Miniled display screen dead pixel detection positioning method, device and system based on deep learning

Abstract

The application provides a Miniled display screen dead pixel detection and positioning method, device and system based on deep learning. A preset driving signal sequence is applied to a pixel array of a Miniled display screen row by row, brightness feedback data of each pixel in different driving stages is obtained, a brightness response sequence is obtained, the brightness response sequence is spatially reorganized according to a physical arrangement structure of the pixel array, a brightness response matrix is generated, local area feature difference analysis is performed on the brightness response matrix, a brightness response abnormal area is identified, a candidate dead pixel position set is generated, a pre-trained dead pixel identification network is called to perform feature learning processing on the candidate dead pixel position set, a dead pixel probability corresponding to each candidate dead pixel position is output, the candidate dead pixel position set is screened based on this, candidate positions with a dead pixel probability lower than a preset condition are removed, and a dead pixel positioning result set is generated. The application can effectively improve the accuracy and reliability of Miniled display screen dead pixel detection and positioning.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and in particular to a method, apparatus and system for detecting and locating defective pixels on a Miniled display screen based on deep learning. Background Technology

[0002] With the continuous development of display technology, Miniled displays, with their advantages of high brightness, high contrast, and fine display effects, are widely used in various display devices. However, dead pixel detection, as a key quality control step in the Miniled display manufacturing process, directly affects the display effect and user experience. Currently, common methods for detecting dead pixels in Miniled displays typically involve applying a driving signal with fixed parameters to the pixel array, acquiring single-shot brightness feedback data for each pixel, and then determining whether a pixel is a dead pixel by setting a fixed brightness threshold. Alternatively, simple linear data sequences can be used to store brightness information and perform single-point comparison analysis. These traditional methods struggle to capture the dynamic response characteristics of dead pixels under different driving conditions, leading to the missed detection of some dead pixels that only exhibit abnormalities in specific driving phases. This makes it difficult to effectively distinguish between real dead pixels and false anomalies, resulting in insufficient accuracy and reliability of the detection results. Summary of the Invention

[0003] In view of this, the present invention provides a method, apparatus, and system for detecting and locating defective pixels on a minimized display screen based on deep learning. The technical solution of the embodiments of the present invention is implemented as follows:

[0004] On one hand, embodiments of the present invention provide a method for detecting and locating defective pixels in a minimized display screen based on deep learning. The method includes: applying a preset driving signal sequence to the pixel array of the minimized display screen row by row, obtaining brightness feedback data of each pixel at different driving stages, and obtaining a brightness response sequence containing pixel position information and corresponding brightness feedback data; spatially reorganizing the brightness response sequence according to the physical arrangement structure of the pixel array to generate a brightness response matrix, wherein the row dimension and column dimension of the brightness response matrix correspond to the horizontal pixel index and vertical pixel index of the display screen, respectively; performing local region feature difference analysis on the brightness response matrix, identifying brightness response abnormal regions that meet preset abnormal conditions, and generating a candidate defective pixel location set, wherein the candidate defective pixel location set contains multiple pixel coordinates with abnormal brightness response features; calling a pre-trained defective pixel recognition network to perform feature learning processing on the candidate defective pixel location set, outputting the defective pixel probability corresponding to each candidate defective pixel location, wherein the defective pixel probability is used to represent the probability that the candidate defective pixel location is a real defective pixel; filtering the candidate defective pixel location set according to the defective pixel probability, removing candidate locations with defective pixel probabilities lower than preset conditions, and generating a defective pixel location result set containing real defective pixel coordinates.

[0005] On the other hand, embodiments of the present invention provide a detection and positioning device, comprising:

[0006] The data acquisition module is used to apply a preset driving signal sequence to the pixel array of the Miniled display line by line, acquire the brightness feedback data of each pixel at different driving stages, and obtain a brightness response sequence containing pixel position information and corresponding brightness feedback data.

[0007] The spatial recombination module is used to spatially recombine the brightness response sequence according to the physical arrangement structure of the pixel array to generate a brightness response matrix. The row dimension and column dimension of the brightness response matrix correspond to the horizontal pixel index and vertical pixel index of the display screen, respectively.

[0008] The difference analysis module is used to perform local region feature difference analysis on the brightness response matrix, identify the brightness response abnormal regions that meet the preset abnormal conditions, and generate a candidate bad pixel location set. The candidate bad pixel location set contains the coordinates of multiple pixels with abnormal brightness response characteristics.

[0009] The feature learning module is used to call the pre-trained bad pixel recognition network to perform feature learning processing on the candidate bad pixel location set, and output the bad pixel probability corresponding to each candidate bad pixel location. The bad pixel probability is used to represent the probability that the candidate bad pixel location is a real bad pixel.

[0010] The defective pixel location module is used to filter the set of candidate defective pixel locations based on the defective pixel probability, remove candidate locations with defective pixel probabilities lower than preset conditions, and generate a set of defective pixel location results containing the actual defective pixel coordinates.

[0011] Thirdly, embodiments of the present invention provide a computer system including a memory and a processor. The memory stores a computer program that can run on the processor, and the processor executes the program to implement the steps in the above method.

[0012] This invention captures the dynamic response process of defective pixels throughout the complete driving cycle by applying a preset driving signal sequence row by row to a pixel array and acquiring the brightness feedback data of each pixel at different driving stages. This overcomes the limitation of traditional single-excitation, single-sampling methods that only record steady-state brightness values, providing a more comprehensive data foundation for subsequent accurate identification. Furthermore, by spatially reorganizing the brightness response sequence according to the physical arrangement of the pixel array to generate a brightness response matrix, a direct mapping between physical space and data structure is constructed, ensuring a one-to-one correspondence between the data structure and the physical location of pixels. This provides a structured data carrier for analyzing local feature differences, avoiding the cumbersome process of additionally associating physical locations required for conventional linear sequence storage. The system employs several techniques: First, it analyzes local feature differences in the brightness response matrix to identify abnormal areas. Second, it distinguishes between real defective pixels and noise interference through regional correlation analysis, overcoming the problem of misjudgment caused by interference in traditional thresholding methods. Third, it calls a pre-trained defective pixel identification network to learn features from the candidate defective pixel location set and output defective pixel probabilities, realizing a hierarchical detection logic of coarse screening and fine judgment. Fourth, it enhances decision robustness by quantifying defective pixel confidence through probability values. Fifth, it generates a defective pixel location result set based on defective pixel probabilities and dynamically adjusts the screening conditions according to probability statistics to ensure that the final result only includes high-confidence real defective pixels, effectively improving the accuracy and reliability of defective pixel detection and location for Miniled displays. Attached Figure Description

[0013] Figure 1 This is a schematic diagram illustrating the implementation process of a deep learning-based method for detecting and locating defective pixels on a minimized display screen, as provided in an embodiment of the present invention.

[0014] Figure 2 This is a schematic diagram of the composition of a detection and positioning device provided in an embodiment of the present invention.

[0015] Figure 3 This is a schematic diagram of the hardware entity of a computer system provided in an embodiment of the present invention. Detailed Implementation

[0016] This invention provides a deep learning-based method for detecting and locating defective pixels on a minimized display screen. This method can be executed by a computer system's processor. The computer system can refer to devices with data processing capabilities, such as servers, laptops, tablets, desktop computers, and distributed computers.

[0017] Figure 1 This is a schematic diagram illustrating the implementation process of a deep learning-based method for detecting and locating defective pixels on a minimized display screen, as provided in an embodiment of the present invention. Figure 1 As shown, the method includes:

[0018] Step S100: Apply a preset driving signal sequence to the pixel array of the Miniled display line by line, obtain the brightness feedback data of each pixel at different driving stages, and obtain a brightness response sequence containing pixel position information and corresponding brightness feedback data.

[0019] The preset drive signal sequence is a series of electrical signals pre-set according to specific rules and parameters based on the characteristics and detection requirements of the Miniled display screen. These signals have parameters such as timing, amplitude, and frequency, and their purpose is to drive the pixel array of the Miniled display screen, causing the pixels to exhibit different brightness states at different driving stages. After receiving the drive signal, each pixel generates corresponding brightness feedback based on its own physical characteristics. This feedback data reflects the actual light emission of the pixel at each driving stage. Pixel position information is used to identify the specific position of each pixel in the display screen's pixel array. It can be represented in the form of two-dimensional coordinates, such as horizontal row index and vertical column index, which allows for accurate positioning of each pixel on the display screen.

[0020] In actual operation, a preset sequence of driving signals is applied row by row to the pixel array of the minimized display based on the driving circuit. The driving circuit can be implemented using a programmable logic device (such as an FPGA). By programming the FPGA, the timing, amplitude, and frequency of the signals can be precisely controlled to ensure that the driving signals are applied to the pixel array according to the preset requirements. Simultaneously with the application of the driving signals, a high-sensitivity brightness sensor is used to acquire the brightness feedback data of each pixel. The brightness sensor can be a photodiode based on the photoelectric effect or a CMOS image sensor, which converts the light signal into an electrical signal, and then processes it through analog-to-digital conversion to obtain digitized brightness feedback data. Combining the position information of each pixel with the corresponding brightness feedback data in a certain order yields the brightness response sequence. For example, the pixels can be arranged sequentially according to their row and column indices, combining the position information of each pixel with the brightness feedback data into an ordered sequence.

[0021] Step S200: The brightness response sequence is spatially reorganized according to the physical arrangement structure of the pixel array to generate a brightness response matrix. The row dimension and column dimension of the brightness response matrix correspond to the horizontal pixel index and vertical pixel index of the display screen, respectively.

[0022] In one implementation, step S200 may specifically include the following steps S210 to S260:

[0023] Step S210: Parse the pixel position information in the brightness response sequence, extract the physical driving line path identifier and array coordinate code of each pixel, and convert the driving line path identifier into the initial horizontal index and the initial vertical index through the preset line path-coordinate mapping rules to generate the original pixel association table containing the original coordinates and brightness feedback data.

[0024] Pixel position information includes crucial information such as the physical connection path of the pixel in the driving circuit and its position code in the pixel array. The physical driving line path identifier is a number or code used to uniquely identify the driving line to which the pixel is connected, reflecting the pixel's specific position and connection relationship in the driving circuit. Different pixels may be connected to different driving lines; the physical driving line path identifier allows for accurate determination of the pixel's position in the driving circuit. Array coordinate encoding is an encoded representation of the pixel's position in the pixel array, using, for example, binary or decimal numbers.

[0025] When parsing pixel position information, a string parsing algorithm can be used. If the pixel position information is stored as a string in a preset format, such as "line path identifier - array coordinate code", a string splitting function can be used to split it into two parts: the physical driving line path identifier and the array coordinate code. Then, according to the preset line path-coordinate mapping rules, the physical driving line path identifier is converted into initial horizontal and initial vertical indices by querying a mapping table or performing preset calculations. This mapping rule can be stored in a data table or database, and the conversion is achieved through lookup and matching. For example, the mapping table records the initial horizontal and initial vertical indices corresponding to each physical driving line path identifier. When a physical driving line path identifier is obtained, the corresponding index can be looked up in the mapping table. Finally, the original coordinates (initial horizontal and initial vertical indices) are combined with the corresponding brightness feedback data to obtain the original pixel association table. This table can be stored in tabular form, with each row representing a pixel and containing the pixel's original coordinates and brightness feedback data.

[0026] Step S220: Obtain the physical topology map of the pixel array of the Miniled display screen, identify the row driving line distribution and column driving line distribution, determine the actual arrangement order of pixels in each row and the actual arrangement order of pixels in each column, and establish a physical arrangement calibration model.

[0027] A pixel array physical topology diagram is a graphical representation used to describe the connection relationships between pixels and drive lines in a pixel array of a minimized display. It shows the distribution of row and column drive lines, and how each pixel is connected to these lines. Row and column drive line distributions refer to the layout and orientation of row and column drive lines in the pixel array, respectively. Understanding these distributions is crucial for accurately determining the actual pixel arrangement order. The actual arrangement order of pixels in each row and column is the true order in which pixels are arranged in each row and column, which may differ from the initial index order. This is because the actual physical connections and wiring methods can cause some differences in pixel arrangement. A physical layout calibration model is a model used to convert initial coordinate indices into actual physical position coordinates. It considers the actual physical layout structure of the pixel array. This model can be used to calibrate the initial coordinates, obtaining more accurate pixel position information.

[0028] In one implementation, step S220 may specifically include the following steps S221 to S226:

[0029] Step S221: Call the display screen physical parameter database, read the drive circuit design drawing data of the target Miniled display screen, and extract the circuit topology data containing row drive circuit number, column drive circuit number and corresponding pixel connection relationship.

