a local residual error extraction module and a posture self-compensation reasoning module. According to the method, continuous coordinate sorting and displacement difference operation are carried out on edge pixels of a display screen, offset gradient distribution reflecting a real edge change trend is constructed, tiny alignment deviation is amplified and presented in a spatial dimension, then the gradient distribution is mapped into multi-dimensional features for weighted discrimination, and therefore the accuracy of the alignment deviation is improved. According to the method, non-uniform deviation forms are effectively distinguished, and meanwhile, attitude reasoning basis is established by combining the mass center of an abnormal area and the deviation amplitude, so that a detection result has clear spatial directivity and interpretability, the stability and correction accuracy of alignment judgment under the rotating working condition are improved, the automatic detection effect under the complex working condition is further optimized, and the detection efficiency is improved. And the adaptive capacity of the system to a dynamic change environment is enhanced.

Description

Technical Field

[0001] This invention relates to the field of machine vision technology, and in particular to a display screen alignment detection system and its detection method. Background Technology

[0002] Machine vision technology refers to the field of technology that utilizes computer vision principles and algorithms to enable computers to acquire, analyze, recognize, and process image information through camera equipment. Its core aspects include image acquisition, image processing, target recognition, localization, and tracking. Machine vision is widely used in industrial automation, quality inspection, security monitoring, and intelligent transportation. Through the processing and analysis of image data, machine vision systems can automatically perform image recognition and decision-making, and execute environmental perception and control. With technological advancements, the accuracy, speed, and application scope of machine vision continue to expand, becoming a crucial foundation for achieving automation and intelligence.

[0003] Traditional display screen alignment detection systems refer to position detection systems used in the display screen manufacturing process. They employ cameras or image sensors to acquire image information of the display screen and analyze its features within the image to determine if it is aligned with a predetermined position. Traditional methods typically rely on a fixed-position camera, using image comparison technology to judge the alignment between the display screen and the detection equipment. This method often uses specific image processing algorithms, such as edge detection and template matching, to determine if the display screen is in the correct position and to provide feedback on any deviations.

[0004] Existing display alignment detection relies on fixed cameras to acquire images and perform overall feature comparison. During operation, it is highly dependent on the stability of the installation position and viewing angle. Under conditions of rotating workstations or changes in posture, feature distortion is prone to occur. Image processing mostly uses static edges or templates as the basis for judgment, which makes it difficult to depict the continuous displacement relationship between edge pixels. It is not responsive enough to subtle shift changes, and deviations are only identified after accumulation. The detection results lack spatial directionality and cannot accurately locate the source of abnormal areas, thus affecting the efficiency of subsequent posture correction and manual review. Summary of the Invention

[0005] To address the technical problems of existing display screen alignment detection systems that rely on fixed cameras to acquire images and perform overall feature comparison, which are highly dependent on the stability of the installation position and viewing angle during operation, prone to feature distortion under rotating positions or posture changes, and whose image processing often uses static edges or templates as the basis for judgment, making it difficult to depict the continuous displacement relationship between edge pixels, insufficient response to subtle offset changes, and deviations only being identified after accumulation, the detection results lack spatial directionality and cannot accurately locate the source of abnormal areas, thus affecting the efficiency of subsequent posture correction and manual verification, this invention provides a display screen alignment detection system and method. The technical solution is as follows: On the one hand, a display screen alignment detection system is provided, the system comprising: The image acquisition and analysis module acquires images of the border of the display screen on the rotating workstation, uses the Canny edge detection algorithm to calculate the edge response, extracts edge pixels, performs spatial coordinate sorting, and generates an edge pixel coordinate sequence. The gradient distribution calculation module calculates the displacement difference of corresponding pixels based on the edge pixel coordinate sequence and the preset reference coordinate sequence, performs differential calculation on the displacement difference of adjacent pixels, and generates the edge offset gradient field of the display screen. The intelligent visual detection module maps the gradient field of the display screen edge offset into a feature tensor, uses a convolutional neural network model for channel weighting and classification, and generates a display screen alignment deviation feature map. The local residual extraction module extracts pixel regions that exceed a preset deviation threshold based on the alignment deviation feature map of the display screen, calculates the spatial centroid coordinates and deviation magnitude of the pixel regions, performs spatial mapping with the edge pixel coordinate sequence, and generates local alignment anomaly regions. The attitude self-compensation inference module calls the local alignment anomaly region, constructs an attitude input set based on the region centroid coordinates and offset amplitude, uses a support vector machine model to analyze attitude correction and performs manual re-judgment, and generates display screen alignment detection results.

[0006] As a further aspect of the present invention, the edge pixel coordinate sequence includes pixel position, edge identifier and spatial sorting order; the display edge offset gradient field includes offset amplitude, angle gradient and displacement change rate; the display alignment deviation feature map includes classification label, feature weight and detection reliability; the local alignment anomaly region includes anomaly position, deviation value and region deviation direction; and the display alignment detection result includes correction magnitude, re-judgment decision and pass / fail judgment result.

[0007] As a further aspect of the present invention, the image acquisition and parsing module includes: The image acquisition submodule acquires frame sequences by acquiring images of the border of the display screen on the rotating workstation, performs pixel array reading and grayscale value mapping on the frame images, records row and column indices according to the imaging scanning order, performs order verification on the frame sequences and aggregates them to generate a border grayscale image matrix. The edge response submodule, based on the grayscale image matrix of the border, uses the Canny algorithm to calculate the grayscale difference between adjacent pixels, performs difference calculations in the horizontal and vertical directions respectively, synthesizes the difference results, classifies pixels according to the amplitude distribution and records the row and column coordinates, and generates an edge response pixel set. The coordinate sorting submodule calls the edge response pixel set, extracts the row and column coordinate values ​​corresponding to the pixels, performs ascending row coordinate sorting on the coordinate set while maintaining the consistency of the column coordinate association, and performs sequential combination of the sorting results to generate an edge pixel coordinate sequence.

