CN-121565070-B - Miniled display screen dead pixel detection and positioning method device and system based on deep learning

Abstract

The invention provides a method, a device and a system for detecting and positioning bad points of Miniled display screens based on deep learning, which are characterized in that a preset driving signal sequence is applied to a pixel array of Miniled display screens row by row, brightness feedback data of each pixel in different driving stages are obtained to obtain a brightness response sequence, the brightness response sequence is subjected to space recombination processing according to a physical arrangement structure of the pixel array to generate a brightness response matrix, local area characteristic difference analysis is performed on the brightness response matrix, a brightness response abnormal area is identified, a candidate bad point position set is generated, a pre-trained bad point identification network is called to perform characteristic learning processing on the candidate bad point position set, the bad point probability corresponding to each candidate bad point is output, candidate positions with the bad point probability lower than a preset condition are removed based on the candidate bad point probability, and a bad point positioning result set is generated. The invention can effectively improve the accuracy and reliability of detecting and positioning the bad points of Miniled display screens.

Inventors

• WANG JIAN
• CHEN YANHUA
• CHEN CHAOCHENG
• Cai Xiaoen
• Yang Guanshui
• YUAN PING

Assignees

• 贵州理工学院
• 深圳市创义信光电科技有限公司

Dates

Publication Date

20260508

Application Date

20251128

Claims (10)

