Visual MicroLED color deviation identification method for improving yield

Abstract

The invention relates to the technical field of image recognition, in particular to a visual MicroLED color deviation recognition method for improving the yield, which comprises the following steps: collecting color data of a MicroLED display screen, continuously reading RGB color values of all points on the display screen, establishing a color distribution database, and calculating chromatic aberration by comparing the color values of all the points with average color distribution. According to the method, the RGB color values of all the pixel points on the MicroLED display screen are collected, the color information is continuously read, and therefore the accurate color distribution database is constructed, the average color distribution serves as the standard, abnormal data points with the color deviation exceeding the set threshold value are recognized and screened out, and the remarkable color difference area is locked. Meanwhile, the significant color deviation points are further subjected to spatial aggregation trend analysis, a high-risk area with the most serious color deviation is rapidly recognized, and the positioning accuracy of the color deviation problem is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of image recognition, and particularly to a visual MicroLED color deviation recognition method for improving the yield rate. Background Art

[0002] Image recognition technology is an important branch in the field of computer vision. It mainly refers to the technology of detecting, classifying, recognizing, or analyzing target objects or features in images by a computer.

[0003] In the prior art, only rough target extraction or simple feature comparison of image information is performed, ignoring the spatial correlation between color data and failing to effectively quantify the spatial distribution of abnormal points. This simple detection method easily leads to unclear positioning of color deviation areas, restricting the rapid positioning and repair efficiency of defect problems. Therefore, improvements are needed. Summary of the Invention

[0004] The purpose of the present invention is to solve the deficiencies in the prior art and propose a visual MicroLED color deviation recognition method for improving the yield rate.

[0005] To achieve the above purpose, the present invention adopts the following technical solutions. A visual MicroLED color deviation recognition method for improving the yield rate includes the following steps: Collect color data of the MicroLED display screen, continuously read the RGB color values of each point on the display screen, establish a color distribution database, calculate the color difference by comparing the color values of each point with the average color distribution, and obtain preliminary color difference data; for the preliminary color difference data, screen the data points with deviations exceeding the set threshold to generate a list of significant color difference points; Perform spatial analysis on the list of significant color difference points. By calculating the distance between each color difference point and its nearest neighbor point, analyze the aggregation trend of the color difference points to obtain the color difference aggregation analysis result; based on the color difference aggregation analysis result, identify the estimated color difference area to obtain a high-risk color difference area; Collect the color deviation images of past MicroLED display screens as input, perform deep learning training to generate a color difference detection model; Apply the color difference detection model to perform color difference detection on the high-risk color difference area to generate a color difference recognition result.

[0006] Preferably, the step of obtaining the preliminary color difference data is as follows: Collect the RGB color values of each pixel point on the MicroLED display screen, record the color values to form a color value data set; Based on the color value data set, calculate the average value of the RGB color values to generate a color difference data set; Based on the color difference data set, calculate the color difference influence factor of each point, and summarize the color difference influence factors to obtain preliminary color difference data. The calculation formula is as follows: ; Wherein, represents the color difference influence factor of the i-th pixel point, represents the value of the i-th pixel in the j-th color of the RGB color spectrum, represents the average value of the average color distribution in the j-th color, represents the standard deviation of the j-th color.

[0007] Preferably, the steps for obtaining the list of significant color difference points are as follows: Classify and sort the color difference influence factors of each pixel point in the preliminary color difference data, extract the color difference interval values between adjacent pixels according to the arrangement order, and generate a color difference interval sequence; Based on the color difference interval sequence, calculate the dynamic threshold. The calculation formula is as follows: ; Wherein, is the k-th color difference interval value, is the maximum value in the color difference interval sequence, is the minimum value in the color difference interval sequence, is the median of the color difference interval sequence, is the total number of color difference interval values, is the red component value of the k-th pixel point, is the green component value of the k-th pixel point, is the dynamic threshold.

[0008] According to the dynamic threshold, screen the pixel points in the preliminary color difference data whose color difference influence factors are greater than the dynamic threshold to form a list of significant color difference points.