[0030] The display screen physical parameter database stores various physical parameter information for the display screen, including drive circuit design drawings. The target Minimized display screen is the specific Minimized display screen for which dead pixel detection and location are required. The drive circuit design drawings record the layout and connection relationships of the display screen's drive circuits, including row drive circuit numbers, column drive circuit numbers, and connection information between each pixel and these circuits. Circuit topology data is key data extracted from the drive circuit design drawings, used to describe the connection relationships between circuits and pixels.

[0031] Step S222: Perform graph structure modeling on the line topology data, using row-driven line numbers and column-driven line numbers as nodes and pixel connection relationships as edges to construct a line-pixel association graph, where each edge attribute contains the connection order number of the pixel in the corresponding line.

[0032] Graph structure modeling is the process of representing and organizing line topology data in the form of a graph. This method can more intuitively show the connection relationships between lines and pixels. Row-driven line numbers and column-driven line numbers serve as nodes in the graph, representing different driving lines. Pixel connections are represented as edges in the graph, connecting row-driven line nodes and column-driven line nodes, indicating which row-driven lines and column-driven lines are interconnected through pixels.

[0033] In one implementation, step S222 may specifically include the following steps S2221 to S2225:

[0034] Step S2221: Parse the row-driven line records in the line topology data, extract the unique identifier number of each row-driven line as the row node ID, extract the unique identifier number of each column-driven line as the column node ID, and store them in the row node set respectively.

[0035] Parsing row-driven line records in the line topology data involves analyzing and processing relevant information about row-driven lines to extract useful information. Each row-driven line has a unique identifier, which can be a number, letter, or a combination thereof, accurately distinguishing different row-driven lines. Similarly, each column-driven line has a unique identifier. Row node IDs and column node IDs identify the row-driven and column-driven lines in the graph structure, respectively, and are stored in the row node set and column node set for subsequent graph structure construction.

[0036] For example, a parsing program is written to scan the line topology data row by row, identifying row-driven and column-driven line records. For each record in the line topology data, the row-driven line number is extracted as the row node ID, and the column-driven line number is extracted as the column node ID. String processing functions can be used to extract the numbering information from the record. For example, if the record format is "Row-driven line: L1, connecting pixels: P1, P2; Column-driven line: C1, connecting pixels: P1, P3", string splitting and matching functions can be used to extract "L1" as the row node ID and "C1" as the column node ID. These row node IDs and column node IDs are stored in a row node set and a column node set, respectively. The row node set and column node set can be implemented using data structures such as lists and sets, for example, a list can be used to store the row node IDs: ["L1", "L2", "L3"].

[0037] Step S2222: Initialize the line-pixel association graph as a directed graph structure. The node set of the directed graph structure is a joint node set composed of the row node set and the column node set. The edge set of the directed graph structure is initially an empty set.

[0038] A directed graph structure is a graph structure in which edges have direction, pointing from one node to another. This structure can more accurately represent the connections between lines and pixels. The line-pixel relationship graph is initialized as a directed graph structure, whose node set is a joint set of row and column nodes. This means that the nodes in the graph include all row-driven and column-driven lines. The edge set is initially empty, indicating that there are no edges in the graph at initialization; edges will be added later based on pixel connections.

[0039] Step S2223: Traverse the pixel connection records in the line topology data. Each record contains the row-driven line number, column-driven line number, pixel identifier, and the connection sequence number of the pixel in the corresponding row line and the connection sequence number in the corresponding column line.

[0040] Traversing the pixel connection records in the line topology data involves sequentially accessing and processing all records related to pixel connection relationships within the data. Each pixel connection record contains the row-driven line number, column-driven line number, pixel identifier, and the connection order number of that pixel in the corresponding row and column lines. This information is crucial for constructing the line-pixel relationship graph. For example, a loop structure is used to traverse the pixel connection records in the line topology data. For instance, for line topology data stored in a list, a for loop is used to sequentially access each record. During the traversal, the row-driven line number, column-driven line number, pixel identifier, and row and column connection order numbers are extracted from each record to prepare for subsequent edge creation.

[0041] Step S2224: Starting from the row node corresponding to the row-driven line number and ending at the column node corresponding to the column-driven line number, create a directed edge and add the pixel identifier, row connection sequence number, and column connection sequence number as the edge attribute data to the edge set of the line-pixel association graph.

[0042] After traversing each pixel connection record, the corresponding row and column nodes are located in the initialized line-pixel association graph based on the row-driven line number and column-driven line number in the record. Directed edges are created with the row node as the starting point and the column node as the ending point. The pixel identifier, row connection sequence number, and column connection sequence number are added to the edge set of the line-pixel association graph as attribute data. In this way, the edges not only represent the connection relationship between row-driven and column-driven lines but also contain relevant pixel information.

[0043] For example, given a pixel connection record "Row-driven line L1 connects pixel P1, column-driven line C1 connects pixel P1, P1's connection sequence number in L1 is 1, and its connection sequence number in C1 is 1", find the row node corresponding to row-driven line number L1 and the column node corresponding to column-driven line number C1 in the graph. Create directed edges starting from the row node and ending at the column node, and add the pixel identifier P1, row connection sequence number 1, and column connection sequence number 1 as attribute data to the edge set of the line-pixel association graph.

[0044] Step S2225: After completing the creation of edges for all pixel connection records, perform integrity verification on the line-pixel association graph. Check whether each row node and column node has at least one associated edge. If there is an isolated node, mark it as an abnormal line node and record the abnormal line node information.

[0045] An isolated node is a node in the graph that has no associated edges. These nodes may indicate erroneous or missing connections in the line topology data. An abnormal line node is a row or column node marked as abnormal. Recording information about abnormal line nodes is for subsequent troubleshooting and processing.

[0046] In one implementation, step S2225 may specifically include the following steps S22251 to S22256:

[0047] Step S22251: Traverse the set of row nodes in the line-pixel association graph. For each row node, count the number of edges that serve as its starting point, and record it as the row node edge count. If the row node edge count is zero, mark the row as an isolated row node.

[0048] Traversing the set of row nodes in the line-pixel relationship graph involves sequentially visiting and processing all row nodes in the graph. Counting the number of edges that serve as the starting point of each row node is used to determine whether that node has connections with other nodes. A row node with zero edges indicates that it has no associated edges and is marked as an isolated row node.

[0049] For example, a loop is used to traverse the set of row nodes in the line-pixel association graph. For each row node, the number of edges that serve as its starting point is counted using either the graph's adjacency matrix or adjacency list. An adjacency matrix is ​​a two-dimensional matrix where elements indicate whether edges exist between nodes; an adjacency list is a linked list structure that stores information about each node's neighboring nodes. For instance, when using an adjacency list, the linked list corresponding to a row node can be traversed, and the number of elements in the list can be counted, which represents the number of edges in the row node. If a row node has zero edges, it is marked as an isolated row node, and its node ID and corresponding line number are recorded.

[0050] Step S22252: Traverse the set of column nodes in the same way. For each column node, count the number of edges that serve as its endpoints and record it as the column node edge count. If the column node edge count is zero, mark the column as an isolated column node.

[0051] Similar to step S22251, iterate through the set of column nodes and count the number of edges that serve as endpoints for each column node. A column node with zero edge counts indicates that the node has no associated edges and is marked as an isolated column node.

[0052] Step S22253: Collect all isolated row nodes and isolated column nodes, record their node IDs and corresponding line numbers, and generate a list of abnormal line nodes.

[0053] After identifying isolated row and column nodes, the information of these nodes is collected and organized. The node ID and corresponding line number of each isolated row and column node are recorded, and this information is then compiled to generate a list of abnormal line nodes.

[0054] For example, data structures such as lists or tables can be used to store information about abnormal line nodes. For instance, a list can be used to store information for each abnormal line node, where each element is a tuple containing a node ID and a line number. Information for all isolated row nodes and isolated column nodes is then added to the list sequentially, ultimately generating a list of abnormal line nodes, which can be used for subsequent problem analysis and processing.

[0055] Step S22254: Calculate the proportion of isolated row nodes to the total number of row nodes and the proportion of isolated column nodes to the total number of column nodes. If either proportion exceeds a preset threshold, it is determined that the line topology data has an integrity failure.

[0056] Calculating the proportion of isolated row nodes and isolated column nodes relative to the total number of row nodes and column nodes is to assess the severity of isolated nodes in the line topology data. The preset threshold is a pre-defined percentage value used to determine if the line topology data exhibits integrity failure. If the proportion of isolated row nodes or isolated column nodes exceeds the preset threshold, it indicates a large number of isolated nodes in the line topology data, potentially suggesting erroneous or missing connections, thus classifying the line topology data as having integrity failure.

[0057] For example, first, the total number of row nodes and the total number of column nodes are counted, and then the number of isolated row nodes and isolated column nodes are calculated separately. The number of isolated row nodes is divided by the total number of row nodes to obtain the percentage of isolated row nodes, and the number of isolated column nodes is divided by the total number of column nodes to obtain the percentage of isolated column nodes. These two percentages are compared with preset thresholds. If either percentage exceeds the preset threshold, it is determined that the line topology data has an integrity failure.

[0058] Step S22255: Based on the list of abnormal line nodes and the proportional calculation results, determine the integrity level of the line topology data. The integrity level is used to guide whether it is necessary to reacquire the line topology data.

[0059] Based on the list of abnormal line nodes and the calculated proportions, the integrity of the line topology data is comprehensively evaluated to determine its integrity level. Integrity levels can be categorized into different grades, such as high, medium, and low, each indicating a different degree of data integrity. The integrity level guides whether it's necessary to reacquire the line topology data. A low integrity level indicates significant problems with the data, necessitating reacquisition to ensure the accuracy of subsequent analysis. For example, the integrity level can be determined based on factors such as the number, proportion, and distribution of abnormal line nodes. For instance, if the number of isolated nodes is small, the proportion does not exceed a preset threshold, and the distribution is relatively dispersed, the integrity level can be considered high; if the number of isolated nodes is large, the proportion exceeds the preset threshold, and they are concentrated in certain areas, the integrity level can be considered low. Based on the determined integrity level, a decision is made regarding whether to reacquire the line topology data.

[0060] Step S22256: If the integrity failure is determined, the process of re-acquiring the line topology data is triggered; otherwise, the subsequent processing continues and the abnormal node information is recorded in the system log.

[0061] If the line topology data is determined to be incomplete based on the integrity level, a re-acquisition process for the line topology data needs to be triggered. Re-acquiring the line topology data can be achieved by re-accessing the display screen's physical parameter database or by re-detecting it using a testing device, to obtain more accurate line topology data. If the line topology data is determined not to be incomplete, subsequent processing steps continue, and abnormal node information is recorded in the system log for later troubleshooting and analysis.

[0062] For example, appropriate processing is performed based on the integrity level determination result. If the integrity is determined to be invalid, the process of re-acquiring the line topology data is initiated, such as re-querying the database or restarting the detection equipment. If the integrity is determined to be normal, subsequent steps are performed, such as further analysis and processing of the line-pixel correlation diagram, and abnormal node information is recorded in the system log, which may include node ID, line number, abnormal condition description, etc.

[0063] Step S223: Traverse all row-driven line nodes in the line-pixel association graph, extract the pixel connection sequence number connected to each row node, sort the numbers in ascending order to obtain the actual pixel arrangement order corresponding to the row-driven line, and generate a row arrangement order table.

[0064] For example, a loop is used to traverse all row-driven line nodes in the line-pixel association graph. For each row node, the pixel connection order numbers connected to it are extracted using the graph's edge set information. These numbers can be stored in a list, and then the list is sorted, for example, using a sorting algorithm (such as bubble sort, quick sort, etc.) to arrange the numbers in the list in ascending order. The sorted list of numbers is then added to the row order table as the actual pixel arrangement order corresponding to that row-driven line. The row order table can be stored using a data structure such as a table or a dictionary; for example, a dictionary can be used where the keys are the row-driven line numbers and the values ​​are the corresponding lists of actual pixel arrangement orders.

[0065] Step S224: Traverse all column-driven line nodes in the same way, extract the pixel connection sequence number connected to each column node, sort the numbers in ascending order to obtain the actual pixel arrangement order corresponding to the column-driven line, and generate a column arrangement order table.