[0008] As a further aspect of the present invention, the gradient distribution calculation module includes: The displacement difference acquisition submodule obtains the row and column coordinates of the corresponding pixel points based on the edge pixel coordinate sequence and the preset reference coordinate sequence. It performs coordinate matching for pixels with the same index and generates a pixel displacement difference vector set by subtracting the edge pixel row and column coordinates from the reference row and column coordinates. The differential sequence construction submodule calls the pixel displacement difference vector set, obtains the adjacent pixel index based on the edge pixel coordinate sequence, performs differential operation on the displacement components under the adjacent index, binds the differential result with the index position and arranges them in order to generate displacement difference sequence data; The gradient field generation submodule obtains the row vector difference and column vector difference corresponding to the pixel index based on the displacement difference sequence data, constructs a grid according to the coordinate mapping relationship, fills the grid nodes with difference components and verifies the index consistency, organizes the node difference distribution relationship, and generates the display screen edge offset gradient field.

[0009] As a further aspect of the present invention, the intelligent visual detection module includes: The gradient mapping submodule maps the gradient field offset of the display edge to a feature tensor, obtains the row and column gradient values ​​corresponding to the pixel, writes the gradient values ​​into the tensor channel according to the pixel index consistency, verifies the tensor dimension and numerical integrity, and generates the edge gradient feature tensor. The convolutional feature extraction submodule calls the edge gradient feature tensor, performs convolutional sliding operation on the pixel region in the tensor, records the positional convolutional response value, aggregates the response according to different scales, rearranges the channel index of the aggregation result, and generates a multi-scale spatial feature matrix. Note that the discrimination generation submodule obtains channel feature values ​​based on the multi-scale spatial feature matrix, calculates weight coefficients and performs weighted superposition to form a discrimination vector, performs spatial index mapping and numerical normalization on the discrimination vector, and generates a display screen alignment deviation feature map.

[0010] As a further aspect of the present invention, writing gradient values ​​into the tensor channel based on pixel index consistency means that the gradient value writing operation is performed only when the pixel row index and column index simultaneously meet the preset edge index range. The preset edge index range is the area that is no more than a fixed pixel width from the physical edge of the display screen.

[0011] As a further aspect of the present invention, the local residual extraction module includes: The residual data extraction submodule collects pixel deviation data based on the alignment deviation feature map of the display screen, compares the deviation value with a preset deviation threshold, filters pixel areas that exceed the threshold, records the pixel coordinates and deviation magnitude, and generates local residual data. The deviation amplitude calculation submodule extracts pixel coordinate information based on the local residual data, performs weighted calculations according to the coordinate distribution, calculates the spatial centroid coordinates, and combines the corresponding deviation amplitude to obtain the spatial distribution data of the deviation amplitude. The abnormal region generation submodule, based on the spatial distribution data of the deviation amplitude, calls the edge pixel coordinate sequence, performs spatial mapping operation according to spatial adjacency and geometric relationship, calculates the spatial boundary of the corresponding region in combination with the edge pixel coordinate, performs region stitching processing, and generates a local alignment abnormal region.

[0012] As a further aspect of the present invention, the preset deviation threshold is based on the multi-frame alignment deviation feature map obtained by the display screen in the calibration state. The pixel deviation data in each frame alignment deviation feature map is statistically processed to calculate the mean and dispersion parameters of the pixel deviation data. A deviation statistical interval is constructed based on the mean and dispersion parameters, and the deviation value corresponding to the upper limit of the deviation statistical interval is set as the preset deviation threshold.

[0013] As a further aspect of the present invention, the attitude self-compensation inference module includes: The regional data receiving submodule acquires the local alignment anomaly region, collects the region centroid coordinates and offset amplitude data, performs consistency verification on the centroid coordinates according to the coordinate axis definition rules, performs dimensional consistency detection and effective interval filtering on the offset amplitude, and generates an abnormal region attitude basic dataset. The attitude input construction submodule extracts the centroid coordinates as spatial position parameters based on the attitude base dataset of the abnormal region, performs normalization processing on the offset amplitude, combines the position and amplitude to form a vector concatenation, performs consistency judgment according to the vector dimension standard, and generates an attitude inference input vector set. The alignment determination generation submodule calculates the vector interval parameter based on the posture inference input vector set, classifies and marks the posture correction direction according to the distribution of the interval parameter, integrates the classification results based on the manual review information, summarizes the posture state and determination label, and generates the display screen alignment detection result.

[0014] On the other hand, a display screen alignment detection method, which is based on the aforementioned display screen alignment detection system, includes the following steps: S1: Acquire the image of the border of the display screen on the rotating workstation, use the Canny edge detection algorithm to calculate the edge response, extract the edge pixels and perform spatial coordinate sorting to generate an edge pixel coordinate sequence; S2: Based on the edge pixel coordinate sequence and the preset reference coordinate sequence, calculate the displacement difference of the corresponding pixel point, perform differential calculation on the displacement difference of adjacent pixel points, and generate the display screen edge offset gradient field. S3: Map the edge offset gradient field of the display screen into a feature tensor, and use a convolutional neural network model to perform channel weighting and classification to generate a display screen alignment deviation feature map. S4: Based on the alignment deviation feature map of the display screen, extract the pixel region that exceeds the preset deviation threshold, calculate the spatial centroid coordinates and deviation magnitude of the pixel region, perform spatial mapping with the edge pixel coordinate sequence, and generate a local alignment anomaly region. S5: Call the local alignment anomaly region, construct an attitude input set based on the region centroid coordinates and offset amplitude, use a support vector machine model to analyze the attitude correction and perform manual re-judgment, and generate the display screen alignment detection result.

[0015] The beneficial effects of the technical solutions provided by the embodiments of the present invention include at least the following: By performing continuous coordinate sorting and displacement difference calculation on the edge pixels of the display screen, an offset gradient distribution reflecting the actual edge change trend is constructed, which amplifies and presents small alignment deviations in the spatial dimension. The gradient distribution is then mapped to multi-dimensional features for weighted discrimination, which effectively distinguishes non-uniform offset patterns. At the same time, the centroid of the abnormal region and the offset amplitude are combined to establish the attitude reasoning basis, so that the detection results have clear spatial orientation and interpretability, thereby improving the stability and accuracy of alignment judgment under rotation conditions. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the accompanying drawings without creative effort.