[0009] Preferably, the steps for obtaining the color difference aggregation analysis result are as follows: Extract the two-dimensional coordinates of each point in the list of significant color difference points, calculate the Euclidean distance of each point to the nearest neighbor point, and obtain a distance data set; Based on the distance data set, calculate the spatial aggregation degree of the color difference points. The calculation formula is as follows: ; Wherein, represents the distance between the i-th color difference point and the nearest neighbor point, is the number of points in the distance data set, is the spatial aggregation degree.

[0010] According to the spatial aggregation degree, determine the spatial aggregation trend of the color difference points and generate a color difference aggregation analysis result.

[0011] Preferably, the steps for obtaining the high-risk color difference region are as follows: Extract the boundary coordinate ranges of each aggregation block from the color difference aggregation analysis results, and count the number of color difference points within each boundary range to generate an aggregation region space coverage set; Based on the aggregation region space coverage set, calculate the risk score value for each region. The calculation formula is: ; Wherein, and are the coordinates of the upper left corner and the lower right corner of the color difference aggregation region respectively, is the number of color difference points within this region, and are the coordinates of the two farthest color difference points within this region, and are the maximum and minimum values of the brightness of the color difference points within this region, is the risk score value of this region; According to the risk score value, filter out the spatial regions with risk score values higher than the set threshold to form high-risk color difference regions.

[0012] Preferably, the steps for obtaining the color difference detection model are as follows: Collect the color deviation images of past MicroLED displays, and extract the RGB channel color difference variances, the number of abnormal pixels, the edge change amplitudes, the gray channel discrete values, and the regional saturation extreme differences of each frame of image to form an image feature structure set; Based on the image feature structure set, calculate the image training effectiveness score value. The calculation formula is: ; Wherein, and and are the color difference variance values of the red, green, and blue channels respectively, is the gray channel discrete value, is the number of abnormal pixel points, is the maximum saturation value within the region, is the minimum saturation value within the region, is the number of continuous edge segments in the image, is the image training effectiveness score value; According to the image training effectiveness score value, filter out the image frames for deep learning training to obtain a color difference detection model.

[0013] Preferably, the steps for obtaining the color difference recognition result are as follows: Deploy the color difference detection model to the production line; Obtain the image data of the high-risk color difference area, and input the image data of the high-risk color difference area into the color difference detection model; The color difference detection model processes the input image data of the high-risk color difference area, analyzes and identifies the type and degree of color difference, and generates a color difference identification result.

[0014] Compared with the prior art, the advantages and positive effects of the present invention are as follows: The present invention collects the RGB color values of each pixel point on the MicroLED display screen, continuously reads these color information, thereby constructing an accurate color distribution database. Based on the average color distribution, it identifies and filters out abnormal data points with color deviations exceeding the set threshold, and locks the significant color difference area. At the same time, by further analyzing the spatial aggregation trend of these significant color difference points, it quickly identifies the high-risk areas with the most serious color deviations, improving the positioning accuracy of color deviation problems. In addition, by using the color deviation image as input data, deep learning training is carried out to generate a color difference detection model, realizing the identification and rapid response to abnormal situations in the high-risk color difference area, improving the product yield and quality control efficiency, reducing the manual detection cost, and enhancing the overall production efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a step schematic diagram of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0016] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention.

[0017] Please refer to Figure 1 , the present invention provides a technical solution, a visual MicroLED color deviation identification method for improving the yield rate, including the following steps: Collect the color data of the MicroLED display screen, continuously read the RGB color values of each point on the display screen, establish a color distribution database, calculate the color difference by comparing the color values of each point with the average color distribution, and obtain preliminary color difference data; for the preliminary color difference data, filter out the data points with deviations exceeding the set threshold, and generate a list of significant color difference points; Conduct a spatial analysis on the list of significant color difference points. By calculating the distance between each color difference point and its nearest neighbor point, analyze the aggregation trend of the color difference points to obtain the color difference aggregation analysis result; according to the color difference aggregation analysis result, identify the estimated color difference area to obtain the high-risk color difference area; Collect the color deviation images of past MicroLED display screens as input, conduct deep learning training, and generate a color difference detection model; Apply the color difference detection model to detect the color difference in high-risk color difference areas and generate color difference recognition results.