1. 1. A method for detecting and positioning a bad point of a Miniled display screen based on deep learning is characterized by comprising the following steps: applying a preset driving signal sequence to a pixel array of the Miniled display screen line by line, and acquiring brightness feedback data of each pixel in different driving stages to obtain a brightness response sequence containing pixel position information and corresponding brightness feedback data; Performing spatial recombination processing on the brightness response sequence according to a physical arrangement structure of the pixel array to generate a brightness response matrix, wherein the row dimension and the column dimension of the brightness response matrix respectively correspond to a horizontal pixel index and a longitudinal pixel index of the display screen; Carrying out local area characteristic difference analysis on the brightness response matrix, identifying a brightness response abnormal area meeting a preset abnormal condition, and generating a candidate dead point position set which comprises a plurality of pixel coordinates with abnormal brightness response characteristics; Invoking a pre-trained dead point identification network to perform feature learning processing on the candidate dead point set, and outputting dead point probability corresponding to each candidate dead point, wherein the dead point probability is used for representing the probability that the candidate dead point position is a true dead point; And screening the candidate dead point set according to the dead point probability, removing candidate positions with the dead point probability lower than a preset condition, and generating a dead point positioning result set containing real dead point coordinates.
2. 2. The method of claim 1, wherein the spatially reorganizing the luminance response sequence according to a physical arrangement of the pixel array to generate a luminance response matrix, where a row dimension and a column dimension of the luminance response matrix correspond to a horizontal pixel index and a vertical pixel index of the display screen, respectively, includes: analyzing pixel position information in the brightness response sequence, extracting a physical driving line path identifier and an array coordinate code of each pixel, converting the driving line path identifier into an initial transverse index and an initial longitudinal index through a preset line path-coordinate mapping rule, and generating a pixel original association table containing original coordinates and brightness feedback data; acquiring Miniled a pixel array physical topological graph of the display screen, identifying row driving line distribution and column driving line distribution in the pixel array physical topological graph, determining the actual arrangement sequence of each row of pixels and the actual arrangement sequence of each column of pixels, and establishing a physical arrangement calibration model; Inputting the initial transverse index and the initial longitudinal index in the pixel original association table into a physical arrangement calibration model, carrying out coordinate mapping correction according to the line driving line distribution and the column driving line distribution to obtain a calibrated transverse coordinate index and a calibrated longitudinal coordinate index, and updating the pixel original association table into a pixel calibration association table; Determining a target row dimension parameter and a target column dimension parameter of a brightness response matrix according to a nominal resolution parameter of Miniled display screens and the number of effective pixels in a physical topological graph, wherein the target row dimension parameter is equal to the number of effective transverse pixels, and the target column dimension parameter is equal to the number of effective longitudinal pixels; Creating a blank brightness response matrix, wherein the row number of the matrix is a target row dimension parameter, the column number is a target column dimension parameter, the initial value of a matrix element is set to be a preset invalid value, traversing each pixel entry in a pixel calibration association table, positioning a filling position in the blank matrix according to the calibrated transverse coordinate index and longitudinal coordinate index, and writing corresponding brightness feedback data into the filling position; After filling all pixel items, carrying out neighborhood interpolation processing on the position which is still an invalid value in the blank brightness response matrix, and calculating filling values based on brightness feedback data of adjacent effective pixels to obtain a complete brightness response matrix.
3. 3. The method of claim 2, wherein the obtaining Miniled a physical topology of the pixel array of the display screen, identifying a row driving line distribution and a column driving line distribution therein, determining an actual arrangement order of each row of pixels and an actual arrangement order of each column of pixels, and establishing the physical arrangement calibration model includes: calling a display screen physical parameter database, reading driving line design drawing data of a target Miniled display screen, and extracting line topology data comprising a line driving line number, a column driving line number and a corresponding pixel connection relation; carrying out graph structure modeling on the line topology data, taking a line driving line number and a column driving line number as nodes, and taking a pixel connection relationship as an edge, and constructing a line-pixel association graph, wherein each edge attribute comprises a connection sequence number of pixels in a corresponding line; Traversing all line driving line nodes in the line-pixel association diagram, extracting the serial numbers of pixel connection sequences connected with each line node, and arranging the serial numbers in ascending order to obtain the actual arrangement sequence of pixels corresponding to the line driving line, so as to generate a line arrangement sequence table; Traversing all column driving circuit nodes in the same way, extracting the serial numbers of the pixel connection sequences connected with each column node, and arranging the serial numbers in ascending order to obtain the actual arrangement sequence of the pixels corresponding to the column driving circuit, so as to generate a column arrangement sequence table; Establishing a mapping function of an initial coordinate index and an actual arrangement sequence according to the row arrangement sequence table and the column arrangement sequence table, wherein the transverse mapping function inputs the initial transverse index and outputs a calibrated transverse coordinate index, and the longitudinal mapping function inputs the initial longitudinal index and outputs a calibrated longitudinal coordinate index; And integrating the transverse mapping function and the longitudinal mapping function into a physical arrangement calibration model, inputting the model into an initial transverse index and an initial longitudinal index, and outputting the model into a calibrated transverse coordinate index and a calibrated longitudinal coordinate index, wherein model parameters comprise mapping relation data of a row arrangement sequence table and a column arrangement sequence table.