[0066] Similar to step S223, all column-driven line nodes are traversed in the same way. The pixel connection sequence number connected to each column node is extracted, and the numbers are sorted in ascending order to obtain the actual pixel arrangement order corresponding to that column-driven line. The column arrangement order table is a table that summarizes the actual pixel arrangement order corresponding to each column-driven line. Through this table, the actual arrangement of pixels in each column can be clearly understood.

[0067] Step S225: Based on the row arrangement order table and the column arrangement order table, establish a mapping function between the initial coordinate index and the actual arrangement order. The horizontal mapping function takes the initial horizontal index as input and outputs the calibrated horizontal coordinate index, and the vertical mapping function takes the initial vertical index as input and outputs the calibrated vertical coordinate index.

[0068] The initial coordinate index is a preliminary numbering of pixel positions without considering differences in the actual physical layout, while the actual arrangement order reflects the true arrangement of pixels on the display screen. By establishing a mapping function, the initial coordinate index can be converted into a calibrated coordinate index that better reflects the actual situation.

[0069] In practice, one approach to establishing mapping functions is to construct a lookup table. Taking a horizontal mapping function as an example, the information from the row order table is organized into a lookup table, where the key is the initial horizontal index and the value is the calibrated horizontal coordinate index. When an initial horizontal index is input, the corresponding calibrated horizontal coordinate index is directly looked up in the lookup table. For example, if the row order table shows that the pixel with an initial horizontal index of 1 is actually in the 3rd position, then the corresponding value for key 1 in the lookup table is 3. Another approach is to use mathematical function fitting. By analyzing the data patterns in the row and column order tables, a mathematical function is used to describe the relationship between the initial and calibrated coordinate indices. For example, if a linear relationship is found between the initial and calibrated horizontal coordinate indices, a linear function y = ax + b can be used for fitting, where x is the initial horizontal index, y is the calibrated horizontal coordinate index, and a and b are coefficients obtained through linear regression analysis of the data in the row order table.

[0070] For the vertical mapping function, lookup tables or mathematical functions can also be used for fitting. A suitable mapping relationship is determined by analyzing the data in the column arrangement table. If the actual arrangement of column pixels is complex, polynomial or piecewise functions can be used to describe the relationship between the initial vertical index and the calibrated vertical coordinate index.

[0071] When establishing a mapping function, boundary cases and outliers should also be considered. For example, when the initial input coordinate index exceeds the preset range, corresponding processing rules need to be defined, such as returning a default value or displaying an error message. Simultaneously, for any abnormal data that may exist in the row-order and column-order tables, preprocessing is required, such as removing erroneous data or correcting the data, to ensure the accuracy and reliability of the mapping function.

[0072] Step S226: Integrate the horizontal mapping function and the vertical mapping function into a physical layout calibration model. The model input is the initial horizontal index and the initial vertical index, and the output is the calibrated horizontal coordinate index and the calibrated vertical coordinate index. The model parameters include the mapping relationship data of the row arrangement order table and the column arrangement order table.

[0073] For example, the horizontal and vertical mapping functions can be encapsulated in a single model class, which includes an input interface, computational logic, and an output interface. The input interface receives the initial horizontal and vertical indices. The computational logic calls the horizontal and vertical mapping functions respectively based on the mapping relationship data of the row and column order tables. The output interface outputs the calibrated horizontal and vertical coordinate indices. In this way, the horizontal and vertical mapping functions are integrated into a complete physical layout calibration model.

[0074] Step S230: Input the initial horizontal index and initial vertical index from the original pixel association table into the physical layout calibration model, perform coordinate mapping correction according to the row driving line distribution and column driving line distribution, obtain the calibrated horizontal coordinate index and the calibrated vertical coordinate index, and update the original pixel association table to the pixel calibration association table.

[0075] The initial horizontal and vertical indices from the original pixel association table are input into the physical layout calibration model to calibrate the initial coordinates. The physical layout calibration model establishes a mapping relationship between the initial coordinates and the actual physical positions based on the row-driven and column-driven line distributions. This model can convert the initial coordinates into more accurate calibrated coordinates. Coordinate mapping correction based on the row-driven and column-driven line distributions is necessary because the actual physical connections and wiring methods may cause some differences in pixel arrangement, requiring calibration to obtain accurate coordinates. After obtaining the calibrated horizontal and vertical coordinate indices, they are updated in the original pixel association table, replacing the original initial horizontal and vertical indices, resulting in the pixel calibration association table.

[0076] For example, each row of the original pixel association table is traversed, and the initial horizontal and vertical indices are extracted. These are then passed as input to the physical layout calibration model. The model calculates the calibrated horizontal and vertical coordinate indices based on the mapping data between the row and column order tables. These calibrated coordinate indices are then updated to their corresponding positions in the original pixel association table, resulting in the final pixel calibration association table. The coordinate information in the pixel calibration association table more accurately reflects the actual position of the pixels on the display screen, providing a more reliable data foundation for subsequent analysis and processing.

[0077] Step S240: Based on the nominal resolution parameters of the Miniled display screen and the number of effective pixels in the physical topology diagram, determine the target row dimension parameters and target column dimension parameters of the brightness response matrix, where the target row dimension parameters are equal to the number of effective horizontal pixels and the target column dimension parameters are equal to the number of effective vertical pixels.

[0078] The effective pixel count in the physical topology diagram refers to the number of pixels that can function correctly in the actual physical connections and layout. Determining the target row and column dimension parameters of the luminance response matrix ensures that the size of the luminance response matrix matches the actual pixel layout of the display screen. The target row dimension parameter equals the effective horizontal pixel count, and the target column dimension parameter equals the effective vertical pixel count. This ensures that the luminance response matrix accurately reflects the luminance information of each effective pixel on the display screen.

[0079] Step S250: Create a blank brightness response matrix. The number of rows in the matrix is ​​the target row dimension parameter, and the number of columns is the target column dimension parameter. The initial values ​​of the matrix elements are set to preset invalid values. Traverse each pixel entry in the pixel calibration association table, locate the filling position in the blank matrix according to the calibrated horizontal coordinate index and vertical coordinate index, and write the corresponding brightness feedback data into the filling position.

[0080] The blank luminance response matrix is ​​a matrix initially filled with no actual data. Its size is determined by the target row and column dimension parameters, ensuring the matrix matches the actual pixel layout of the minimized display. A preset invalid value is a special value that indicates the matrix element initially lacks valid luminance feedback data; this could be a flag value, such as -1, or another value outside the actual luminance data range. When traversing the pixel calibration association table, each pixel entry contains the calibrated horizontal and vertical coordinate indices, along with the corresponding luminance feedback data. These coordinate indices allow for precise location of the filling position within the blank luminance response matrix, where the corresponding luminance feedback data is then written.

[0081] Step S260: After filling all pixel entries, perform neighborhood interpolation on the positions in the blank luminance response matrix that are still invalid values, calculate the filling value based on the luminance feedback data of adjacent valid pixels, and obtain the complete luminance response matrix.

[0082] For example, a neighborhood interpolation algorithm is used to process the positions in the blank luminance response matrix that are still invalid values. Common neighborhood interpolation algorithms include linear interpolation and bilinear interpolation. Taking linear interpolation as an example, for an invalid value position, its neighboring valid pixels are found. Based on the luminance feedback data of these valid pixels and their distances from the invalid value position, a weighted average is calculated as the fill value. When performing neighborhood interpolation, the matrix boundary needs to be considered. For invalid values ​​at boundary positions, special processing methods may be required, such as considering only internally adjacent valid pixels. After filling all invalid value positions, a complete luminance response matrix is ​​obtained, which contains the luminance feedback data of each pixel on the display screen, providing a complete data foundation for subsequent analysis.

[0083] Step S300: Perform local region feature difference analysis on the brightness response matrix, identify the brightness response abnormal regions that meet the preset abnormal conditions, and generate a candidate bad pixel location set. The candidate bad pixel location set contains the coordinates of multiple pixels with abnormal brightness response characteristics.

[0084] Local area feature difference analysis involves extracting and comparing features from different local areas within the luminance response matrix to identify feature differences between regions. Preset anomaly conditions are a series of pre-defined rules or thresholds used to determine whether the luminance response of a specific local area is abnormal. An abnormal luminance response region refers to a region in the luminance response matrix whose luminance characteristics differ significantly from the surrounding areas, satisfying the preset anomaly conditions. The candidate defective pixel location set is a collection of pixel coordinates with abnormal luminance response characteristics. These pixels may be defective pixels on the display screen, but further analysis and verification are required.

[0085] In one implementation, step S300 may specifically include the following steps S310 to S360:

[0086] Step S310: Divide the brightness response matrix into multi-granularity regions and set up three different analysis windows with different analysis granularities to capture local detail features, regional distribution features and global correlation features respectively. Each analysis window slides in the matrix at a preset step size to generate multiple sets of local region sequences.

[0087] Multi-granularity region partitioning involves dividing the luminance response matrix into different granularities to obtain feature information at different levels. Three different analysis windows with varying sizes and coverage areas are used to capture different types of features. The local detail window captures the luminance features of a single pixel and its immediate neighbors, reflecting pixel-level detail. The region distribution window captures the luminance distribution statistics of multi-pixel clusters with a coverage level higher than the local detail window, such as the mean and variance of luminance. The global correlation window captures the luminance correlation features of large-scale pixel clusters with a coverage level higher than the region distribution window, reflecting the overall luminance distribution trend of the display screen. The preset step size refers to the distance the analysis window slides in the matrix each time. When setting the step size, it is necessary to ensure that adjacent windows overlap to capture feature information more comprehensively. Generating multiple local region sequences involves arranging the regions covered by each analysis window in the matrix in a specific order. Each sequence contains the coordinate range and window size information of multiple local regions.

[0088] In one implementation, step S310 may specifically include the following steps S311 to S316:

[0089] Step S311: Based on the target row dimension parameter and target column dimension parameter of the brightness response matrix, preset the size specifications of the analysis window for three analysis granularities, so that the window size is odd to maintain central symmetry. The local detail window is used to capture the brightness features of a single pixel and its immediate neighbors, the regional distribution window is used to capture the brightness distribution statistical characteristics of multi-pixel clusters with a coverage level higher than that of the local detail window, and the global correlation window is used to capture the brightness correlation features of large-scale pixel clusters with a coverage level higher than that of the regional distribution window.

[0090] The size of the analysis window is preset based on the target row and column dimension parameters of the luminance response matrix to ensure that the analysis window can reasonably cover different regions within the matrix. Setting the window size to an odd number maintains central symmetry, allowing the window to capture the feature information of surrounding pixels more accurately by using the central pixel as a reference during feature extraction. The local detail window is smaller, focusing on a single pixel and its immediate neighbors to capture pixel-level luminance features such as luminance values ​​and variations. The regional distribution window is larger, covering clusters of multiple pixels to statistically analyze the luminance distribution characteristics of the cluster, such as the mean and variance of luminance, reflecting the overall distribution of pixel luminance within the region. The global correlation window is even larger, covering a wide range of pixel clusters to capture the luminance correlation features between these pixels, such as whether the luminance change trends are consistent across different regions.

[0091] Step S312: Calculate the sliding step size for each type of analysis window. The step size is set to ensure that there is an overlapping area between adjacent windows. The step size for the local detail window is set according to the minimum feature capture requirement, the step size for the regional distribution window is set according to the regional feature analysis requirement, and the step size for the global association window is set according to the global distribution analysis requirement.

[0092] The step size for each analysis window is calculated to control the distance the window slides within the luminance response matrix, ensuring overlap between adjacent windows to avoid missing important information during feature extraction. The step size for the local detail window is set based on the minimum feature capture requirement. Since local detail windows primarily capture the luminance features of individual pixels and their immediate neighbors, the step size needs to be relatively small to ensure sufficient detail is captured. The step size for the region distribution window is set based on the region's feature analysis requirements, taking into account the pixel distribution and feature variation within the region. The step size can be relatively larger, but it must still ensure some overlap between adjacent windows to accurately analyze the region's luminance distribution characteristics. The step size for the global correlation window is set based on the global distribution analysis requirements. Because the global correlation window covers a larger area, the step size can be set even larger to improve analysis efficiency, but it's still crucial to ensure overlap between adjacent windows to capture global luminance correlation features.

[0093] Step S313: Starting from the beginning position of the luminance response matrix, slide the local detail window, moving the step size pixels horizontally each time. When the horizontal slide reaches the matrix boundary, move the step size pixels vertically and reset the horizontal position to the starting column until the entire matrix is ​​covered. Record the starting and ending coordinates of each window to generate a sequence of local detail regions.