[0017] Figure 1 This is a system schematic diagram of the present invention; Figure 2 This is a schematic diagram of the system framework of the present invention; Figure 3 This is a flowchart of the image acquisition and parsing module in this invention; Figure 4 This is a flowchart of the gradient distribution calculation module in this invention; Figure 5 This is a flowchart of the intelligent visual detection module in this invention; Figure 6 This is a flowchart of the local residual extraction module in this invention; Figure 7 This is a flowchart of the attitude self-compensation inference module in this invention; Figure 8 This is a flowchart of the method of the present invention. Detailed Implementation

[0018] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0019] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0020] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0021] In this embodiment of the invention, sometimes the subscript such as W1 is written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0022] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0023] This invention provides a display screen alignment detection system, such as... Figure 1-2 The diagram shown illustrates a display screen alignment detection system, which includes: The image acquisition and analysis module acquires images of the border of the display screen on the rotating workstation, uses the Canny edge detection algorithm to calculate the edge response, extracts edge pixels, performs spatial coordinate sorting, and generates an edge pixel coordinate sequence. The gradient distribution calculation module calculates the displacement difference of corresponding pixels based on the edge pixel coordinate sequence and the preset reference coordinate sequence, performs differential calculation on the displacement difference of adjacent pixels, and generates the edge offset gradient field of the display screen. The intelligent visual inspection module maps the gradient field of the display edge offset into a feature tensor, uses a convolutional neural network model for channel weighting and classification, and generates a display alignment deviation feature map. The local residual extraction module extracts pixel regions that exceed a preset deviation threshold based on the display alignment deviation feature map, calculates the spatial centroid coordinates and deviation magnitude of the pixel region, performs spatial mapping with the edge pixel coordinate sequence, and generates local alignment anomaly regions. The attitude self-compensation inference module calls the local alignment anomaly region, constructs the attitude input set based on the region centroid coordinates and offset amplitude, uses the support vector machine model to analyze attitude correction and performs manual re-judgment, and generates the display alignment detection result.

[0024] The edge pixel coordinate sequence includes pixel position, edge identifier and spatial sorting order; the display edge offset gradient field includes offset magnitude, angle gradient and displacement change rate; the display alignment deviation feature map includes classification label, feature weight and detection reliability; the local alignment anomaly region includes anomaly position, deviation value and regional deviation direction; the display alignment detection result includes correction magnitude, re-judgment decision and pass / fail judgment result.

[0025] Specifically, such as Figure 2 , 3 As shown, the image acquisition and parsing module includes: The image acquisition submodule acquires frame sequences by acquiring images of the border of the display screen on the rotating workstation, performs pixel array reading and grayscale value mapping on the frame images, records row and column indices according to the imaging scanning order, performs order verification on the frame sequences and aggregates them to generate a border grayscale image matrix. Turn on the industrial area scan camera installed directly above the rotary worktable and set the camera's acquisition resolution to [resolution value missing]. As the rotating workstation disk drives the display screen under test... The camera rotates in steps, triggering the shutter at each stop to capture a sequence of raw frame images of the display bezel. This activates the memory buffer, allowing the camera to read the color values ​​of the pixel array in each frame image sequentially, targeting the coordinates within the array... Extract the red channel component from any pixel. Green channel weight and blue channel components The values ​​are set as follows: the weighting coefficients for grayscale mapping are the weights of the red channel. Green channel weight and the weight of the blue channel Perform linear weighted summation operation For example, for a pixel located in the border area Its original RGB values ​​were collected as follows: The grayscale value is obtained through calculation. Round the floating-point value down to the nearest integer. The data is written to a temporary grayscale array, and the row index counter is initialized based on the raster path of the imaging scan. With column index counter From the top left corner of the image Start until the bottom right corner Each calculated grayscale value is bound to the row and column index of the current scan position and stored. After completing single-frame processing, the timestamp information and frame number in the frame sequence header are read. Check if the numbering of adjacent frames is consecutive, for example, by comparing the current frame. Compared to the previous frame satisfy If the verification passes, the grayscale data blocks corresponding to multiple frames are vertically stacked and stitched in memory according to the acquisition time order to generate a grayscale image matrix of the border.

[0026] The edge response submodule, based on the grayscale image matrix of the border, uses the Canny algorithm to calculate the grayscale difference between adjacent pixels, performs difference calculations in the horizontal and vertical directions respectively, synthesizes the difference results, classifies pixels according to the amplitude distribution and records the row and column coordinates, and generates an edge response pixel set. The system retrieves the grayscale image matrix of the border from memory and first performs Gaussian smoothing on it to reduce the impact of random noise and high-frequency interference on the edge detection results. Gaussian smoothing uses a two-dimensional Gaussian kernel function to perform a weighted summation of the grayscale values ​​in the neighborhood of the target pixel, generating a smoothed grayscale image matrix, which is then used as input data for subsequent gradient calculations. After Gaussian smoothing, based on the smoothed grayscale image matrix, arbitrary pixels within the matrix are... Using the center point as the reference point, the gradient calculation step in the Canny edge detection algorithm is executed. The gradient operator in the horizontal direction is defined as reading the adjacent pixels to the right of this point. The grayscale value and subtract the current point The grayscale value is used to obtain the horizontal grayscale difference: ; Define the gradient operator operation in the vertical direction as reading the adjacent pixels below the point. The grayscale value and subtract the current point The grayscale value is used to obtain the vertical grayscale difference: ; Subsequently, regarding the horizontal grayscale difference... Vertical grayscale difference Perform amplitude synthesis operation to obtain pixel points. Gradient magnitude: ; Based on gradient magnitude and gradient direction information, non-maximum suppression (NMS) processing is performed on the gradient magnitude matrix. The NMS processing includes: selecting the preceding and following pixels along the gradient direction of the current pixel; comparing the gradient magnitude of the current pixel with the gradient magnitudes of the adjacent pixels; and suppressing the gradient magnitude of the current pixel to zero when it is not a local maximum in that direction, thereby refining the edge width and eliminating spurious responses in non-edge regions. After NMS, to distinguish between real edges and background noise, a selection is made based on the non-border area of ​​the display screen. Using a pixel range as a background reference sample, the arithmetic mean of the pixel gradient magnitudes within this background sample region is statistically calculated. with standard deviation And set a high threshold accordingly. At the same time, a low threshold is set according to a preset ratio. The proportionality coefficient It is a positive number less than 1. Then, a double-threshold edge connection process is performed on the gradient magnitude matrix after non-maximum suppression: when the gradient magnitude of a pixel is greater than the higher threshold... When the gradient value of a pixel is within a low threshold, the pixel is classified as a strong edge point; when the gradient value of a pixel is within a low threshold, the pixel is classified as a strong edge point. With high threshold If a pixel is spatially connected to a strong edge point, it is considered a valid edge point; the remaining pixels are considered background noise and discarded. The final set of valid edge pixels constitutes the edge response pixel set. Table 1: Example of edge response calculation based on Canny edge detection algorithm