4. 4. A method according to claim 3, wherein the modeling the line topology data by using the line driving line number and the column driving line number as nodes and the pixel connection relationship as edges to construct a line-pixel association graph comprises: analyzing line driving line records in line topology data, extracting unique identification numbers of driving lines of each line as line node IDs, extracting unique identification numbers of driving lines of each column as column node IDs, and respectively storing the unique identification numbers in a line node set; Initializing a line-pixel association graph to form a directed graph structure, wherein a node set of the directed graph structure is a joint node set consisting of a row node set and a column node set, and an edge set of the directed graph structure is initially an empty set; traversing pixel connection records in the line topology data, wherein each record comprises a line driving line number, a column driving line number, a pixel identifier, a connection sequence number of the pixel in a corresponding line and a connection sequence number of the pixel in a corresponding column line; Creating a directed edge by taking a row node corresponding to a row driving line number as a starting point and a column node corresponding to a column driving line number as an end point, and adding pixel identifications, row connection sequence numbers and column connection sequence numbers as attribute data of the edge to an edge set of a line-pixel association graph; After the creation of the edges of all the pixel connection records is completed, carrying out integrity check on the line-pixel association graph, checking whether at least one associated edge exists in each row node and each column node, marking the line node as an abnormal line node if an isolated node exists, and recording the information of the abnormal line node.
5. 5. The method of claim 4, wherein after the creating of the edges of all pixel connection records is completed, performing an integrity check on the line-pixel association graph, checking whether at least one associated edge exists in each row node and each column node, if so, marking the line node as an abnormal line node, and recording abnormal line node information, wherein the method comprises the steps of: Traversing a line node set of the line-pixel association graph, counting the number of edges taking each line node as a starting point for each line node, marking the line node as the line node edge number, and marking the line node with isolated line if the line node edge number is zero; Traversing the column node set in the same way, counting the number of edges taking each column node as a terminal point for each column node, marking the column node as an isolated column node if the number of the edges of the column node is zero; Collecting all isolated row nodes and isolated column nodes, recording node IDs and corresponding line numbers of the isolated row nodes and the isolated column nodes, and generating an abnormal line node list; calculating the proportion of the isolated row nodes to the total row nodes and the proportion of the isolated column nodes to the total column nodes, and judging that the line topology data has an integrity failure condition if any proportion exceeds a preset threshold value; determining the integrity level of the line topology data according to the abnormal line node list and the proportion calculation result, wherein the integrity level is used for guiding whether the line topology data needs to be acquired again or not; If the integrity failure condition is judged, triggering a circuit topology data reacquiring process, otherwise, continuing the subsequent processing and recording abnormal node information in a system log.
6. 6. The method of claim 1, wherein the performing the local region feature difference analysis on the luminance response matrix identifies a luminance response anomaly region in which a preset anomaly condition is satisfied, and generating the candidate dead point set comprises: Dividing the brightness response matrix into multiple granularity areas, setting three analysis windows with different analysis granularities, respectively capturing local detail features, area distribution features and global association features, sliding each analysis window in the matrix according to a preset step length, and generating a plurality of groups of local area sequences; For each local region in each group of local region sequences, extracting the brightness time sequence characteristics of pixels in the region, including the rising edge change rate, the falling edge change rate, the duration of a stable phase and the fluctuation amplitude of a brightness response curve, and constructing a region time sequence feature vector; calculating the spatial neighborhood correlation characteristic of each local area, respectively calculating the absolute value of the brightness difference value between each local area and surrounding adjacent pixels by taking the central pixel of the area as a reference, and obtaining a mean value to obtain a spatial correlation index; Inputting the regional time sequence feature vector and the spatial association degree index into a preset abnormal regional judgment model, and outputting an abnormal probability score of the local region, wherein the abnormal probability score comprehensively considers the degree of deviation of the time sequence feature from a normal range and the abnormal degree of the spatial association degree index; Respectively carrying out anomaly possibility scoring calculation on the local area sequences divided by the three analysis granularity windows, then carrying out weighted fusion on anomaly possibility scores of the same pixel coordinate under different analysis granularities, and presetting weights according to suitability of window analysis granularities and pixel densities; And setting an anomaly scoring threshold according to the anomaly possibility scores after weighted fusion, marking the pixel coordinates with scores higher than the threshold as candidate dead points, collecting the pixel coordinates of all marks, and generating a candidate dead point set after removing the repeated coordinates.
7. 7. The method of claim 6, wherein the performing multi-granularity region division on the luminance response matrix sets three analysis windows with different analysis granularities, each of which is used for capturing local detail features, region distribution features and global association features, and each of which slides in the matrix according to a preset step length, and generating a plurality of groups of local region sequences, including: Presetting the size specification of analysis windows with three analysis granularities according to the target row dimension parameter and the target column dimension parameter of the brightness response matrix, so that the window size is odd so as to maintain central symmetry; Calculating the sliding step length of each analysis window, wherein the step length value is set to ensure that the adjacent windows have overlapping areas; sliding the local detail windows from the initial position of the brightness response matrix, moving the step-length pixels along the transverse direction each time, moving the step-length pixels longitudinally and resetting the transverse position as an initial column when the step-length pixels transversely slide to the matrix boundary until the whole matrix is covered, recording the initial coordinate and the end coordinate of each window, and generating a local detail region sequence; respectively carrying out sliding treatment on the region distribution window and the global association window in the same starting position and sliding mode to generate a region distribution region sequence and a global association region sequence, wherein the coordinate range of each local region is jointly determined by window starting coordinates, window side lengths and matrix boundaries; performing boundary verification on the region coordinates in each local region sequence, and if the window slides to the edge of the matrix to cause the partial region to exceed the range of the matrix, setting the brightness value of the pixel exceeding the range as the average value of the brightness values of the existing pixels in the window so as to ensure that the size of each local region accords with the preset window side length; storing a local detail region sequence, a region distribution region sequence and a global association region sequence respectively, wherein each sequence comprises coordinate ranges of a plurality of local regions and corresponding window size information, and obtaining a multi-granularity region division result; Extracting, for each local region in each group of local region sequences, a luminance time sequence feature of pixels in the region, including a rising edge change rate, a falling edge change rate, a duration of a stable phase, and a fluctuation