[0018] The steps for obtaining preliminary color difference data are as follows: Collect the RGB color values of each pixel on the MicroLED display screen, record the color values, and form a color value dataset; Based on the color value dataset, calculate the average value of the RGB color values to generate a color difference dataset; Based on the color difference dataset, calculate the color difference influence factor for each point, and summarize the color difference influence factors to obtain preliminary color difference data. The calculation formula is: ; where, represents the color difference influence factor of the i-th pixel point, represents the value of the i-th pixel in the j-th color of the RGB color spectrum, represents the average value of the average color distribution in the j-th color, represents the standard deviation of the j-th color.

[0019] Specifically, based on the pixel distribution of the MicroLED display screen, first perform a point-by-point scan on the entire display screen area in the actual production environment. Use a detection device integrated with red, green, and blue light intensity receiving components and calibrate it to ensure that the RGB color values of each pixel remain stable when read. Subsequently, sequentially collect the original RGB values of the pixel points in the order of horizontal coordinates from 0 to 4095 and vertical coordinates from 0 to 2175. When scanning, use a step interval of 1 to traverse all pixel points. If the RGB signal reception intensity of some pixel points exceeds the pre-determined high brightness threshold of 500 cd / m², then compare the color information of this pixel point with the existing brightness distribution reference table. By recording the difference between the RGB values of this pixel point and the corresponding interval in the brightness distribution reference table, to confirm whether the acquisition and reading are accurate. To further accurately distinguish the pixel point information in different brightness intervals, use a light intensity meter to calibrate the offset of the detection device during sampling to ensure that the RGB values read during repeated scans do not drift. Then integrate the RGB values of all pixel points into a centrally managed color value record list for sequentially summarizing and storing the pixel numbers and three-channel values, and record the corresponding coordinate positions during sampling in the table. For example, the first 100×100 pixels in the upper left corner coordinate area of the display screen are all recorded in the first row of the list, and then the RGB values of the subsequent pixels are recorded in this coordinate order. After the acquisition is completed, merge all the record rows to obtain a color value dataset containing all pixel information of the entire display screen.

[0020] Based on the color value dataset obtained previously, it is necessary to sum up the RGB values of all pixel points and perform addition operations separately. When reading each record of this dataset, traverse it sequentially from left to right and then from top to bottom according to the pixel coordinates. For the red channel value R, the green channel value G, and the blue channel value B, accumulate and sum them respectively. After completing the sequential accumulation of all pixels, the red channel sum ΣR, the green channel sum ΣG, and the blue channel sum ΣB are obtained. Then, count the total number of pixels N. Divide ΣR by N to get the red channel average μR, divide ΣG by N to get the green channel average μG, and divide ΣB by N to get the blue channel average μB. Then, write these three groups of averages into a new data sequence in the record order as the average RGB color value list. When detecting that the data of some pixel points have abnormal values, they will be excluded during the reading stage. For example, when the RGB values in a certain record are all 0 or all exceed 255, it is regarded as an invalid record. After uniform illumination is completed, the available valid record sequence forms the final average RGB data sequence. Subsequently, perform element-level difference calculations between each point in this average RGB color value list and the original color value data, and organize the deviations between all pixels and the average value to form a color difference dataset.

[0021] Formula: , the benefit of the formula is that it comprehensively considers the deviation degree between pixel points and the average color distribution, and introduces the standard deviation of each channel at the denominator position. After cumulative summation of the three-channel ratio and then performing the cube root operation, the dispersion of the RGB three channels is uniformly measured, avoiding excessive deviation of a single channel alone and taking into account the comprehensive differences of multiple channels, and being able to more balancedly reflect the pixel deviation situation when monitoring the MicroLED color difference, thereby providing a solid numerical basis for subsequent judgment of whether there is abnormal color difference.

[0022] The steps to obtain the parameter are: This parameter represents the actual acquisition value of the th pixel in the th channel, usually including three groups of values, corresponding to the red, green, and blue channels respectively. For example, during a certain measurement, the peak signal of the red channel of the 220th pixel is converted to get , and the green and blue channels are converted to get and respectively.