[0094] The local detail window is slid across the entire matrix starting from the initial position of the brightness response matrix to ensure coverage of every pixel. Each horizontal movement is a step size of pixels. When the horizontal movement reaches the matrix boundary, the vertical movement is also a step size of pixels, and the horizontal position is reset to the starting column. This sliding method guarantees that the window fully covers the matrix. Recording the start and end coordinates of each window accurately determines the location and extent of each local detail region. The generated sequence of local detail regions contains this coordinate information, providing a foundation for subsequent feature extraction and analysis.

[0095] For example, a loop structure is used to implement the sliding of the local detail window. Starting from the top left corner of the matrix (the starting position), the local detail window is placed at this position, and the starting and ending coordinates of the window are recorded. Then, the window is moved horizontally by a preset horizontal step, and the new starting and ending coordinates are recorded after each movement. When the horizontal position of the window reaches the matrix boundary, the vertical position of the window is moved down by a preset vertical step, and the horizontal position is reset to the starting column, continuing the horizontal sliding. This process is repeated until the window covers the entire matrix. Finally, all the recorded starting and ending coordinates are arranged into a sequence to generate a sequence of local detail regions.

[0096] Step S314: Using the same starting position and sliding method, slide the regional distribution window and the global association window respectively to generate the regional distribution region sequence and the global association region sequence. The coordinate range of each local region is determined by the starting coordinates of the window, the side length of the window, and the matrix boundary.

[0097] Sliding the region distribution window and the globally related window with the same starting position and sliding method is to maintain the same traversal order as the local detail window, facilitating simultaneous analysis of regions with different granularities. The coordinate range of each local region is determined by the window's starting coordinates, window side length, and matrix boundaries. The window's starting coordinates determine its initial position in the matrix, the window side length determines its size, and the matrix boundaries limit the maximum range of the window. During the sliding process, when the window exceeds the matrix boundaries, the excess portion is specially handled to ensure that the coordinate range of each local region is reasonable.

[0098] For example, starting from the beginning of the luminance response matrix, the regional distribution window and the global association window are slidable respectively. The same sliding method as the local detail window is used, i.e., moving a step size pixel horizontally each time. When the horizontal sliding reaches the matrix boundary, the vertical sliding step size pixel is moved and the horizontal position is reset to the starting column. For each window, its starting and ending coordinates are recorded, and the coordinate range of each local region is determined based on the window side length and the matrix boundary. During the sliding process, cases where the window exceeds the matrix boundary are handled, such as setting the luminance value of the pixels exceeding the boundary to the average of the existing pixel luminance values ​​within that window. Finally, the recorded starting, ending, and coordinate ranges of the regional distribution window and the global association window are organized into sequences to generate the regional distribution region sequence and the global association region sequence, respectively.

[0099] Step S315: Perform boundary verification on the region coordinates in each local region sequence. If the window slides to the edge of the matrix and causes some regions to exceed the matrix range, the pixel brightness value of the exceeding part is set to the average of the existing pixel brightness values ​​in the window, so that the size of each local region conforms to the preset window side length.

[0100] Boundary checks are performed on the coordinates of each local region sequence to ensure that the coordinate range of each local region is within the valid range of the matrix. When the window slides to the edge of the matrix, some regions may exceed the matrix range. In this case, special handling is required for the excess parts. Setting the pixel brightness value of the excess part to the average of the existing pixel brightness values ​​within the window ensures that the size of each local region conforms to the preset window side length, while maintaining the continuity and consistency of brightness information within the region. For example, the coordinates of each local region sequence are traversed, and the coordinate range of each region is checked to see if it exceeds the matrix boundary. If it is found that part of the coordinates of a region exceeds the matrix range, the brightness values ​​of the existing pixels within the window are extracted, and their average is calculated. Then, the brightness value of the excess pixel is set to this average value. In this way, boundary checks and processing are performed on each local region to ensure that the size and brightness information of each local region are reasonable.

[0101] Step S316: Store the local detail region sequence, the regional distribution region sequence, and the global related region sequence respectively. Each sequence contains the coordinate range of multiple local regions and the corresponding window size information to obtain the multi-granularity region division result.

[0102] For example, a suitable data structure (such as a list or dictionary) is used to store each sequence. For each sequence, the coordinate range and window size information of each local region are combined into a data element and added to the corresponding sequence. For instance, a list can be used to store the sequences of local detail regions, where each element is a tuple containing the coordinate range and window size information of the local detail region.

[0103] Step S320: For each local region in each set of local region sequences, extract the brightness time series features of the pixels in the region, including the rising edge change rate, falling edge change rate, duration of the stable phase and fluctuation amplitude of the brightness response curve, and construct the regional time series feature vector.

[0104] Extracting luminance time-series features for each local region within each set of local region sequences is to analyze the luminance changes of pixels within that region during different driving phases. The rising edge rate of the luminance response curve reflects the speed at which pixel luminance rises from its initial value to its peak, the falling edge rate reflects the speed at which pixel luminance falls from its peak to its final value, the duration of the stable phase indicates the length of time pixel luminance remains stable near its peak, and the fluctuation amplitude indicates the degree of luminance change within the stable phase. Constructing a regional temporal feature vector combines these features into a single vector to describe the luminance time-series characteristics of that local region, providing a basis for subsequent anomaly region identification.

[0105] In one implementation, step S320 may specifically include the following steps S321 to S326:

[0106] Step S321: Traverse each local region in the local region sequence, extract the brightness feedback data of all pixels in the local region from the brightness response matrix according to its coordinate range, and obtain the regional brightness data set, where each pixel corresponds to a set of brightness response data that changes with the driving stage.

[0107] Traversing each local region in the sequence allows for sequential processing of each region. Extracting the luminance feedback data of all pixels within a local region from the luminance response matrix based on its coordinate range is crucial for obtaining luminance information for those pixels at different driving stages. The resulting set of regional luminance data contains the luminance response data for each pixel, with each pixel corresponding to a set of luminance values ​​that vary with the driving stage. This data provides the foundation for subsequent feature extraction and analysis.

[0108] Step S322: For each pixel in the region luminance data set, arrange its luminance response data in the order of the driving stages to obtain the luminance response curve of the pixel. The order of the driving stages is determined according to the application order of the preset driving signal sequence.

[0109] Arranging the luminance response data of each pixel in the region luminance dataset according to the driving stage sequence is to organize the pixel luminance data in chronological order, forming a continuous luminance response curve. The driving stage sequence is determined by the application order of the preset driving signal sequence, because the luminance change of a pixel occurs with the application of the driving signal. Arranging the luminance data according to the application order of the driving signal can accurately reflect the luminance change process of the pixel.

[0110] Step S323: Identify the rising phase, the stable phase, and the falling phase of the brightness response curve. The rising phase is when the brightness rises from the initial value to the peak value, the stable phase is when the brightness remains near the peak value, and the falling phase is when the brightness falls from the peak value to the end value.

[0111] Identifying the rising, settling, and falling phases of the luminance response curve is crucial for analyzing the characteristics of different stages of pixel brightness change. The rising phase reflects the rapid increase in pixel brightness from its initial value to its peak, the settling phase indicates that the pixel brightness remains relatively stable near its peak, and the falling phase is the process of the pixel brightness gradually decreasing from its peak to its final value. Accurately identifying these phases is essential for subsequent calculations of features such as the rising edge rate of change, the falling edge rate of change, and the duration of the settling phase.

[0112] In one implementation, step S323 may specifically include the following steps S3231 to S3236:

[0113] Step S3231: Smooth the brightness response curve of the pixel to obtain a smoothed brightness response curve.

[0114] For example, a smoothing algorithm is used to process the luminance response curve. Common smoothing algorithms include moving average and Gaussian filtering. Taking the moving average method as an example, for each data point in the luminance response curve, the average value of a certain number of data points before and after it is calculated, and this average value is used as the smoothed value for that data point. In this way, the entire luminance response curve is smoothed to obtain the smoothed luminance response curve.

[0115] Step S3232: Extract all local maxima from the smoothed luminance response curve, determine the point with the largest value as the peak point, and record the luminance value of the peak point and the corresponding driving stage time, which are denoted as peak luminance and peak time.

[0116] Extracting all local maxima from the smoothed luminance response curve is to find the peak position in the curve. A local maximum is a point in the curve where the luminance value is greater than the luminance values ​​of its neighboring points. The point with the largest value is identified as the peak point because it represents the maximum pixel luminance. The luminance value of the peak point and the corresponding driving phase time are recorded, denoted as peak luminance and peak time, respectively. This information is crucial for subsequent analysis of the rising and falling phases of the luminance response curve.

[0117] Step S3233: Determine the initial value as the brightness value of the first data point of the smooth curve and the ending value as the brightness value of the last data point of the smooth curve. Calculate the lower limit of the peak threshold as a preset ratio of the peak brightness.

[0118] Determining the initial value as the brightness value of the first data point on the smoothed curve and the ending value as the brightness value of the last data point on the smoothed curve is to clearly define the starting and ending brightness of the brightness response curve. Calculating the lower limit of the peak threshold as a preset proportion of the peak brightness is to determine a threshold used to judge the boundary between the rising and stable phases of the brightness response curve.

[0119] Step S3234: Starting from the beginning of the curve, traverse to the peak time, determine the first driving stage time when the brightness value reaches the lower limit of the peak threshold as the starting point of the rising stage, and the ending point of the rising stage is the peak time. The rising stage is the stage from the rising starting point to the peak time.

[0120] The purpose of traversing the curve from its starting point to the peak time is to find the position where the brightness value reaches the lower limit of the peak threshold during the rise. The starting point of the rising phase is determined when the first brightness value reaches the lower limit of the peak threshold, because from this point onwards, the pixel brightness begins to rise rapidly. The ending point of the rising phase is the peak time. The rising phase is the stage from the starting point to the peak time, thus accurately defining the rising phase of the brightness response curve.

[0121] Step S3235: Starting from the peak moment, traverse backwards to determine the first driving phase moment where the brightness value is lower than the lower limit of the peak threshold as the end point of the stable phase. The stable phase is the phase from the peak moment to the end point of the stable phase.

[0122] The process of traversing backwards from the peak value is to find the point where the brightness value begins to decrease in the stable phase. The moment when the first driving phase value falls below the lower limit of the peak threshold is determined as the end of the stable phase, because from this moment on, the pixel brightness begins to decrease from near the peak. The stable phase is the period from the peak value to the end of the stable phase, thus accurately defining the stable phase of the brightness response curve.

[0123] Step S3236: Starting from the end of the stable phase, traverse backwards to the end of the curve. This segment is the descent phase, where the starting point of the descent phase is the end of the stable phase, and the end of the descent phase is the end time of the curve.

[0124] The process of traversing from the end of the stable phase to the end of the curve is to determine the descent phase of the luminance response curve. The starting point of the descent phase is the end of the stable phase because from this moment, the pixel brightness begins to decrease from its peak. The end of the descent phase is the end of the curve; this segment is the descent phase, thus accurately defining the descent phase of the luminance response curve.

[0125] Step S324: Select a first starting point and a first ending point during the rising phase. The first starting point is the moment when the brightness reaches the first proportion of the peak value, and the first ending point is the moment when the brightness reaches the second proportion of the peak value. Divide the first difference between the brightness of the first ending point and the brightness of the first starting point by the second difference between the moment of the first ending point and the moment of the first starting point, and determine the rising edge change rate by the division result.

[0126] Selecting a first start point and a first end point during the rising phase is for calculating the rising edge change rate. The first start point is the moment when the brightness reaches a first proportion of the peak value, and the first end point is the moment when the brightness reaches a second proportion of the peak value. These two points define an interval within the rising phase. The first difference between the brightness at the first end point and the brightness at the first start point is divided by the second difference between the time at the first end point and the time at the first start point. The result of this division is determined as the rising edge change rate, which reflects the speed at which the pixel brightness changes during the rising phase.

[0127] Step S325: Select a second starting point and a second ending point during the falling phase. The second starting point is the moment when the brightness drops from the peak to the third proportion, and the second ending point is the moment when it drops to the fourth proportion. Divide the third difference between the brightness of the second ending point and the brightness of the second starting point by the fourth difference between the second ending point and the second starting point. Determine the falling edge change rate by the result of the division.

[0128] The selection of a second starting point and a second ending point during the falling phase is for calculating the falling edge rate of change. The second starting point is the moment when the brightness decreases from the peak to the third proportion, and the second ending point is the moment when it decreases to the fourth proportion. These two points define an interval within the falling phase. The third difference between the brightness at the second ending point and the brightness at the second starting point is divided by the fourth difference between the second ending point and the second starting point. The result of this division is determined as the falling edge rate of change, which reflects the speed at which the pixel brightness changes during the falling phase.