amplitude of a luminance response curve, and constructing a region time sequence feature vector, including: Traversing each local area in the local area sequence, extracting brightness feedback data of all pixels in the local area from a brightness response matrix according to the coordinate range of the local area to obtain an area brightness data set, wherein each pixel corresponds to a group of brightness response data changing along with a driving stage; For each pixel in the regional brightness data set, arranging brightness response data of the pixel according to a driving stage sequence to obtain a brightness response curve of the pixel, wherein the driving stage sequence is determined according to the application sequence of a preset driving signal sequence; identifying a rising stage, a stabilizing stage and a falling stage of the brightness response curve, wherein the rising stage is a stage of rising the brightness from an initial value to a peak value, the stabilizing stage is a stage of maintaining the brightness near the peak value, and the falling stage is a stage of falling the brightness from the peak value to an end value; selecting a first starting point and a first end point in the rising stage, wherein the first starting point is the moment when the brightness reaches a first proportion of a peak value, the first end point is the moment when the brightness reaches a second proportion of the peak value, dividing a first difference value between the brightness of the first end point and the brightness of the first starting point by a second difference value between the moment of the first end point and the moment of the first starting point, and determining a division result as a rising edge change rate; selecting a second starting point and a second ending point in the falling stage, wherein the second starting point is the moment when the brightness falls from the peak value to the third proportion, the second ending point is the moment when the brightness falls to the fourth proportion, dividing the third difference value between the brightness of the second ending point and the brightness of the second starting point by the fourth difference value between the moment of the second ending point and the moment of the second starting point, and determining the division result as the falling edge change rate; The method comprises the steps of determining a time interval from the fifth proportion of brightness reaching a peak value to the end of the proportion as a stable phase duration, determining a difference value between a maximum value and a minimum value of brightness in a stable phase as a fluctuation amplitude, taking a rising edge change rate, a falling edge change rate, the stable phase duration and the fluctuation amplitude as time sequence characteristics of the pixel, and taking an average value of the time sequence characteristics of all pixels in the region to obtain a region time sequence characteristic vector.
8. 8. The method of claim 1, wherein invoking the pre-trained dead point recognition network to perform feature learning processing on the candidate dead point set, and outputting a dead point probability corresponding to each candidate dead point comprises: Intercepting a three-dimensional neighborhood data block with a preset size from a brightness response matrix for each candidate position in a candidate dead point position set by taking the coordinates of the candidate position as a center, wherein three-dimensional dimensions are a transverse pixel index, a longitudinal pixel index and a driving stage respectively, and the three-dimensional neighborhood data block comprises brightness feedback data of candidate positions and peripheral pixels thereof in different driving stages; Performing characteristic enhancement processing on the three-dimensional neighborhood data blocks, performing differential operation along the dimension of the driving stage to obtain brightness change rate data blocks, performing edge detection along the transverse dimension and the longitudinal dimension respectively to obtain transverse edge data blocks and longitudinal edge data blocks, and splicing the original data blocks, the brightness change rate data blocks, the transverse edge data blocks and the longitudinal edge data blocks into multi-channel input data according to the channel dimension; Inputting the multichannel input data into a feature extraction module of a pre-trained dead pixel identification network, wherein the feature extraction module comprises a multi-stage convolution layer, a batch normalization layer and an activation function are connected behind each layer of convolution layer, the former stage convolution layer is used for extracting local detail features, and the latter stage convolution layer is used for extracting global context features; inputting the high-dimensional feature map output by the feature extraction module into a attention fusion module of the network, wherein the attention fusion module comprises a spatial attention branch and a channel attention branch, the spatial attention branch generates a spatial weight map through global average pooling and convolution operation, the channel attention branch generates a channel weight map through global maximum pooling and convolution operation, and the two weight maps are multiplied with the high-dimensional feature map to obtain an attention enhancement feature map; Inputting the attention-enhancing feature map into a time-sequence dependent modeling layer of a network, wherein the time-sequence dependent modeling layer comprises a plurality of two-way memory units, performing time-sequence dependent modeling on the feature map along the dimension of a driving stage, capturing dynamic features of brightness response changing along with the driving stage, and outputting time-sequence enhancing feature vectors; and inputting the time sequence enhanced feature vector into a classification output layer of the network, and outputting a single probability value, namely the dead pixel probability corresponding to the candidate dead pixel by the classification output layer through the weighted combination and the nonlinear transformation of a plurality of neurons.
9. 9. A test positioning device, comprising: The data acquisition module is used for applying a preset driving signal sequence to the pixel array of the Miniled display screen row by row to acquire brightness feedback data of each pixel in different driving stages, and acquiring a brightness response sequence containing pixel position information and corresponding brightness feedback data; The space reorganization module is used for carrying out space reorganization processing on 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 the column dimension of the brightness response matrix respectively correspond to the horizontal pixel index and the longitudinal pixel index of the display screen; The difference analysis module is used for carrying out local area characteristic difference analysis on the brightness response matrix, identifying a brightness response abnormal area meeting a preset abnormal condition, and generating a candidate dead point position set which contains a plurality of pixel coordinates with abnormal brightness response characteristics; The feature learning module is used for calling a pre-trained dead point identification network to perform feature learning processing on the candidate dead point set and outputting dead point probability corresponding to each candidate dead point, wherein the dead point probability is used for representing the probability that the candidate dead point position is a true dead point; And the dead pixel positioning module is used for screening the candidate dead pixel set according to the dead pixel probability, eliminating candidate positions with the dead pixel probability lower than a preset condition, and generating a dead pixel positioning result set containing real dead pixel coordinates.
10. 10. A computer system comprising a memory and a processor, the memory storing a computer program executable on the processor, wherein the processor implements the steps of the method of any one of claims 1 to 8 when the program is executed.