[0023] The steps to obtain the parameter are: This is a parameter used to identify the average color distribution of the th channel. It is necessary to first sum up the channel values of all pixels detected in the same batch, and then divide by the number of pixels to obtain the channel mean. When obtaining it, select a complete display screen for full-pixel scanning, accumulate and sum up the RGB values of each pixel to get the red channel total , the sum of the green channels , the sum of the blue channels , and record the total number of pixels , then let , , , A total of 5 million pixels were collected on the display screen. The total sum of the red channels reached 3.81×10^8, the total sum of the green channels was 3.74×10^8, and the total sum of the blue channels was 3.79×10^8. Then , , .

[0024] The steps for obtaining the parameter are as follows: This parameter is the standard deviation of the th channel. To obtain the value, it is necessary to first calculate the square of the deviation between each pixel channel value and the mean based on , and then perform a square root operation after summing them up. The specific process is as follows: For the red channel, first square and sum the difference between the red channel data of each pixel and to obtain , then divide by the total number of pixels and then take the square root to get . The same method can be used for the green and blue channels. This value measures the degree of dispersion of the entire channel. For example, among the 5 million pixels summarized previously, the statistical value of the sum of the squared deviations of the red channel is approximately 2.85×10^9. After dividing by the total number of pixels and then taking the square root, we get . The results of the green and blue channels can also be obtained in the same way.

[0025] Calculation process: Let the red channel of the 210th pixel be , the green channel , the blue channel . Combining the mean and standard deviation examples mentioned above, select , , , and , , . First, calculate the absolute deviation ratios respectively: ; Add these three ratios together: ; Then perform a cube root operation: ; This result indicates that the comprehensive deviation between the 210th pixel and the overall average color distribution is approximately 1.63.

[0026] The steps for obtaining the list of significant color difference points are as follows: Classify and sort the color difference influence factors of each pixel point in the preliminary color difference data, extract the color difference interval values between adjacent pixels according to the arrangement order, and generate a color difference interval sequence; Based on the color difference interval sequence, calculate the dynamic threshold, and the calculation formula is: ; Wherein, is the k-th color difference interval value, is the maximum value in the color difference interval sequence, is the minimum value in the color difference interval sequence, is the median of the color difference interval sequence, is the total number of color difference interval values, is the red component value of the k-th pixel point, is the green component value of the k-th pixel point, is the dynamic threshold; According to the dynamic threshold, screen the pixel points in the preliminary color difference data whose color difference influence factors are greater than the dynamic threshold to form a list of significant color difference points.

[0027] Specifically, based on the preliminary color difference data obtained previously, it is necessary to classify and sort the color difference influence factors of all pixels in ascending order. After reading, retrieve the adjacent pixel pairs in sequence according to the sorting sequence and extract the difference between the two as the color difference interval value. By using the index correspondence method, compare the sorting positions of each pixel with the sorting position of the next pixel to determine the actual difference between the two and mark it in a list of color difference interval values. Subsequently, check whether each interval value is within the previously set valid range, for example, compare the value range between 0 and 20. If there is an interval value exceeding this range, classify it as an outlier interval and record the corresponding pixel pair coordinates during subsequent processing. To ensure the accuracy of the interval value extraction process, it is necessary to statistically analyze the data distribution of the color difference influence factors in advance to determine whether there are extremely large values. If extreme values appear, they should be specifically marked at this stage according to the specific situation of the value and temporarily placed in the abnormal inspection area, and the sorting position of the extreme value should be retained in the list of color difference interval values so that it can be associated with the overall sequence during subsequent summarization. Combining the generated list of color difference interval values can intuitively display the degree of color difference between adjacent pixels, thereby completing the extraction and recording process of all adjacent pixel intervals.

[0028] Formula: The advantage of the formula is that it incorporates both the overall dispersion of the color difference interval sequence and the differences between the red and green component values of each pixel into the calculation structure. Through the superimposed method of segmented multiplication and segmented summation, it can comprehensively evaluate different pixel interval sizes and different pixel component differences, and obtain a numerical dynamic threshold after calculation. This threshold can be used to quickly identify those pixels that exceed the normal interval range in the subsequent process, thereby assisting in accurately selecting pixels with significant color difference problems in large-scale display screen detection.