[0129] Step S326: Determine the duration of the stable phase as the time interval from when the brightness reaches the fifth peak ratio to when it decreases to the end of that ratio; determine the fluctuation amplitude as the difference between the maximum and minimum brightness values ​​within the stable phase; use the rising edge rate of change, falling edge rate of change, duration of the stable phase, and fluctuation amplitude as the temporal features of the pixel, and take the average of the temporal features of all pixels in the region to obtain the regional temporal feature vector.

[0130] The duration of the stabilization phase is defined as the time interval from when the brightness reaches the fifth peak value to when it decreases to the end of that value. This is to accurately calculate the length of the stabilization phase. The difference between the maximum and minimum brightness values ​​within the stabilization phase is defined as the fluctuation amplitude, reflecting the degree of change in pixel brightness within the stabilization phase. The rising edge rate of change, falling edge rate of change, stabilization phase duration, and fluctuation amplitude are used as the temporal features of the pixel. The average of the temporal features of all pixels within the region is taken to obtain the regional temporal feature vector, which is used to describe the overall brightness time series characteristics of the local region.

[0131] Step S330: Calculate the spatial neighborhood association features of each local region. Taking the center pixel of the region as the reference, calculate the absolute value of the brightness difference between it and its surrounding neighboring pixels, and take the average value to obtain the spatial association index.

[0132] Calculating the spatial neighborhood association features of each local region is to analyze the degree of brightness correlation between pixels within the region. Using the center pixel of the region as a reference, the absolute value of the brightness difference between the center pixel and its surrounding neighboring pixels is calculated to measure the brightness difference between the center pixel and its neighbors. The average value is taken to obtain the spatial association index, which reflects the consistency of pixel brightness within the local region. For example, for each local region, the position of the center pixel is determined. Then, the neighboring pixels around the center pixel are identified, and the absolute value of the brightness difference between the center pixel and each neighboring pixel is calculated. These absolute values ​​of difference are summed and then divided by the number of neighboring pixels to obtain the spatial association index. A loop structure can be used to iterate through the neighboring pixels, calculate the absolute value of the brightness difference, and sum the results.

[0133] Step S340: Input the regional temporal feature vector and spatial correlation index into the preset abnormal region judgment model, and output the abnormality probability score of the local region. The abnormality probability score comprehensively considers the degree of deviation of the temporal feature from the normal range and the degree of abnormality of the spatial correlation index.

[0134] Inputting regional temporal feature vectors and spatial correlation indices into a pre-defined anomaly region determination model aims to assess the likelihood of anomalies in local areas. The anomaly region determination model is a pre-trained model that, based on the input regional temporal feature vectors and spatial correlation indices, comprehensively considers the degree to which the temporal features deviate from the normal range and the degree of anomaly in the spatial correlation indices, outputting an anomaly probability score for the local area.

[0135] Step S350: Calculate the anomaly probability score for the local region sequences divided by the three analysis granularity windows respectively, and then perform weighted fusion of the anomaly probability scores of the same pixel coordinate under different analysis granularities. The weights are preset according to the adaptability of window analysis granularity and pixel density.

[0136] Anomaly probability scores are calculated for local region sequences divided by three different analysis granularities to assess the anomaly probability of local regions from different granularity perspectives. Then, the anomaly probability scores for the same pixel coordinate under different analysis granularities are weighted and fused to comprehensively consider the analysis results of different granularities and obtain a more accurate anomaly probability assessment. The weights are preset based on the adaptability of the window analysis granularity to pixel density. Different analysis granularities have different capabilities in capturing pixel density; setting appropriate weights can make the fusion result more reasonable.

[0137] Step S360: Based on the weighted fusion anomaly probability score, set an anomaly score threshold, mark the pixel coordinates with scores higher than the threshold as candidate bad points, collect all marked pixel coordinates, remove duplicate coordinates, and generate a set of candidate bad point locations.

[0138] Setting an anomaly score threshold based on the weighted fusion anomaly probability score is to determine a standard for identifying which pixel coordinates might be defective. Pixel coordinates with scores higher than the threshold are marked as candidate defective pixels because these pixels have a higher probability of being defective. All marked pixel coordinates are collected, and duplicate coordinates are removed to generate a candidate defective pixel location set. This set contains the pixel coordinates that might be defective pixels, providing candidate objects for subsequent defective pixel identification.

[0139] Step S400: Call the pre-trained bad pixel recognition network to perform feature learning processing on the candidate bad pixel location set, and output the bad pixel probability corresponding to each candidate bad pixel location. The bad pixel probability is used to represent the probability that the candidate bad pixel location is a real bad pixel.

[0140] Calling a pre-trained defective pixel detection network to perform feature learning on the candidate defective pixel location set is to enable further analysis and judgment of these locations. The pre-trained network is trained on a large amount of data and can learn the features of candidate defective pixel locations, outputting the defective pixel probability for each location. This probability represents the likelihood that the candidate location is a true defective pixel, providing a basis for subsequent defective pixel selection.

[0141] In one implementation, step S400 may specifically include the following steps S410 to S460:

[0142] Step S410: For each candidate location in the candidate bad pixel location set, take its coordinates as the center and extract a three-dimensional neighborhood data block of a preset size from the brightness response matrix. The three dimensions are the horizontal pixel index, the vertical pixel index and the driving stage, respectively. The three-dimensional neighborhood data block contains the brightness feedback data of the candidate location and its surrounding pixels in different driving stages.

[0143] For each candidate location in the candidate defect location set, a three-dimensional neighborhood data block of a preset size is extracted from the luminance response matrix, centered on its coordinates. This is to obtain the luminance feedback data of the candidate location and its surrounding pixels at different driving stages. The three dimensions of the three-dimensional neighborhood data block are the horizontal pixel index, the vertical pixel index, and the driving stage, which can comprehensively describe the luminance changes of the candidate location and its surrounding pixels.

[0144] In one embodiment, step S410 may specifically include the following steps S411 to S417:

[0145] Step S411: Read a candidate position from the candidate bad spot position set and obtain its horizontal coordinate index and vertical coordinate index.

[0146] Reading a candidate location from the candidate defect location set is necessary to process each candidate location sequentially. Obtaining its horizontal and vertical coordinate indices is to determine the specific position of the candidate location in the brightness response matrix, providing a basis for subsequent 3D neighborhood data block extraction. For example, a loop structure is used to traverse the candidate defect location set, reading one candidate location at a time. Extracting the horizontal and vertical coordinate indices from the candidate location can be achieved using list or dictionary indexing operations.

[0147] Step S412: Determine the spatial dimensions of the three-dimensional neighborhood data block, wherein the horizontal and vertical dimensions are preset according to the requirements of capturing the correlation between pixels around the candidate position, and the driving stage dimension includes the entire application process of the driving signal sequence.

[0148] Determining the spatial dimensions of the 3D neighborhood data block is to clarify the size and range of the data block to be extracted. The horizontal and vertical dimensions are preset based on the requirement to capture the correlation between pixels surrounding candidate locations, taking into account the influence range of surrounding pixels on candidate locations to ensure sufficient pixel correlation information can be captured. The driving phase dimension encompasses the entire application process of the driving signal sequence, because pixel brightness changes occur with the application of the driving signal; including the entire driving phase provides a comprehensive description of pixel brightness changes.

[0149] Step S413: Calculate the intercept range of the spatial dimension. The horizontal start index is the horizontal coordinate index minus the number of surrounding pixels captured, and the horizontal end index is the horizontal coordinate index plus the number of surrounding pixels captured. The vertical start index and end index are calculated in the same way.

[0150] Calculating the spatial dimension's cut-off range is to determine the specific location of the three-dimensional neighborhood data block within the brightness response matrix. The horizontal start index is the horizontal coordinate index minus the number of surrounding pixels captured, and the horizontal end index is the horizontal coordinate index plus the number of surrounding pixels captured. The vertical start and end indices are calculated in the same way. This ensures that the cut-off three-dimensional neighborhood data block includes the candidate location and its surrounding pixels.

[0151] Step S414: The truncation range of the driving phase dimension is all phases from the initial driving phase to the end driving phase, ensuring that complete brightness feedback data of candidate positions are included throughout the driving process.

[0152] The truncation range for the driving phase dimension covers all stages from the initial driving phase to the final driving phase. This is to ensure that the truncated 3D neighborhood data block contains complete brightness feedback data for candidate locations throughout the driving process. Since pixel brightness changes occur with the application of the driving signal, including the entire driving phase provides a comprehensive description of pixel brightness variations. For example, the initial driving phase and the final driving phase are determined based on the application process of a preset driving signal sequence. These two phases are used as the truncation range for the driving phase dimension to ensure that the truncated data block contains brightness feedback data for all driving phases.

[0153] Step S415: Extract the submatrix defined by the adjusted start index and end index from the luminance response matrix as the original neighborhood matrix.

[0154] Extracting a submatrix defined by adjusted start and end indices from the luminance response matrix as the original neighborhood matrix is ​​to obtain luminance feedback data of candidate locations and their surrounding pixels at different driving stages. The adjusted start and end indices are determined after considering matrix boundary conditions to ensure that the submatrix is ​​within the effective range of the matrix.

[0155] Step S416: When the size of the original neighborhood matrix is ​​smaller than the preset neighborhood size, perform boundary filling processing on the original neighborhood matrix, using the brightness value of the neighborhood edge to fill the missing area, ensuring that the size of the filled matrix is ​​consistent with the preset neighborhood size.

[0156] When the size of the original neighborhood matrix is ​​smaller than the preset neighborhood size, boundary padding is performed on the original neighborhood matrix to ensure that the matrix size meets the preset requirements. Filling missing regions with the brightness values ​​of the neighborhood edges maintains the continuity and consistency of brightness information within the matrix. Ensuring that the size of the padded matrix matches the preset neighborhood size provides a unified data format for subsequent feature enhancement processing.

[0157] For example, the size of the original neighborhood matrix is ​​compared with a preset neighborhood size. If the size of the original neighborhood matrix is ​​smaller than the preset neighborhood size, the location and size of the missing region are determined. Then, the missing region is filled using the brightness values ​​of the neighborhood edges. A loop structure can be used to traverse the missing region, copying the brightness values ​​of the neighborhood edges to the corresponding positions in the missing region.

[0158] Step S417: Determine the filled neighborhood matrix as the three-dimensional neighborhood data block of the candidate bad pixel location, and keep the relative positional relationship of each element in the matrix consistent with the spatial relationship in the brightness response matrix.

[0159] The padded neighborhood matrix is ​​used to define the 3D neighborhood data block of the candidate defective pixel location so that the processed matrix can be used as input for subsequent feature enhancement processing. Maintaining the relative positions of the elements in the matrix consistent with the spatial relationships in the brightness response matrix ensures that the spatial information of the data is preserved, facilitating subsequent feature learning and analysis. For example, the padded neighborhood matrix is ​​assigned to the 3D neighborhood data block variable of the candidate defective pixel location. Ensuring the relative positions of the elements in the matrix consistent with the spatial relationships in the brightness response matrix can be achieved by maintaining the correct index and order during the padded and truncation processes.

[0160] Step S420: Perform feature enhancement processing on the three-dimensional neighborhood data block, perform differential operation along the driving stage dimension to obtain the brightness change rate data block, perform edge detection along the horizontal and vertical dimensions respectively to obtain the horizontal edge data block and the vertical edge data block, and concatenate the original data block, brightness change rate data block, horizontal edge data block and vertical edge data block according to the channel dimension to form multi-channel input data.

[0161] Feature enhancement processing of the 3D neighborhood data block is performed to extract more valuable feature information and improve the accuracy of defective pixel identification. Differential operations are performed along the driving stage dimension to obtain a brightness change rate data block, which reflects the rate of change of pixel brightness in different driving stages. Edge detection is performed along the horizontal and vertical dimensions respectively to obtain horizontal edge data blocks and vertical edge data blocks. These two data blocks can highlight the edge information of pixel brightness, which helps to identify the boundaries of defective pixels. Concatenating the original data block, brightness change rate data block, horizontal edge data block, and vertical edge data block along the channel dimension into multi-channel input data is to combine different types of feature information together, providing richer input for the subsequent defective pixel identification network.