[0027] As shown in Table 1, the data in Table 1 is only used to illustrate the gradient calculation and gradient magnitude synthesis process. Non-maximum suppression and double-threshold edge connection processing are further performed based on the gradient magnitude calculation results. The intermediate results of this subsequent processing are not listed item by item. The gradient magnitude at a point is greater than the calculated threshold, therefore the point is identified and retained as an edge point.

[0028] The coordinate sorting submodule calls the edge response pixel set, extracts the row and column coordinate values ​​corresponding to the pixels, performs ascending row coordinate sorting on the coordinate set while maintaining the consistency of column coordinate relationships, and performs sequential combination of the sorting results to generate an edge pixel coordinate sequence. Read the edge response pixel set from the buffer pool. This set contains several unordered coordinate pairs, for example, the initial state of the set is... Start the traversal program and extract the first value of each pair of tuples as the row coordinate variable. Extract the second value as a column coordinate variable. Construct an unsorted array containing struct elements, and perform comparison and swap operations on this array, first sorting by row coordinates. As the primary key, compare two adjacent elements in the array. and The row coordinate value, if Then swap the elements. and In the array's storage location, shifting the smaller row coordinate forward, if This further triggers the secondary comparison logic, using column coordinates. As the second primary key, in comparison and The magnitude of the value, if If the two are swapped, then pixels within the same row are arranged in ascending column order. In the example above, the row coordinates are compared first. The smallest value is ranked first, then the two row coordinates are compared. The elements, based on column coordinates Less than The rules determine their relative order, and the row coordinates are... The element is inserted in the middle, and after multiple rounds of scanning and position swapping, the original set is reorganized into an ordered sequence. Finally, the sorted array indices are locked, and the coordinate tuples are read out sequentially according to the new memory address order and joined end to end to generate the edge pixel coordinate sequence.

[0029] Specifically, such as Figure 2 , 4 As shown, the gradient distribution calculation module includes: The displacement difference acquisition submodule obtains the row and column coordinates of the corresponding pixel points based on the edge pixel coordinate sequence and the preset reference coordinate sequence. It performs coordinate matching for pixels with the same index and generates a set of pixel displacement difference vectors by subtracting the edge pixel row and column coordinates from the reference row and column coordinates. Retrieve the edge pixel coordinate sequence, which has been strictly sorted according to the rule of row priority followed by column priority, to generate an ordered set. As input data, a pre-stored standard design reference coordinate sequence is synchronously loaded from the database. This reference sequence is generated based on the ideal edge path of the CAD drawing, and the coordinates of its corresponding index position are set as follows: Initialize the displacement calculation kernel and set the traversal pointer. from Start by incrementing, performing coordinate matching and difference operations for each index position, first extracting... Edge point of time With reference point Perform subtraction to calculate horizontal displacement With vertical displacement The displacement vector is obtained. Next, process Edge point of time With reference point Calculations yielded , Generate vectors Continue processing Edge point of time With reference point , and thus , Generate vectors Final processing Edge point of time With reference point , and thus , Generate vectors Formula is used here. The aim is to quantify the absolute positional deviation of the actual manufactured edge relative to the ideal design edge, perform physical range verification on each calculated vector component, and set a maximum permissible deviation threshold. Pixels, if or If the value is not found, it is marked as an anomaly and removed. If the current value is within the valid range, the vectors that pass the verification are stored in the linear list in sequence to generate a set of pixel displacement difference vectors.

[0030] The differential sequence construction submodule calls the pixel displacement difference vector set, obtains the index of adjacent pixels based on the edge pixel coordinate sequence, performs differential operation on the displacement components under the adjacent indices, binds the differential results with the index positions and arranges them in order to generate displacement difference sequence data; To analyze the microscopic continuity of edge deformation, the pixel displacement difference vector set in the buffer pool is accessed. First-order difference operations are performed on adjacent displacement vectors. This operation reveals local curvature or stretching changes by calculating the difference between the successor vector and the current vector. A difference operator is then defined. Specifically, this can be broken down into horizontal differences. and vertical difference Based on the generated vector set Start the loop calculation when the index At that time, read and Execute calculation , The results show that in the index to Vertical direction exists within the transition section A positive deformation mutation of units, when index At that time, read and ,calculate , This indicates that the displacement of this segment remains constant, with no relative deformation, when the index... At that time, read and ,calculate , The difference pairs obtained from the above calculations The data is bound to the original index position, and the specific values ​​are recorded in Table 2. This result intuitively reflects the deformation trend of the edge within different micro-segments. The results showed that there was a small step in the vertical direction at this point, and the key-value pairs were finally encapsulated in order to generate displacement difference sequence data.

[0031] Table 2: Data for Calculation of Displacement Difference Sequence

[0032] As shown in Table 2, the table lists the detailed values ​​for performing the difference operation on adjacent displacement vectors, where the index... The calculation results at this location show a non-zero difference in the vertical direction, revealing the local deformation characteristics at that location.

[0033] The gradient field generation submodule obtains the row vector difference and column vector difference corresponding to the pixel index based on the displacement difference sequence data, constructs a grid according to the coordinate mapping relationship, fills the grid nodes with difference components and verifies the index consistency, organizes the node difference distribution relationship, and generates the display edge offset gradient field. Based on the displacement difference sequence data, a resolution of is constructed. A discrete grid coordinate system for pixels is used to divide the edge area of ​​the display screen into several spatial grid nodes, and an index is established. To grid node The mapping relationship, setting the mapping rule as per Each consecutive index corresponds to one grid node, for example, index Data mapping to nodes , index and subsequent data mapping to nodes Extract the difference components within the node and perform mean calculation. For the node... ,extract gather and gather Calculate the horizontal gradient components Vertical gradient component Formula here The aim is to eliminate the impact of single-point measurement noise on the overall deformation assessment through statistical averaging, specifically targeting the calculated gradient vector. Set gradient consistency check threshold This threshold is determined by the gradient magnitude of the original good product data. quantile setting, i.e., if If so, it is determined that there is a risk of stress concentration in the area. In this example... The verification passed, and this result indicates that the node... The deformation within the region is within a safe and controllable range. The calculation and verification of the grid nodes are completed sequentially, and the gradient vectors of the nodes are filled into the matrix structure according to the spatial topological relationship to generate the edge offset gradient field of the display screen.