[0029] The steps to obtain the parameter are as follows: This parameter represents the th color difference interval value, which needs to be obtained by extracting the color difference values between adjacent pixels after sorting all pixels according to the color difference influence factor. In order to obtain , the absolute difference will be calculated between the pixel at the th sorting position and the pixel at the th sorting position, forming an interval value and storing it in the interval value set. If there are a total of pixels that have been sorted, then interval values will be obtained. After sequentially extracting the color difference influence factors on a certain MicroLED display screen and sorting them from small to large, 5000 valid records are obtained and the differences between adjacent records are calculated in sequence, thus generating 4999 values. If the absolute value of the difference between the 200th and 201st in the sorting sequence is 3.5, then this value is stored as , and the other interval values are obtained in the same way.

[0030] The steps to obtain the parameter are as follows: This parameter is the maximum value in the color difference interval sequence and needs to find the element with the largest value from the list formed in the previous step. When obtaining it, first read all and compare them one by one to determine the largest one. For example, among the 4999 interval values obtained from the above 5000 valid records, after comparison, the largest interval value is approximately 8.2, that is .

[0031] The steps to obtain the parameter are as follows: This parameter is the minimum value in the color difference interval sequence. For example, the minimum value among the aforementioned 4999 interval values is approximately 0.1, then it is determined as .

[0032] The steps to obtain the parameter are as follows: This parameter represents the median of the color difference interval sequence and requires sorting all interval values from small to large and then selecting the value at the middle position. During the obtaining process, it will first ensure that the number of the interval value list is in a known state. If If it is odd, directly select the th value. If is even, then average the th value and the th value after sorting to obtain the median. For specific calculation examples, refer to the above scenario of 4999 interval values. is odd. Take out the 2500th value after sorting and obtain . For example, after calculation .

[0033] The steps for obtaining the parameter are as follows: This parameter refers to the total number of color difference interval values, usually equal to the number of sorted pixels minus 1. When obtaining it, it needs to be determined based on the previously obtained number of pixels . Generally , if there are some abnormal pixels excluded in the previous detection, the total number of available pixels finally generated will be slightly less than the initial number of pixels, thus affecting value. Specifically, the number of effective pixels obtained after each classification and sorting can be recorded, and the difference can be directly calculated to obtain . In the above example scenario, 4999 interval values can be obtained from 5000 valid records. Therefore .

[0034] The steps for obtaining the parameter are as follows: This parameter is the red component value of the th pixel point. At a certain detection site, the signal of the red channel of the 1700th pixel is collected multiple times, and after denoising and conversion, is obtained.

[0035] The steps for obtaining the parameter are as follows: This parameter represents the green component value of the th pixel point. Similar to , it is obtained by collecting the green channel brightness signal. If the signal intensity of the green channel of the 1700th pixel is recorded as about 183 during the detection, then through the same denoising and compensation methods, can be finally obtained.

[0036] Calculation process: Suppose 4999 interval values are obtained after collection, that is , where , , . And record the red component and green component of a certain segment of pixels. The average absolute value of the difference is about 4. Specifically, for each item in these 4999 interval values, calculate the average value, and after summarization, , then calculate for each pixel point , compare these values with the sum of , take the average after comparison, and this average result is about 2.3 for example. Finally, add these two parts together and perform and operations: ; Then take the fourth root of this result: ; After that, add another part to the above result: ; Finally, take the square root of the whole sum: ; This result shows that the dynamic threshold obtained at this time is about 2.08. When the color difference influence factor is greater than 2.08, the pixel points will be judged as those with a relatively high degree of color difference, which can be included in the subsequent screening process of significant color difference points, and the color difference distribution of different display screens can be further compared in large-scale detection scenarios.

[0037] Based on the dynamic threshold obtained in the previous step, it is necessary to select the color difference influence factors of all pixel points from the preliminary color difference data obtained before, and compare the influence factor values corresponding to each pixel with the dynamic threshold one by one. Retrieve all pixel records with influence factor values exceeding 2.08, and then establish a coordinate list of significant color difference points. During the execution process, first read the preliminary color difference data table that has been uniformly stored before, check the color difference influence factors of each pixel point one by one in a sequential manner, and compare the size relationship between these values and the threshold in real time. When it is detected that the value of a pixel point is greater than 2.08, extract the row and column coordinates of the pixel point and store them in the screening result. To avoid data indexing errors, it is necessary to match the record order of the preliminary color difference data table with the order of pixel numbers on the display screen before retrieval, and then traverse in the order from top to bottom and then from left to right. Whenever the threshold is exceeded, append a significant color difference point record to the result list. At the same time, the coordinate positions recorded will also be highlighted, which can be located when comparing the visual coordinate map of the pixel distribution on the display screen. Then merge all the result entries to obtain the final significant color difference point list.