[0162] For example, feature enhancement processing is performed on the three-dimensional neighborhood data blocks. For the brightness change rate data block, a difference algorithm can be used to calculate the brightness difference between adjacent driving stages. For the horizontal and vertical edge data blocks, edge detection algorithms (such as the Sobel operator) can be used for edge detection. Finally, a concatenation function is used to concatenate the original data block, the brightness change rate data block, the horizontal edge data block, and the vertical edge data block along the channel dimension to obtain multi-channel input data.

[0163] Step S430: Input the multi-channel input data into the feature extraction module of the pre-trained bad pixel recognition network, which contains multiple levels of convolutional layers. Each convolutional layer is followed by a batch normalization layer and an activation function. The first level of convolutional layer is used to extract local detail features using fine-grained 3D convolutional kernels, and the second level of convolutional layer is used to extract global context features using coarse-grained 3D convolutional kernels.

[0164] Multi-channel input data is fed into the feature extraction module of a pre-trained bad pixel detection network to process the input data using the network's feature extraction capabilities. The feature extraction module contains multiple convolutional layers, each followed by a batch normalization layer and an activation function. This structure enhances the network's feature representation capabilities. The initial convolutional layers use fine-grained 3D convolutional kernels to extract local detail features, capturing subtle features and local variations in the data; the subsequent convolutional layers use coarse-grained 3D convolutional kernels to extract global contextual features, grasping the overall structure and global information of the data.

[0165] In one implementation, step S430 may specifically include the following steps S431 to S436:

[0166] Step S431: Input the multi-channel input data into the first-level convolutional layer. This convolutional layer contains a preset number of fine-grained three-dimensional convolutional kernels to perform convolution operations on the input data and generate multiple feature maps. The size of the feature maps is the same as the spatial and driving stage dimensions of the input data.

[0167] Multi-channel input data is fed into the first-level convolutional layer to begin feature extraction. This convolutional layer contains a predetermined number of fine-grained 3D convolutional kernels, which capture local details in the input data. Convolution operations are performed on the input data to generate multiple feature maps. The size of these feature maps is the same as the spatial and driving stage dimensions of the input data, ensuring that the spatial and temporal structure of the data is not altered during feature extraction.

[0168] For example, multi-channel input data is passed to the first-level convolutional layer. Fine-grained 3D convolutional kernels in the convolutional layer perform convolution operations with the input data, generating multiple feature maps. Each convolutional kernel corresponds to one feature map, and the size of the feature map is the same as the spatial and driving stage dimensions of the input data.

[0169] Step S432: Perform batch normalization on the feature map output by the first-level convolutional layer, calculate the mean and variance of each feature channel, standardize the feature values ​​to a distribution with a mean of 0 and a variance of 1, and then perform nonlinear transformation through an activation function to enhance the feature representation capability.

[0170] Batch normalization of the feature map output from the first-level convolutional layer is performed to accelerate network convergence and improve model stability. The mean and variance of each feature channel are calculated, and the feature values ​​are standardized to a distribution with a mean of 0 and a variance of 1. This ensures that the feature values ​​across different channels have the same scale, facilitating subsequent processing. Then, a nonlinear transformation is performed using an activation function to enhance feature representation, enabling the network to learn more complex features. For example, a batch normalization layer is used to process the feature map output from the first-level convolutional layer. The batch normalization layer calculates the mean and variance of each feature channel and standardizes the feature values. Then, an activation function (such as ReLU) is used to perform a nonlinear transformation on the standardized feature values.

[0171] Step S433: Input the activated feature map into the second-level convolutional layer, which contains more fine-grained 3D convolutional kernels than the first-level convolutional layer. After the convolution operation, more feature maps are generated, batch normalized and activated, and the feature size remains unchanged.

[0172] The activated feature map is fed into the second-level convolutional layer to further extract local detail features from the input data. This convolutional layer contains a greater number of fine-grained 3D convolutional kernels than the first-level convolutional layer, enabling it to capture richer local feature information. The convolution operation generates a larger number of feature maps, increasing the feature dimensionality. Batch normalization and activation ensure that feature values ​​have the same scale and enhance feature expressiveness, while maintaining the feature size for easier subsequent processing. For example, the activated feature map is passed to the second-level convolutional layer. The fine-grained 3D convolutional kernels in the convolutional layer perform convolution operations with the input data, generating a larger number of feature maps. Then, a batch normalization layer normalizes the convolutional output, followed by a non-linear transformation using an activation function. These operations can be implemented using convolutional layers, batch normalization layers, and activation function layers from a deep learning framework.

[0173] Step S434: Input the output of the second-level convolutional layer into the third-level convolutional layer. This convolutional layer contains a preset number of coarse-grained three-dimensional convolutional kernels. After the convolution operation, multiple feature maps are generated, batch normalized and activated. The spatial and driving stage dimensions of the feature maps are adjusted according to the convolution stride.

[0174] The output of the second-level convolutional layer is fed into the third-level convolutional layer to begin extracting global contextual features from the input data. This convolutional layer contains a predetermined number of coarse-grained 3D convolutional kernels that capture the global structure and information in the input data. Multiple feature maps are generated after the convolution operation, batch normalized, and activated to ensure that feature values ​​have the same scale and enhance feature expressiveness. The spatial and driving stage dimensions of the feature maps are adjusted according to the convolution stride to reduce feature dimensionality and improve computational efficiency. For example, the output of the second-level convolutional layer is passed to the third-level convolutional layer. The coarse-grained 3D convolutional kernels in the convolutional layer perform convolution operations with the input data, generating multiple feature maps. The spatial and driving stage dimensions of the feature maps are adjusted according to the convolution stride. Then, the convolution output is normalized using a batch normalization layer and then subjected to a non-linear transformation through an activation function.

[0175] Step S435: Input the output of the third-level convolutional layer into the fourth-level convolutional layer, which contains more coarse-grained 3D convolutional kernels than the third-level convolutional layer. After the convolution operation, more feature maps are generated, batch normalized, and activated.

[0176] The output of the third-level convolutional layer is fed into the fourth-level convolutional layer to further extract global contextual features from the input data. This convolutional layer contains a greater number of coarse-grained 3D convolutional kernels than the third-level layer, enabling it to capture richer global feature information. The convolution operation generates a larger number of feature maps, increasing the dimensionality of the features. Batch normalization and activation ensure that feature values ​​have the same scale and enhance feature expressiveness.

[0177] For example, the output of the third-level convolutional layer is passed to the fourth-level convolutional layer. The coarse-grained 3D convolutional kernels in the convolutional layers perform convolution operations with the input data, generating a larger number of feature maps. Then, a batch normalization layer is used to normalize the convolutional output, followed by a non-linear transformation through an activation function.

[0178] Step S436: Use the feature map output by the fourth convolutional layer as the output of the feature extraction module. These feature maps contain multi-scale features under different receptive fields, where the first two convolutional layers extract local detail features and the last two convolutional layers extract global context features.

[0179] Using the feature map output from the fourth-level convolutional layer as the output of the feature extraction module is to summarize and output the features extracted by multiple convolutional layers. These feature maps contain multi-scale features under different receptive fields. The first two convolutional layers use fine-grained 3D convolutional kernels to extract local detail features, which can capture subtle features and local changes in the data; the last two convolutional layers use coarse-grained 3D convolutional kernels to extract global contextual features, which can grasp the overall structure and global information of the data.

[0180] Step S440: Input the high-dimensional feature map output by the feature extraction module into the attention fusion module of the network. The attention fusion module includes a spatial attention branch and a channel attention branch. The spatial attention branch generates a spatial weight map through global average pooling and convolution operations, and the channel attention branch generates a channel weight map through global max pooling and convolution operations. Multiply the two weight maps with the high-dimensional feature map to obtain the attention-enhanced feature map.

[0181] The high-dimensional feature map output from the feature extraction module is input into the network's attention fusion module to enhance the feature map using an attention mechanism. The attention fusion module includes spatial attention and channel attention branches, which focus on and weight the feature map from both spatial and channel dimensions, respectively. The spatial attention branch generates a spatial weight map through global average pooling and convolution operations, reflecting the importance of different spatial locations in the feature map. The channel attention branch generates a channel weight map through global max pooling and convolution operations, reflecting the importance of different channels in the feature map. Multiplying the two weight maps with the high-dimensional feature map yields an attention-enhanced feature map, which highlights important spatial locations and channels in the feature map, improving the expressive power of the features.

[0182] For example, the high-dimensional feature map output by the feature extraction module is passed to the attention fusion module. The spatial attention branch performs global average pooling on the feature map, averaging it across the spatial dimension to obtain a one-dimensional vector. This vector is then processed through convolution to generate a spatial weight map. The channel attention branch performs global max pooling on the feature map, max pooling it across the channel dimension to obtain a one-dimensional vector. This vector is then processed through convolution to generate a channel weight map. Finally, the spatial weight map and the channel weight map are multiplied by the high-dimensional feature map to obtain the attention-enhanced feature map.

[0183] Step S450: Input the attention-enhanced feature map into the temporal dependency modeling layer of the network. This temporal dependency modeling layer contains multiple bidirectional memory units, performs temporal dependency modeling on the feature map along the driving stage dimension, captures the dynamic features of brightness response changing with the driving stage, and outputs the temporal enhancement feature vector.

[0184] Inputting the attention-enhanced feature map into the network's temporal dependency modeling layer aims to further explore the temporal dependencies and dynamic features of the brightness response data. This temporal dependency modeling layer contains multiple bidirectional memory units, which can simultaneously consider past and future information, capturing the temporal features of the data more comprehensively compared to unidirectional memory units. Performing temporal dependency modeling on the feature map along the driving stage dimension means analyzing the correlation and changes of the feature map between different driving stages, using the driving stage as the time axis. This modeling approach captures the dynamic features of brightness response changes with the driving stage. The final output temporal enhanced feature vector is a feature representation obtained after temporal modeling of the input attention-enhanced feature map; it integrates information from the temporal dimension and better reflects the dynamic characteristics of pixel brightness changes.

[0185] For example, a temporal dependency modeling layer containing multiple bidirectional memory units is constructed using a deep learning framework. For instance, a Long Short-Term Memory (LSTM) network or a Gated Recurrent Unit (GRU) can be used as the bidirectional memory unit. The attention-enhanced feature map is split according to the driving stage dimension and sequentially input into the bidirectional memory unit. The bidirectional memory unit calculates and updates its state based on the current input and past and future information. After processing all driving stages, the hidden state at the last time step is output as the temporal enhanced feature vector. In this process, the bidirectional memory unit learns the dependencies and change patterns of the brightness response between different driving stages, thereby enhancing the temporal information representation capability of the feature vector.

[0186] Step S460: Input the temporal enhancement feature vector into the classification output layer of the network. The classification output layer combines low-level features into high-level semantic features through weighted combination and nonlinear transformation of multiple neurons, and finally outputs a single probability value, which is the probability of the bad point corresponding to the candidate bad point position.

[0187] The purpose of inputting the temporal enhancement feature vector into the classification output layer of the network is to classify whether a candidate bad pixel location is a true bad pixel based on the previously extracted and enhanced features. The classification output layer combines low-level features into high-level semantic features through weighted combinations and nonlinear transformations of multiple neurons. Low-level features refer to the relatively basic features extracted by the preceding modules, while high-level semantic features are obtained by further abstracting and combining these low-level features, resulting in more representative and discriminative features. In this way, the classification output layer can learn higher-level patterns and rules from complex features, thus achieving more accurate classification. The final output single probability value represents the probability that the candidate bad pixel location is a true bad pixel; this probability value can be used as a basis for subsequent bad pixel screening.

[0188] For example, the classification output layer is typically implemented using a fully connected layer. Each neuron in the fully connected layer is connected to every element in the input temporal augmentation feature vector. By weighted summing of these connections and then transforming the result through a non-linear activation function (such as the sigmoid function), the output value is mapped to a probability range of 0 to 1. The closer this probability value is to 1, the greater the probability that the candidate bad pixel location is a real bad pixel; the closer it is to 0, the less likely it is to be a real bad pixel. During training, using a large amount of labeled data, optimization algorithms (such as stochastic gradient descent) are used to adjust the weights of the neurons in the fully connected layer, enabling the network to learn the correct classification rules and improve the accuracy of bad pixel probability prediction.

[0189] Step S500: Filter the candidate bad point location set according to the bad point probability, remove candidate locations with bad point probabilities lower than preset conditions, and generate a bad point location result set containing the actual bad point coordinates.