Description

Miniled display screen dead pixel detection and positioning method device and system based on deep learning Technical Field The invention relates to the technical field of deep learning, in particular to a Miniled display screen dead point detection and positioning method device and system based on deep learning. Background Along with the continuous development of display technology, miniled display screens are widely applied to various display devices by virtue of the advantages of high brightness, high contrast, fine display effect and the like, and dead point detection is used as a key quality control link in the Miniled display screen production process, so that the display effect and user experience of products are directly affected. At present, a common method for detecting bad pixels of a Miniled display screen generally applies a driving signal with fixed parameters to a pixel array to obtain single brightness feedback data of each pixel, and then judges whether the pixel is a bad pixel or not by setting a fixed brightness threshold value, or stores brightness information by adopting a simple linear data sequence and performs single-point comparison analysis. The traditional methods are difficult to capture the dynamic response characteristics of the dead pixels under different driving conditions, so that partial dead pixels which are abnormal in performance only in a specific driving stage are missed to be detected, real dead pixels and false abnormal points are difficult to effectively distinguish, and the accuracy and reliability of detection results are insufficient. Disclosure of Invention In view of the above, the invention provides a method, a device and a system for detecting and positioning a defective pixel of a Miniled display screen based on deep learning. The technical scheme of the embodiment of the invention is realized as follows: On one hand, the embodiment of the invention provides a method for detecting and positioning a bad pixel of a Miniled display screen based on deep learning, which comprises the steps of applying a preset driving signal sequence to a pixel array of the Miniled display screen row by row, obtaining brightness feedback data of each pixel in different driving stages to obtain a brightness response sequence containing pixel position information and corresponding brightness feedback data, carrying out space recombination processing on the brightness response sequence according to a physical arrangement structure of the pixel array to generate a brightness response matrix, wherein row dimension and column dimension of the brightness response matrix respectively correspond to a horizontal pixel index and a longitudinal pixel index of the display screen, carrying out local area characteristic difference analysis on the brightness response matrix, identifying a brightness response abnormal area meeting preset abnormal conditions, generating a candidate bad pixel position set, wherein the candidate bad pixel position set comprises a plurality of pixel coordinates with abnormal brightness response characteristics, calling a pre-trained bad pixel identification network to carry out feature learning processing on the candidate bad pixel position set, outputting bad pixel probability corresponding to each candidate bad pixel position, wherein the bad pixel probability is used for representing the probability that the candidate bad pixel position is a true bad pixel, carrying out screening on the candidate bad position set according to the bad pixel probability, and removing the bad pixel position probability is lower than the preset conditions, and generating a candidate bad position set containing true bad position result. In another aspect, an embodiment of the present invention provides a detection positioning device, including: The data acquisition module is used for applying a preset driving signal sequence to the pixel array of the Miniled display screen row by row to acquire brightness feedback data of each pixel in different driving stages, and acquiring a brightness response sequence containing pixel position information and corresponding brightness feedback data; The space reorganization module is used for carrying out space reorganization processing on the brightness response sequence according to the physical arrangement structure of the pixel array to generate a brightness response matrix, and the row dimension and the column dimension of the brightness response matrix respectively correspond to the horizontal pixel index and the longitudinal pixel index of the display screen; The difference analysis module is used for carrying out local area characteristic difference analysis on the brightness response matrix, identifying a brightness response abnormal area meeting preset abnormal conditions, and generating a candidate dead point position set which contains a plurality of pixel coordinates with abnormal brightness response characteristics; The feature learnin