[0038] The steps to obtain the color difference aggregation analysis result are as follows: Extract the two-dimensional coordinates of each point in the significant color difference point list, calculate the Euclidean distance of each point to its nearest neighbor point, and obtain a distance data set; Based on the distance data set, calculate the spatial aggregation degree of the color difference points, and the calculation formula is: ; in, Represents the distance between the ith color difference point and its nearest neighbor, is the number of points in the distance dataset, is the spatial aggregation degree.

[0039] According to the spatial aggregation degree, the spatial aggregation trend of the color difference points is determined and the color difference aggregation analysis results are generated.

[0040] Specifically, based on the previously obtained list of significant color difference points, it is necessary to read the pixel coordinate information stored in the list one by one, and accurately check the coordinates in combination with the physical resolution and pixel layout range of the display screen. The coordinates of each pixel point are compared in pairs with the pixel coordinates of its left neighbor, right neighbor, upper neighbor, lower neighbor and diagonal neighbor. By reading the markings of these adjacent pixels in the list of significant color difference points, it is determined whether there is a comparable reference object. Subsequently, the row and column coordinate differences between the significant color difference point and the nearest neighbor pixel are calculated and squared and summed. The square root of the result is then taken to obtain the Euclidean distance. For a screen with a resolution of 3200×1800, the horizontal coordinate range can be set to 0 to 3199 and the vertical coordinate range can be set to 0 to 1799. The coordinates of each significant color difference point are searched within this range to ensure that there is no invalid coordinate data beyond the boundary. When performing the calculation, a coordinate correspondence table is first established, and the row and column indexes and the front, back, left and right positions of each pixel are recorded for each pixel. The coordinate value on the right is used to locate the nearest neighbor pixel when reading the significant color difference point. For any significant color difference point, such as row index 520 and column index 1085, first find out whether there is a pixel coordinate in the eight positions of row index 520 and column index 1084, row index 520 and column index 1086, row index 519 and column index 1085, row index 521 and column index 1085, row index 519 and column index 1084, row index 519 and column index 1086, row index 521 and column index 1084, and row index 521 and column index 1086. If the coordinate is completely within the range, the distance is obtained by calculating the square sum of the row and column differences, and then the smallest one is selected from these calculated distances as the nearest neighbor distance of the significant color difference point. If the coordinates in some directions exceed the valid range, the calculation of the direction is abandoned and the boundary situation is noted when recording. After the distance calculation of all significant color difference points is completed, these minimum distance value sets are summarized and stored correspondingly according to the serial number, and finally a distance data set is formed.

[0041] formula: , The advantage of the formula is that by combining the operations of geometric mean distance, summing the reciprocals of distances, and the arctangent function, it can quantitatively analyze the spatial closeness between significant color difference points from multiple perspectives. It not only measures the relative distribution of points in geometric space but also takes into account the impact of extremely large and extremely small distances on the overall result, so as to use this value to evaluate the aggregation trend of color difference points in subsequent steps.

[0042] The steps to obtain the parameter are as follows: This parameter represents the distance between the th color difference point and its nearest neighbor pixel. To obtain , it is necessary to square and sum the differences between the two coordinates and then take the square root on the basis of knowing the row and column coordinates of the th color difference point and the row and column coordinates of the adjacent pixel. The mathematical form is: , where represents the coordinates of the th color difference point, represents the coordinates of its nearest neighbor pixel. For example, when the coordinates of the 620th color difference point are (610, 1350), the nearest adjacent pixel coordinates after comparison are (609, 1350), then the row and column differences are 1 and 0 respectively, and calculate , and finally .

[0043] The steps to obtain the parameter are as follows: This parameter represents the number of points in the distance dataset, corresponding to the number of significant color difference points. If a certain number of significant color difference points are determined in the previous steps, then the corresponding distance values can be generated equivalently. For example, when 230 coordinates are recorded in the significant color difference point list and all the corresponding nearest neighbor pixels are found during the detection process, can be determined.