[0190] Filtering the candidate defect location set based on defect probability aims to identify the locations that are actually defective. The preset criteria are pre-defined standards used to determine whether a candidate defect location is a genuine defect. By eliminating candidate locations with defect probabilities lower than the preset criteria, false positives are reduced, improving the accuracy of defect location. The final defect location result set contains the coordinates of the identified genuine defective locations, which can be used for subsequent repair or further analysis.

[0191] In one implementation, step S500 may specifically include the following steps S510-S560:

[0192] Step S510: Collect all candidate locations and their corresponding bad point probabilities in the candidate bad point location set, and construct a candidate location-probability lookup table, where each entry contains a horizontal coordinate index, a vertical coordinate index, and the corresponding bad point probability.

[0193] Collecting all candidate locations and their corresponding bad pixel probabilities from the candidate bad pixel location set is for organizing and summarizing this information to facilitate subsequent analysis and filtering. A candidate location-probability lookup table is constructed, combining the horizontal and vertical coordinate indices of each candidate location with its corresponding bad pixel probability into an entry and storing it in the table. This lookup table clearly shows the coordinates of each candidate location and its corresponding bad pixel probability, facilitating comparison and filtering. For example, a suitable data structure (such as a list or dictionary) can be used to construct the candidate location-probability lookup table. The candidate bad pixel location set is traversed, and for each candidate location, its horizontal and vertical coordinate indices and corresponding bad pixel probability are extracted, and this information is combined into an entry and added to the lookup table. For example, a dictionary can be used to store each entry, with the coordinate information as the key and the bad pixel probability as the value.

[0194] Step S520: Perform statistical analysis on the probability of all bad points in the candidate position-probability comparison table, calculate the median and interquartile range of the probability distribution, and determine the initial screening threshold based on the median and interquartile range.

[0195] Statistical analysis of all bad pixel probabilities in the candidate position-probability comparison table is performed to understand the distribution of bad pixel probabilities. The median, located in the middle after arranging all bad pixel probabilities in ascending order, reflects the central tendency of the data. The interquartile range (ICM), the difference between the upper and lower quartiles, measures the dispersion of the data. Determining the initial screening threshold based on the median and ICM is a screening method based on data distribution characteristics, which can more reasonably determine the screening criteria and reduce misjudgments.

[0196] Step S530: Perform preliminary screening of candidate positions based on the initial screening threshold, retain candidate positions with a bad point probability greater than or equal to the initial screening threshold, generate a preliminary screening set, and record the coordinates and bad point probabilities of candidate positions in the preliminary screening set.

[0197] The initial screening of candidate locations based on the initial screening threshold aims to eliminate locations with low probability of defective pixels from the candidate defective pixel location set, reducing the workload of subsequent processing. Candidate locations with a defective pixel probability greater than or equal to the initial screening threshold are retained; these locations are more likely to be actual defective pixels. A preliminary screening set is generated, and the coordinates and defective pixel probabilities of these retained candidate locations are recorded to provide a basis for further screening.

[0198] For example, each entry in the candidate location-probability lookup table is iterated through, and the probability of a bad pixel is compared to an initial screening threshold. If the probability of a bad pixel is greater than or equal to the initial screening threshold, the coordinates of the candidate location corresponding to that entry and the probability of a bad pixel are added to the preliminary screening set. The preliminary screening set can be stored using a list or dictionary, where each element contains the coordinates of the candidate location and the probability of a bad pixel.

[0199] Step S540: Perform spatial clustering analysis on the candidate locations in the preliminary screening set. Use density clustering algorithm, set the neighborhood range and minimum number of contained points, and cluster candidate locations with spatial distance less than the neighborhood range into one class to obtain multiple bad point clusters.

[0200] Spatial clustering analysis of candidate locations in the initial screening set aims to discover their spatial clustering characteristics. Density clustering is a clustering method based on data point density, which divides data points into different clusters according to the distance and density relationships between them. Setting the neighborhood range and the minimum number of contained points are two important parameters of density clustering. The neighborhood range defines the spatial extent surrounding a data point, while the minimum number of contained points specifies the minimum number of data points a cluster must contain. Clustering candidate locations with spatial distances smaller than the neighborhood range into one class yields multiple bad point clusters. These clusters may represent different bad point regions, which helps in further analyzing the distribution patterns of bad points.

[0201] For example, a density-based clustering algorithm, such as DBSCAN (Density-Based Spatial Clustering), is chosen. The spatial distance between candidate locations in the initial selection set is calculated. A neighborhood range and a minimum number of contained points are set, and the DBSCAN algorithm is used to cluster the candidate locations. The algorithm clusters candidate locations whose spatial distance is less than the neighborhood range and meet the minimum number of contained points into one class, forming different bad point clusters. The DBSCAN algorithm can be implemented using a clustering library in a programming language (such as scikit-learn).

[0202] Step S550: For each bad point cluster, calculate the average bad point probability of all candidate positions within the cluster, use the average as the cluster probability of the cluster, then retain candidate positions whose bad point probability is greater than the cluster probability and remove candidate positions whose bad point probability is less than the cluster probability, thus optimizing the candidate positions of each cluster.

[0203] For each cluster of defective pixels, the average probability of defective pixels at all candidate locations within the cluster is calculated and used as the cluster probability. This measure assesses the overall likelihood of defective pixels within the cluster. Retaining candidate locations with a defective pixel probability greater than the cluster probability and discarding those with a probability less than the cluster probability is an optimization method based on intra-cluster probability comparison. Since candidate locations with a defective pixel probability greater than the cluster probability are more likely to be actual defective pixels, this filtering method can further improve the accuracy of defective pixel location and optimize the candidate locations for each cluster.

[0204] For example, each cluster of defective locations is iterated through, and the sum of the defective location probabilities for all candidate locations within the cluster is calculated. This sum is then divided by the number of candidate locations to obtain the cluster probability. Next, each candidate location within a cluster is iterated through, and its defective location probability is compared to the cluster probability. If the defective location probability is greater than the cluster probability, the candidate location is retained; otherwise, it is removed from the cluster. A list or dictionary can be used to store the candidate locations for each cluster, and filtering and optimization are implemented through looping and comparison operations.

[0205] Step S560: Collect the candidate location coordinates in all optimized clusters, remove duplicate coordinates generated by cluster optimization, arrange them in ascending order of horizontal coordinate index and ascending order of vertical coordinate index, and generate a set of bad spot location results containing the coordinates of real bad spots. Each coordinate in the set corresponds to a real bad spot that has been screened and confirmed.

[0206] For example, all optimized clusters are traversed, and the candidate location coordinates in each cluster are extracted and added to a temporary set. Duplicate coordinates are removed using the properties of the set data structure. Then, the coordinates in the set are sorted in ascending order of horizontal coordinate index and then in ascending order of vertical coordinate index. The resulting ordered set is the set of bad spot location results containing the coordinates of the actual bad spots.

[0207] Figure 2 This is a schematic diagram of the composition structure of a detection and positioning device provided in an embodiment of the present invention, as shown below. Figure 2 As shown, the detection and positioning device 200 includes:

[0208] The data acquisition module 210 is used to apply a preset driving signal sequence to the pixel array of the Miniled display line by line, acquire the brightness feedback data of each pixel at different driving stages, and obtain a brightness response sequence containing pixel position information and corresponding brightness feedback data.

[0209] The spatial recombination module 220 is used to spatially recombine the brightness response sequence according to the physical arrangement structure of the pixel array to generate a brightness response matrix. The row dimension and column dimension of the brightness response matrix correspond to the horizontal pixel index and vertical pixel index of the display screen, respectively.

[0210] The difference analysis module 230 is used to perform local region feature difference analysis on the brightness response matrix, identify the brightness response abnormal regions that meet the preset abnormal conditions, and generate a candidate bad pixel location set. The candidate bad pixel location set contains the coordinates of multiple pixels with abnormal brightness response features.

[0211] The feature learning module 240 is used to call the pre-trained bad pixel recognition network to perform feature learning processing on the candidate bad pixel location set, and output the bad pixel probability corresponding to each candidate bad pixel location. The bad pixel probability is used to represent the probability that the candidate bad pixel location is a real bad pixel.

[0212] The defective pixel location module 250 is used to filter the set of candidate defective pixel locations based on the defective pixel probability, remove candidate locations with defective pixel probabilities lower than preset conditions, and generate a set of defective pixel location results containing the actual defective pixel coordinates.

[0213] The descriptions of the apparatus embodiments above are similar to those of the method embodiments above, and have similar beneficial effects. In some embodiments, the functions or modules included in the apparatus provided by the present invention can be used to perform the methods described in the method embodiments above. For technical details not disclosed in the apparatus embodiments of the present invention, please refer to the descriptions of the method embodiments of the present invention for understanding.

[0214] Figure 3 A hardware entity diagram of a computer system provided as an embodiment of the present invention, such as... Figure 3 As shown, the hardware entity of the computer system 1000 includes a processor 1001 and a memory 1002, wherein the memory 1002 stores a computer program that can run on the processor 1001, and the processor 1001 executes the program to implement the steps in the method of any of the above embodiments.

Claims

1. A method for detecting and locating defective pixels on a minimized display screen based on deep learning, characterized in that, The method includes: A preset driving signal sequence is applied row by row to the pixel array of the Miniled display, and the brightness feedback data of each pixel at different driving stages is obtained to obtain a brightness response sequence containing pixel position information and corresponding brightness feedback data. The brightness response sequence is spatially reorganized according to the physical arrangement structure of the pixel array to generate a brightness response matrix. The row dimension and column dimension of the brightness response matrix correspond to the horizontal pixel index and vertical pixel index of the display screen, respectively. Local region feature difference analysis is performed on the brightness response matrix to identify brightness response abnormal regions that meet preset abnormal conditions, and a candidate bad pixel location set is generated. The candidate bad pixel location set contains multiple pixel coordinates with abnormal brightness response characteristics. The pre-trained defective pixel recognition network is invoked to perform feature learning processing on the candidate defective pixel location set, and the defective pixel probability corresponding to each candidate defective pixel location is output. The defective pixel probability is used to represent the probability that the candidate defective pixel location is a real defective pixel. The candidate bad spot location set is filtered according to the bad spot probability, and candidate locations with bad spot probabilities lower than preset conditions are eliminated to generate a bad spot location result set containing the actual bad spot coordinates.

2. The method according to claim 1, characterized in that, The step of spatially reorganizing the brightness response sequence according to the physical arrangement structure of the pixel array to generate a brightness response matrix, wherein the row dimension and column dimension of the brightness response matrix correspond to the horizontal pixel index and vertical pixel index of the display screen, respectively, includes: The pixel position information in the brightness response sequence is parsed, the physical driving line path identifier and array coordinate code of each pixel are extracted, and the driving line path identifier is converted into the initial horizontal index and the initial vertical index through the preset line path-coordinate mapping rule to generate the original pixel association table containing the original coordinates and brightness feedback data. Obtain the physical topology map of the pixel array of the Miniled display, identify the row driving line distribution and column driving line distribution, determine the actual arrangement order of pixels in each row and the actual arrangement order of pixels in each column, and establish a physical arrangement calibration model. The initial horizontal and initial vertical indices in the original pixel association table are input into the physical layout calibration model. Coordinate mapping correction is performed based on the row-driven line distribution and column-driven line distribution to obtain the calibrated horizontal and vertical coordinate indices. The original pixel association table is then updated to the pixel calibration association table. Based on the nominal resolution parameters of the Miniled display and the number of effective pixels in the physical topology diagram, determine the target row dimension parameters and target column dimension parameters of the brightness response matrix, where the target row dimension parameter is equal to the number of effective horizontal pixels and the target column dimension parameter is equal to the number of effective vertical pixels. Create a blank brightness response matrix with the number of rows and columns being the target row dimension parameter and the number of columns being the target column dimension parameter. Set the initial values ​​of the matrix elements to preset invalid values. Iterate through each pixel entry in the pixel calibration association table, locate the filling position in the blank matrix according to the calibrated horizontal and vertical coordinate indices, and write the corresponding brightness feedback data into the filling position. After filling all pixel entries, neighborhood interpolation is performed on the positions in the blank luminance response matrix that are still invalid values. The filling value is calculated based on the luminance feedback data of the adjacent valid pixels to obtain the complete luminance response matrix.