[0034] Specifically, such as Figure 2 , 5 As shown, the intelligent visual inspection module includes: The gradient mapping submodule maps the gradient field offset of the display edge to a feature tensor, obtains the row and column gradient values ​​of the corresponding pixels, writes the gradient values ​​into the tensor channel according to the consistency of the pixel index, verifies the tensor dimension and numerical integrity, and generates the edge gradient feature tensor. The generated display edge offset gradient field is retrieved, which stores the gradient vectors of the grid nodes in the form of a two-dimensional matrix. To construct a feature data structure suitable for deep analysis, an initial dimension is set. 3D edge gradient feature tensor The first two dimensions correspond to the spatial resolution of the grid, and the third dimension corresponds to the horizontal and vertical channels. A preset edge index range determination logic is set, and an effective edge width parameter is defined. Pixels, traversing the coordinates of each node in the gradient field Read its corresponding original pixel physical index Perform an interval determination operation for coordinates of The node, whose pixel index is and Determine whether the index value meets the requirement of distance from the physical edge. No more than The conditions of the pixel, i.e., the verification If the condition is true, the horizontal component of the node in the gradient field is then extracted. With vertical component , will the value Write Tensor coordinates Position, value Write coordinates Location, relative to coordinates The node, assuming its corresponding pixel index is ,check If this condition is not met, the region is determined to be outside the valid edge interest area, and the corresponding position of the tensor is forcibly moved. Fill with floating-point numbers After completing the full field mapping, start the tensor integrity verification program to count the number of non-zero elements in the tensor. Calculate the proportion of non-zero values Set an integrity threshold ,like This triggers an alarm and prompts that the mesh division needs to be readjusted. In the current example, the calculated mesh is... , percentage Once the requirements are met, the validated data structure is solidified, generating the edge gradient feature tensor.

[0035] The convolutional feature extraction submodule calls the edge gradient feature tensor, performs convolutional sliding operation on the pixel region in the tensor, records the positional convolutional response value, aggregates the response according to different scales, rearranges the channel index of the aggregation result, and generates a multi-scale spatial feature matrix. Load the edge gradient feature tensor in memory To capture the multi-scale spatial features of edge deformation, the system constructs separate systems with dimensions of [size missing]. small-scale convolution kernels With size Large-scale convolution kernels Set the convolution kernel The center weight is The weight of the four neighboring areas is The rest are Perform a sliding window operation to align the center of the convolution kernel with the coordinates of the tensor space plane. Extract the horizontal channel data block of the tensor within the local window. Perform element-wise multiplication and summation operations; the calculation formula is as follows: Assuming the center value of a local data block is The value above is Below is The left side is The right side is Substitute into numerical calculation The result Record as location The convolution response value, similarly used Perform the operation on the same location, assuming the response value is... ,Will The generated response matrix and The generated response matrix Depth-oriented stacking aggregation is performed, with specific computational data shown in Table 3. After aggregation, channel index rearrangement is executed, changing the data dimensions from... Adjusted to To adapt to subsequent processing, the channel is read first. The data forms the first feature surface, and then the channels are read. The data constitutes the second feature surface, generating a multi-scale spatial feature matrix.

[0036] Table 3: Multi-scale Convolution Feature Calculation Data Table

[0037] As shown in Table 3, the table lists the process of performing convolution kernel operations of different scales for specific spatial coordinate points, where the coordinates... In small-scale nuclei It exhibits a strong edge mutation response. In large-scale nuclear The smooth response is These two values ​​together constitute a multi-scale feature description of the point.

[0038] Note that the discrimination generation submodule obtains channel feature values ​​based on the multi-scale spatial feature matrix, calculates weight coefficients and performs weighted superposition to form a discrimination vector, performs spatial index mapping and numerical normalization on the discrimination vector, and generates a display screen alignment deviation feature map. Read the multi-scale spatial feature matrix and extract the first channel from the matrix. (Corresponding to small-scale features) and the second channel (Corresponding to large-scale features) values, calculate the global energy value of the channels to determine the fusion weights, respectively for and Perform the absolute value summation operation; the formula is as follows: Assuming the calculation yields , Calculate the total energy Then, the weight coefficient of the first channel is obtained. Second channel weighting coefficient This coefficient is used to perform a weighted superposition of the feature vectors at each location in space, with respect to the coordinates. eigenvectors (From Table 3), perform the operation This value This is the comprehensive discriminant value at that position. After traversing the matrix nodes and performing weighted calculations, an original discriminant vector set is formed. Subsequently, numerical normalization is performed to eliminate dimensional differences, and the maximum value in the set is found. and minimum value The range standardization formula is used. Regarding the aforementioned results Normalized calculation is The normalized values According to the original spatial index Remapping back to a two-dimensional grid coordinate system, using a bicubic interpolation algorithm to reduce the grid resolution from... Smooth upsampling to To refine the display effect, a final alignment deviation feature map of the display screen is generated.