3. The method according to claim 2, characterized in that, The process of acquiring the pixel array physical topology map of the Miniled display, identifying the row and column drive line distributions, determining the actual arrangement order of pixels in each row and column, and establishing a physical arrangement calibration model includes: Call the display screen physical parameter database, read the drive circuit design drawing data of the target Miniled display screen, and extract the circuit topology data containing row drive circuit number, column drive circuit number and corresponding pixel connection relationship; A graph structure model is performed on the line topology data. The line-pixel association graph is constructed with row-driven line numbers and column-driven line numbers as nodes and pixel connection relationships as edges. Each edge attribute contains the connection order number of the pixel in the corresponding line. Traverse all row-driven line nodes in the line-pixel association graph, extract the pixel connection sequence number connected to each row node, sort the numbers in ascending order to obtain the actual pixel arrangement order corresponding to the row-driven line, and generate a row arrangement order table. Traverse all column-driven line nodes in the same way, extract the pixel connection sequence number connected to each column node, sort the numbers in ascending order to obtain the actual pixel arrangement order corresponding to the column-driven line, and generate a column arrangement order table. Based on the row and column order lists, establish a mapping function between the initial coordinate index and the actual order. The horizontal mapping function takes the initial horizontal index as input and outputs the calibrated horizontal coordinate index, while the vertical mapping function takes the initial vertical index as input and outputs the calibrated vertical coordinate index. The horizontal and vertical mapping functions are integrated into a physical layout calibration model. The model inputs are the initial horizontal and initial vertical indices, and the outputs are the calibrated horizontal and vertical coordinate indices. The model parameters include the mapping relationship data of the row arrangement order table and the column arrangement order table.

4. The method according to claim 3, characterized in that, The graph structure modeling of the line topology data, using row-driven and column-driven line numbers as nodes and pixel connection relationships as edges, constructs a line-pixel association graph, including: Parse the row-driven line records in the line topology data, extract the unique identifier number of each row-driven line as the row node ID, and extract the unique identifier number of each column-driven line as the column node ID, and store them in the row node set respectively; The initial line-pixel association graph is a directed graph structure. The node set of the directed graph structure is a joint node set consisting of the row node set and the column node set. The edge set of the directed graph structure is initially an empty set. Traverse the pixel connection records in the line topology data. Each record contains the row-driven line number, column-driven line number, pixel identifier, and the connection order number of the pixel in the corresponding row line and the connection order number in the corresponding column line. Starting from the row node corresponding to the row-driven line number and ending from the column node corresponding to the column-driven line number, a directed edge is created. The pixel identifier, row connection sequence number, and column connection sequence number are used as the attribute data of the edge and added to the edge set of the line-pixel association graph. After creating the edges for all pixel connection records, perform an integrity check on the line-pixel association graph. Check whether each row node and column node has at least one associated edge. If there is an isolated node, mark it as an abnormal line node and record the abnormal line node information.

5. The method according to claim 4, characterized in that, After creating the edges for all pixel connection records, the integrity of the line-pixel association graph is checked. Each row and column node is examined to ensure it has at least one associated edge. If an isolated node exists, it is marked as an abnormal line node, and its information is recorded, including: Traverse the set of row nodes in the line-pixel relationship graph. For each row node, count the number of edges that serve as its starting point, and record it as the row node edge count. If the row node edge count is zero, mark the row as an isolated row node. Traverse the set of column nodes in the same way. For each column node, count the number of edges that are its endpoints and record it as the column node edge count. If the column node edge count is zero, mark the column as an isolated column node. Collect all isolated row nodes and isolated column nodes, record their node IDs and corresponding line numbers, and generate a list of abnormal line nodes. Calculate the proportion of isolated row nodes to the total number of row nodes and the proportion of isolated column nodes to the total number of column nodes. If either proportion exceeds a preset threshold, it is determined that the line topology data has an integrity failure. Based on the list of abnormal line nodes and the calculation results of the proportion, the integrity level of the line topology data is determined. The integrity level is used to guide whether it is necessary to reacquire the line topology data. If the system is determined to be in a state of integrity failure, the process of re-acquiring the line topology data will be triggered; otherwise, subsequent processing will continue and abnormal node information will be recorded in the system log.

6. The method according to claim 1, characterized in that, The step involves performing local region feature difference analysis on the brightness response matrix to identify brightness response anomaly regions that meet preset anomaly conditions, generating a candidate bad pixel location set, including: The brightness response matrix is ​​divided into multi-granularity regions, and three different analysis windows with different analysis granularities are set to capture local detail features, regional distribution features and global correlation features, respectively. Each analysis window slides in the matrix at a preset step size to generate multiple sets of local region sequences. For each local region in each set of local region sequences, extract the brightness time series features of the pixels within the region, including the rising edge change rate, falling edge change rate, duration of the stable phase, and fluctuation amplitude of the brightness response curve, and construct the regional time series feature vector. Calculate the spatial neighborhood association features of each local region. Taking the center pixel of the region as the benchmark, calculate the absolute value of the brightness difference between it and its surrounding neighboring pixels, and take the average value to obtain the spatial association index. The regional temporal feature vector and spatial correlation index are input into the preset abnormal region judgment model, and the abnormality probability score of the local region is output. The abnormality probability score comprehensively considers the degree of deviation of the temporal feature from the normal range and the degree of abnormality of the spatial correlation index. Anomaly probability scores are calculated for local region sequences divided by three analysis granularity windows. Then, the anomaly probability scores of the same pixel coordinate under different analysis granularities are weighted and fused. The weights are preset according to the adaptability of window analysis granularity and pixel density. Based on the weighted fusion anomaly probability score, an anomaly score threshold is set. Pixel coordinates with scores higher than the threshold are marked as candidate bad pixels. All marked pixel coordinates are collected, and duplicate coordinates are removed to generate a set of candidate bad pixel locations.

7. The method according to claim 6, characterized in that, The brightness response matrix is ​​divided into multi-granularity regions, and three different analysis windows with different granularities are set to capture local detail features, regional distribution features, and global correlation features, respectively. Each analysis window slides in the matrix at a preset step size to generate multiple sets of local region sequences, including: Based on the target row dimension parameters and target column dimension parameters of the brightness response matrix, the size specifications of the analysis window for three analysis granularities are preset to ensure that the window size is an odd number to maintain central symmetry. Calculate the sliding step size for each analysis window, and set the step size value to ensure that there is an overlapping area between adjacent windows; Starting from the beginning of the luminance response matrix, the local detail window is slid, moving horizontally by a step size of pixels each time. When the horizontal sliding reaches the matrix boundary, it moves vertically by a step size of pixels and resets the horizontal position to the starting column, until the entire matrix is ​​covered. The starting and ending coordinates of each window are recorded to generate a sequence of local detail regions. Using the same starting position and sliding method, the regional distribution window and the global association window are slid-processed respectively to generate the regional distribution region sequence and the global association region sequence. The coordinate range of each local region is determined by the window's starting coordinates, the window's side length, and the matrix boundary. Boundary verification is performed on the region coordinates in each local region sequence. If the window slides to the edge of the matrix and causes some regions to exceed the matrix range, the pixel brightness value of the exceeding part is set to the average of the existing pixel brightness values ​​in the window, so that the size of each local region conforms to the preset window side length. The local detail region sequence, the regional distribution region sequence, and the global associated region sequence are stored separately. Each sequence contains the coordinate range of multiple local regions and the corresponding window size information to obtain the multi-granularity region division result. For each local region in each set of local region sequences, the brightness time series features of pixels within the region are extracted, including the rising edge rate of change, falling edge rate of change, duration of the stable phase, and fluctuation amplitude of the brightness response curve, to construct a regional time series feature vector, including: Traverse each local region in the local region sequence, extract the luminance feedback data of all pixels in the local region from the luminance response matrix according to its coordinate range, and obtain the regional luminance data set, where each pixel corresponds to a set of luminance response data that changes with the driving stage; For each pixel in the region luminance data set, its luminance response data is arranged in the order of the driving stages to obtain the luminance response curve of that pixel. The order of the driving stages is determined according to the application order of the preset driving signal sequence. Identify the rising phase, the stable phase, and the falling phase of the brightness response curve. The rising phase is when the brightness rises from the initial value to the peak value, the stable phase is when the brightness remains near the peak value, and the falling phase is when the brightness falls from the peak value to the end value. In the rising phase, a first starting point and a first ending point are selected. The first starting point is the moment when the brightness reaches the first proportion of the peak value, and the first ending point is the moment when the brightness reaches the second proportion of the peak value. The first difference between the brightness of the first ending point and the brightness of the first starting point is divided by the second difference between the moment of the first ending point and the moment of the first starting point. The result of the division is determined as the rising edge change rate. In the falling phase, a second starting point and a second ending point are selected. The second starting point is the moment when the brightness drops from the peak to the third proportion, and the second ending point is the moment when it drops to the fourth proportion. The third difference between the brightness of the second ending point and the brightness of the second starting point is divided by the fourth difference between the second ending point and the second starting point. The result of the division is determined as the falling edge change rate. The time interval from when the brightness reaches the fifth peak value to when it decreases to the end of that value is determined as the duration of the stable phase; the difference between the maximum and minimum brightness values ​​within the stable phase is determined as the fluctuation amplitude; the rising edge rate, falling edge rate, duration of the stable phase, and fluctuation amplitude are used as the temporal features of the pixel, and the average of the temporal features of all pixels in the region is taken to obtain the regional temporal feature vector.

8. The method according to claim 1, characterized in that, The pre-trained defect detection network is invoked to perform feature learning processing on the candidate defect location set, outputting the defect probability corresponding to each candidate defect location, including: For each candidate location in the candidate bad pixel location set, a three-dimensional neighborhood data block of a preset size is extracted from the brightness response matrix with its coordinates as the center. The three dimensions are the horizontal pixel index, the vertical pixel index, and the driving stage, respectively. The three-dimensional neighborhood data block contains the brightness feedback data of the candidate location and its surrounding pixels in different driving stages. Feature enhancement processing is performed on the three-dimensional neighborhood data block, and differential operation is performed along the driving stage dimension to obtain the brightness change rate data block. Edge detection is performed along the horizontal and vertical dimensions to obtain the horizontal edge data block and the vertical edge data block. The original data block, brightness change rate data block, horizontal edge data block and vertical edge data block are concatenated into multi-channel input data according to the channel dimension. Multi-channel input data is fed into the feature extraction module of the pre-trained bad pixel recognition network, which contains multiple levels of convolutional layers. Each convolutional layer is followed by a batch normalization layer and an activation function. The preceding convolutional layer is used to extract local detail features, and the subsequent convolutional layer is used to extract global context features. The high-dimensional feature map output by the feature extraction module is input into the attention fusion module of the network. The attention fusion module includes a spatial attention branch and a channel attention branch. The spatial attention branch generates a spatial weight map through global average pooling and convolution operations, and the channel attention branch generates a channel weight map through global max pooling and convolution operations. The two weight maps are multiplied with the high-dimensional feature map to obtain the attention-enhanced feature map. The attention-enhanced feature map is input into the temporal dependency modeling layer of the network. This temporal dependency modeling layer contains multiple bidirectional memory units, which perform temporal dependency modeling on the feature map along the driving stage dimension, capture the dynamic features of brightness response changing with the driving stage, and output the temporal enhancement feature vector. The temporal enhancement feature vector is input into the classification output layer of the network. This classification output layer outputs a single probability value through a weighted combination and nonlinear transformation of multiple neurons, which is the probability of the bad point corresponding to the candidate bad point location.

9. A detection and positioning device, characterized in that, include: The data acquisition module is used to apply a preset driving signal sequence to the pixel array of the Miniled display line by line, acquire the brightness feedback data of each pixel at different driving stages, and obtain a brightness response sequence containing pixel position information and corresponding brightness feedback data. The spatial recombination module is used to spatially recombine the brightness response sequence according to the physical arrangement structure of the pixel array to generate a brightness response matrix. The row dimension and column dimension of the brightness response matrix correspond to the horizontal pixel index and vertical pixel index of the display screen, respectively. The difference analysis module is used to perform local region feature difference analysis on the brightness response matrix, identify the brightness response abnormal regions that meet the preset abnormal conditions, and generate a candidate bad pixel location set, which contains multiple pixel coordinates with abnormal brightness response characteristics. The feature learning module is used to call a pre-trained bad pixel recognition network to perform feature learning processing on the candidate bad pixel location set, and output the bad pixel probability corresponding to each candidate bad pixel location. The bad pixel probability is used to represent the probability that the candidate bad pixel location is a real bad pixel. The defective pixel location module is used to filter the set of candidate defective pixel locations based on the defective pixel probability, remove candidate locations with defective pixel probabilities lower than preset conditions, and generate a set of defective pixel location results containing the actual defective pixel coordinates.

10. A computer system comprising a memory and a processor, the memory storing a computer program executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method according to any one of claims 1 to 8.

Images