[0039] Specifically, such as Figure 2 , 6 As shown, the local residual extraction module includes: The residual data extraction submodule collects pixel deviation data based on the display screen alignment deviation feature map, compares the deviation value with a preset deviation threshold, filters pixel areas that exceed the threshold, records the pixel coordinates and deviation magnitude, and generates local residual data. Retrieve from memory the bicubic interpolation generated Alignment deviation feature map of display screen at resolution, for coarse-grained grid coordinates The normalized high response value calculated at [location] Based on the interpolation mapping relationship, its corresponding mapping region in the current high-resolution map is determined, that is, by subdividing the grid coordinates. Read the coordinates of the center's neighborhood. Pixel deviation data at the location To determine whether the deviation value falls within the system's allowable process fluctuation range, the data for this model of display screen under standard alignment conditions is loaded from the database. From the original feature map data of each frame, extract the coordinates of each frame's map. The values ​​at each location are used to construct the original sample set. Perform statistical operations and calculate the sample mean. Compared with standard deviation: ; Assume the calculation result is , ,in accordance with The principle is to construct a deviation statistical interval, and the preset deviation threshold calculation formula is set as follows: Substituting the numerical values, we get , set here The standard deviation as a coefficient value is intended to filter out Random Gaussian noise is detected, followed by a numerical comparison operation to determine the current deviation value. Is it greater than the threshold? ,determination If the condition is met, confirm that there is an abnormal residual at this pixel, and immediately record the grid coordinates of this point. and residual amplitude exceeding the threshold The above verification process is performed on the pixels within the map, and the coordinates of the points that fail the verification are recorded. The values ​​are stored in the anomaly list, generating local residual data.

[0040] The deviation amplitude calculation submodule extracts pixel coordinate information based on local residual data, performs weighted calculations according to coordinate distribution, calculates spatial centroid coordinates, and combines the corresponding deviation amplitude to obtain spatial distribution data of deviation amplitude. Load local residual data, which contains discretely distributed information about anomalous pixels. To accurately obtain the geometric center and intensity distribution of the anomalous region, extract the coordinates. The core is a connected pixel cluster, which contains three adjacent pixels. , , The corresponding residual amplitudes are respectively , , Based on the coordinate distribution, a weighted centroid calculation is performed, and the formula for calculating the spatial centroid is defined as follows: and The deviation amplitude is introduced here. As a weighting parameter, its purpose is to make the calculated centroid more biased towards the pixel position with more severe deformation. Substituting the values ​​into the calculation, the total weight is first calculated. Next, calculate the sum of the weighted products of the row coordinates. Calculate the sum of the weighted products of the column coordinates. Finally, the centroid coordinates are obtained. , The calculated floating-point coordinates Mark the spatial center of the anomalous region and the maximum amplitude within the cluster. The regional intensity peaks are recorded, the calculation results of the anomaly clusters are organized, and the spatial distribution data of the deviation amplitude is generated.

[0041] The abnormal region generation submodule, based on the spatial distribution data of deviation amplitude, calls the coordinate sequence of edge pixels, performs spatial mapping operation according to spatial adjacency and geometric relationship, calculates the spatial boundary of the corresponding region by combining the edge pixel coordinates, performs region stitching processing, and generates local alignment abnormal regions. Based on the centroid coordinates determined from the spatial distribution data of the deviation amplitude, it is necessary to establish a spatial mapping relationship between the feature map grid and the physical edge coordinate sequence of the display screen, and call the sorted edge pixel coordinate sequence. Set the mapping ratio from the map grid to the physical coordinates as follows: (Right now Each grid unit corresponds to (physical pixel distance), for the calculated centroid coordinates Perform coordinate reconstruction to obtain the physical mapping center. With this physical center as the origin, the neighborhood correlation radius is set. Pixels, traversal Mid-edge point coordinates Calculate its relationship with The Euclidean distance is used to filter out those with a distance less than 1. The set of edge points, assuming the filtering results cover the edge index from to The section (the physical coordinates of which cover the area mentioned in paragraph 4) (equal points), extract the spatial extrema of the edge point set, determine the physical boundary box of the abnormal region, and the specific values ​​are shown in Table 4, based on the boundary coordinates. and Generate corresponding rectangular bounding boxes in the digital twin model of the display screen. If there are multiple adjacent abnormal regions, the overlapping area of ​​the bounding boxes exceeds [a certain threshold]. Then, a geometric union operation is performed to stitch the regions together, and the final output is region data containing the exact physical range and deviation level, generating local alignment anomaly regions.

[0042] Table 4: Mapping Table of Local Alignment Anomaly Regions

[0043] As shown in Table 4, the table details the process of back-mapping the feature map's grid centroid to the display screen's physical coordinate system, including anomaly numbers. Corresponding to the centroid data calculated above, the generated bounding box clearly defines the area in the physical space that needs to be repaired or monitored.

[0044] Specifically, such as Figure 2 , 7 As shown, the attitude self-compensation inference module includes: The regional data receiving submodule acquires local alignment anomaly regions, collects regional centroid coordinates and offset amplitude data, performs consistency verification on centroid coordinates according to coordinate axis definition rules, performs dimensional consistency detection and effective interval filtering on offset amplitude, and generates anomaly region attitude basic dataset. First, the data transmission interface is activated to receive the generated local alignment anomaly region data. This data packet contains geometric description information of multiple anomaly clusters. The data packet is parsed line by line to extract the physical centroid coordinates of each anomaly cluster. and the corresponding maximum offset value The anomalous cluster in the example For example, extract coordinates With amplitude Abnormal clusters coordinates With amplitude The data consistency verification process is then initiated, setting a valid coordinate space range based on the physical size specifications of the display screen, and setting an upper limit for the horizontal coordinate. Vertical coordinate upper limit For abnormal clusters Execute logical judgment and If the result is true, a dimensional check is performed on the offset amplitude data to confirm that the value type is single-precision floating-point and has no unit appendage. Then, a valid interval filter is performed, and a valid lower limit for the amplitude is set. and upper limit The lower limit value is set based on the original noise data. quantiles, all those below All deviations are considered as sensor thermal noise and discarded, targeting abnormal clusters. amplitude ,check If the data is valid, it will be retained; if any noisy coordinates exist... Amplitude Therefore, Data deemed invalid is discarded. The received region data is then traversed, and the centroid coordinates and amplitude data that have undergone double verification are repackaged to create an index key value and generate a basic dataset of pose for abnormal regions.

[0045] The attitude input construction submodule extracts the centroid coordinates as spatial position parameters based on the attitude base dataset of the abnormal region, normalizes the offset magnitude, combines the position and magnitude into vector concatenation, performs consistency judgment based on the vector dimension standard, and generates the attitude inference input vector set. The goal of calling the anomaly region pose dataset is to transform discrete spatial points into a tensor format that can be recognized by neural networks or logic inference engines. First, it iterates through the valid anomaly points retained in the dataset and extracts the centroid coordinates. As a spatial location parameter, and for the offset magnitude Perform normalization by introducing a normalization formula. In the formula and Use the settings and This formula unifies the magnitude of deviations of different orders of magnitude to a uniform value through linear mapping. Within the interval, eliminate the impact of absolute numerical differences on subsequent weight calculations, and substitute into the outlier cluster. amplitude Calculations were performed, and the results were obtained. Substitute into the abnormal cluster amplitude Calculation Substitute into the abnormal cluster amplitude Calculation After normalization, a vector concatenation operation is performed to concatenate the spatial coordinates. With normalized amplitude Combine to construct a three-dimensional feature vector For abnormal clusters Generate vectors For abnormal clusters Generate vectors The standard for vector dimension is set as follows: The generated vectors are checked for dimension and confirmed to have no missing components. The vectors are then sequentially filled into the input matrix. The specific data construction is shown in Table 5. Finally, the attitude inference input vector set is generated.

[0046] Table 5: Data Table for Constructing Attitude Inference Input Vectors

[0047] As shown in Table 5, the table shows the transformation results from the original physical parameters to the standardized inference vectors. The normalized amplitude accurately reflects the relative weight of outliers in the overall deviation assessment. The constructed vector set provides a standardized input format for subsequent global attitude calculation.

[0048] The alignment determination generation submodule calculates the vector interval parameter based on the posture inference input vector set, classifies and marks the posture correction direction according to the distribution of the interval parameter, integrates the classification results based on the manual review information, summarizes the posture state and determination label, and generates the display screen alignment detection result. To read the attitude inference input vector set, it is necessary to analyze the spatial distribution relationship between the vectors to determine the overall offset trend of the display screen, calculate the vector interval parameter and the direction weighted moment, and define the global offset determination formula as follows: In the formula Set as the physical geometric center of the display screen. , The coordinates of the outlier point. To normalize the amplitude weights, this formula uses the torque principle to calculate the weighted offset vector sum of outliers relative to the screen center. Substituting the data from Table 5, the calculation is performed, targeting the vector... Calculate the position difference Weighted components For vectors ,calculate Weighted components For vectors ,calculate Weighted components Perform vector summation to calculate the total offset vector. , , result vector This indicates that the abnormal center of gravity is significantly shifted to the left (negative X) and slightly upward (negative Y). The attitude is classified based on the magnitude and orientation angle of the resulting vector, and a translation threshold is set. ,because The main defect type is determined to be "horizontal leftward shift", and "slight upward shift" is marked. The original feedback data of the manual review interface is retrieved, and the review record with similar vector features is matched as "confirmed leftward shift". The automatic calculation label and the manual record are integrated to generate the final status description code and generate the display screen alignment detection result.

[0049] Please see Figure 8 The display screen alignment detection method is performed based on the aforementioned display screen alignment detection system and includes the following steps: S1: Acquire the image of the border of the display screen on the rotating workstation, use the Canny edge detection algorithm to calculate the edge response, extract the edge pixels and perform spatial coordinate sorting to generate an edge pixel coordinate sequence; S2: Based on the edge pixel coordinate sequence and the preset reference coordinate sequence, calculate the displacement difference of the corresponding pixel point, perform differential calculation on the displacement difference of adjacent pixel points, and generate the display screen edge offset gradient field. S3: Map the gradient field of the display edge offset to a feature tensor, use a convolutional neural network model to perform channel weighting and classification, and generate a display alignment deviation feature map. S4: Extract pixel regions exceeding the preset deviation threshold based on the display screen alignment deviation feature map, calculate the spatial centroid coordinates and deviation amplitude of the pixel region, perform spatial mapping with the edge pixel coordinate sequence, and generate local alignment anomaly regions; S5: Call the local alignment anomaly area, construct the attitude input set based on the region centroid coordinates and offset amplitude, use the support vector machine model to analyze the attitude correction and perform manual re-judgment, and generate the display screen alignment detection result.

[0050] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of protection of the described technical solutions.

Claims

1. A display screen alignment detection system, characterized in that, The system includes: The image acquisition and analysis module acquires images of the border of the display screen on the rotating workstation, uses the Canny edge detection algorithm to calculate the edge response, extracts edge pixels, performs spatial coordinate sorting, and generates an edge pixel coordinate sequence. The gradient distribution calculation module calculates the displacement difference of corresponding pixels based on the edge pixel coordinate sequence and the preset reference coordinate sequence, performs differential calculation on the displacement difference of adjacent pixels, and generates the edge offset gradient field of the display screen. The intelligent visual detection module maps the gradient field of the display screen edge offset into a feature tensor, uses a convolutional neural network model for channel weighting and classification, and generates a display screen alignment deviation feature map. The local residual extraction module extracts pixel regions that exceed a preset deviation threshold based on the alignment deviation feature map of the display screen, calculates the spatial centroid coordinates and deviation magnitude of the pixel regions, performs spatial mapping with the edge pixel coordinate sequence, and generates local alignment anomaly regions. The attitude self-compensation inference module calls the local alignment anomaly region, constructs an attitude input set based on the region centroid coordinates and offset amplitude, uses a support vector machine model to analyze attitude correction and performs manual re-judgment, and generates display screen alignment detection results.

2. The display screen alignment detection system according to claim 1, characterized in that, The edge pixel coordinate sequence includes pixel position, edge identifier and spatial sorting order; the display edge offset gradient field includes offset amplitude, angle gradient and displacement change rate; the display alignment deviation feature map includes classification label, feature weight and detection reliability; the local alignment anomaly region includes anomaly position, deviation value and regional deviation direction; and the display alignment detection result includes correction magnitude, re-judgment decision and pass / fail judgment result.

3. The display screen alignment detection system according to claim 1, characterized in that, The image acquisition and analysis module includes: The image acquisition submodule acquires frame sequences by acquiring images of the border of the display screen on the rotating workstation, performs pixel array reading and grayscale value mapping on the frame images, records row and column indices according to the imaging scanning order, performs order verification on the frame sequences and aggregates them to generate a border grayscale image matrix. The edge response submodule, based on the grayscale image matrix of the border, uses the Canny algorithm to calculate the grayscale difference between adjacent pixels, performs difference calculations in the horizontal and vertical directions respectively, synthesizes the difference results, classifies pixels according to the amplitude distribution and records the row and column coordinates, and generates an edge response pixel set. The coordinate sorting submodule calls the edge response pixel set, extracts the row and column coordinate values ​​corresponding to the pixels, performs ascending row coordinate sorting on the coordinate set while maintaining the consistency of the column coordinate association, and performs sequential combination of the sorting results to generate an edge pixel coordinate sequence.

4. The display screen alignment detection system according to claim 3, characterized in that, The gradient distribution calculation module includes: The displacement difference acquisition submodule obtains the row and column coordinates of the corresponding pixel points based on the edge pixel coordinate sequence and the preset reference coordinate sequence. It performs coordinate matching for pixels with the same index and generates a pixel displacement difference vector set by subtracting the edge pixel row and column coordinates from the reference row and column coordinates. The differential sequence construction submodule calls the pixel displacement difference vector set, obtains the adjacent pixel index based on the edge pixel coordinate sequence, performs differential operation on the displacement components under the adjacent index, binds the differential result with the index position and arranges them in order to generate displacement difference sequence data; The gradient field generation submodule obtains the row vector difference and column vector difference corresponding to the pixel index based on the displacement difference sequence data, constructs a grid according to the coordinate mapping relationship, fills the grid nodes with difference components and verifies the index consistency, organizes the node difference distribution relationship, and generates the display screen edge offset gradient field.

5. The display screen alignment detection system according to claim 4, characterized in that, The intelligent visual inspection module includes: The gradient mapping submodule maps the gradient field offset of the display edge to a feature tensor, obtains the row and column gradient values ​​corresponding to the pixel, writes the gradient values ​​into the tensor channel according to the pixel index consistency, verifies the tensor dimension and numerical integrity, and generates the edge gradient feature tensor. The convolutional feature extraction submodule calls the edge gradient feature tensor, performs convolutional sliding operation on the pixel region in the tensor, records the positional convolutional response value, aggregates the response according to different scales, rearranges the channel index of the aggregation result, and generates a multi-scale spatial feature matrix. Note that the discrimination generation submodule obtains channel feature values ​​based on the multi-scale spatial feature matrix, calculates weight coefficients and performs weighted superposition to form a discrimination vector, performs spatial index mapping and numerical normalization on the discrimination vector, and generates a display screen alignment deviation feature map.

6. The display screen alignment detection system according to claim 5, characterized in that, The step of writing gradient values ​​to tensor channels based on pixel index consistency means that gradient value writing is performed only when the pixel row index and column index simultaneously meet the preset edge index range. The preset edge index range is the area that is no more than a fixed pixel width from the physical edge of the display screen.

7. The display screen alignment detection system according to claim 5, characterized in that, The local residual extraction module includes: The residual data extraction submodule collects pixel deviation data based on the alignment deviation feature map of the display screen, compares the deviation value with a preset deviation threshold, filters pixel areas that exceed the threshold, records the pixel coordinates and deviation magnitude, and generates local residual data. The deviation amplitude calculation submodule extracts pixel coordinate information based on the local residual data, performs weighted calculations according to the coordinate distribution, calculates the spatial centroid coordinates, and combines the corresponding deviation amplitude to obtain the spatial distribution data of the deviation amplitude. The abnormal region generation submodule, based on the spatial distribution data of the deviation amplitude, calls the edge pixel coordinate sequence, performs spatial mapping operation according to spatial adjacency and geometric relationship, calculates the spatial boundary of the corresponding region in combination with the edge pixel coordinate, performs region stitching processing, and generates a local alignment abnormal region.

8. The display screen alignment detection system according to claim 7, characterized in that, The preset deviation threshold is based on the multi-frame alignment deviation feature map obtained by the display screen in the calibration state. The pixel deviation data in each frame alignment deviation feature map is statistically processed to calculate the mean and dispersion parameters of the pixel deviation data. The deviation statistical interval is constructed based on the mean and dispersion parameters, and the deviation value corresponding to the upper limit of the deviation statistical interval is set as the preset deviation threshold.

9. The display screen alignment detection system according to claim 7, characterized in that, The attitude self-compensation inference module includes: The regional data receiving submodule acquires the local alignment anomaly region, collects the region centroid coordinates and offset amplitude data, performs consistency verification on the centroid coordinates according to the coordinate axis definition rules, performs dimensional consistency detection and effective interval filtering on the offset amplitude, and generates an abnormal region attitude basic dataset. The attitude input construction submodule extracts the centroid coordinates as spatial position parameters based on the attitude base dataset of the abnormal region, performs normalization processing on the offset amplitude, combines the position and amplitude to form a vector concatenation, performs consistency judgment according to the vector dimension standard, and generates an attitude inference input vector set. The alignment determination generation submodule calculates the vector interval parameter based on the posture inference input vector set, classifies and marks the posture correction direction according to the distribution of the interval parameter, integrates the classification results based on the manual review information, summarizes the posture state and determination label, and generates the display screen alignment detection result.

10. A method for detecting the alignment of a display screen, characterized in that, The display screen alignment detection system according to any one of claims 1-9 includes the following steps: S1: Acquire the image of the border of the display screen on the rotating workstation, use the Canny edge detection algorithm to calculate the edge response, extract the edge pixels and perform spatial coordinate sorting to generate an edge pixel coordinate sequence; S2: Based on the edge pixel coordinate sequence and the preset reference coordinate sequence, calculate the displacement difference of the corresponding pixel point, perform differential calculation on the displacement difference of adjacent pixel points, and generate the display screen edge offset gradient field. S3: Map the edge offset gradient field of the display screen into a feature tensor, and use a convolutional neural network model to perform channel weighting and classification to generate a display screen alignment deviation feature map. S4: Based on the alignment deviation feature map of the display screen, extract the pixel region that exceeds the preset deviation threshold, calculate the spatial centroid coordinates and deviation magnitude of the pixel region, perform spatial mapping with the edge pixel coordinate sequence, and generate a local alignment anomaly region. S5: Call the local alignment anomaly region, construct an attitude input set based on the region centroid coordinates and offset amplitude, use a support vector machine model to analyze the attitude correction and perform manual re-judgment, and generate the display screen alignment detection result.

Images