[0044] Calculation process: Now select color difference points as an example. Among them, , , , , . First, calculate : ; Take out each value of ln respectively: , , , , , the sum is 0 + 0.7885 + 0.5878 + 1.1314 + 0.9933 = 3.5009, then divided by 5 to get 0.70018, which is used as the power of the exponential function: ; Then calculate : ; The value is approximately , and then look at : ; ; The sum is approximately 5.4692, then divided by 5 to get 1.09384, and after squaring it is , and the previous parts are incorporated into the denominator: ; Divide the numerator 2.014 by 4.2972: ; This result indicates that when is approximately 0.4688, it means that the spatial distribution of these 5 color difference points within the screen is not extremely dense. By comparing with the test results of other regions, it can be judged that the overall aggregation degree is at a relatively low level at this time. If is closer to 1 or greater than 1, it often means that these color difference points are relatively closely distributed to each other, otherwise a lower value indicates that the intervals between them are relatively dispersed.

[0045] According to the spatial aggregation degree calculation results obtained previously, it is necessary to determine the distribution of color difference points in the current sampling batch or the entire display area. By comparing the calculated aggregation degree value with several pre-established aggregation degree boundary values, for example, the pre-set aggregation degree boundary values are 0.6 and 1.0, and the spatial aggregation degree lower than 0.6 is regarded as a loose distribution, between 0.6 and 1.0 as a medium aggregation, and greater than 1.0 as a tight aggregation. During the execution process, first compare the calculated aggregation degree with 0.6. If the aggregation degree is less than 0.6, mark the current batch record as a loose type, otherwise continue to compare with 1.0. If it is between 0.6 and 1.0, mark it as a medium aggregation, otherwise mark it as a tight aggregation. Subsequently, add the aggregation type to the record entry and centrally store the coordinates and aggregation type of the color difference points. In a scenario with a resolution of 3200×1800, hundreds or even thousands of color difference points may appear. Through the above comparison process, the corresponding aggregation type distribution can be obtained, and then the results are summarized as the color difference aggregation analysis result.

[0046] The steps to obtain the high-risk color difference area are as follows: Extract the boundary coordinate ranges of each aggregation block from the chromatic aberration aggregation analysis results, and count the number of chromatic aberration points within each boundary range to generate a spatial coverage set of the aggregation regions; Based on the spatial coverage set of the aggregation regions, calculate the risk score value for each region. The calculation formula is: ; where, and are the coordinates of the upper left corner and the lower right corner of the chromatic aberration aggregation region respectively, is the number of chromatic aberration points within this region, and are the coordinates of the two farthest chromatic aberration points in this region, and are the maximum and minimum values of the brightness of the chromatic aberration points in this region, is the risk score value of this region; According to the risk score values, filter out the spatial regions with risk score values higher than the set threshold to form high-risk chromatic aberration regions.

[0047] Specifically, based on the summary list of chromatic aberration aggregation analysis results, it is necessary to read one by one the coordinate information of the chromatic aberration points marked as the same aggregation block, obtain the upper left corner boundary of the aggregation block area by confirming the smallest abscissa value and ordinate value among these coordinates, and obtain the lower right corner boundary of the aggregation block area by confirming the largest abscissa value and ordinate value among these coordinates. Subsequently, all pixel points in the aggregation block are included in the statistical scope one by one, and their row and column indexes and corresponding RGB acquisition values are recorded. Then, the number of all chromatic aberration points in the aggregation block is accumulated. If there are duplicate or missing situations found at certain positions, the indexes of the coordinate list are re-compared to correct the data. The obtained boundary coordinate range and the corresponding number of chromatic aberration points are registered in the aggregation block statistical table according to the aggregation block number respectively. Then, continue to process the next aggregation block. After completing the boundary determination and quantity statistics for all aggregation blocks, these information are centrally stored as an aggregation block index list, and the upper left and lower right coordinate values of each aggregation block and the quantity distribution data of all pixel points in this area are retained in the list. For aggregation blocks with a relatively high pixel density, the record of the coordinate difference can be increased during statistics to assist in further area estimation subsequently. To avoid the situation of coordinate out-of-bounds, it is necessary to strictly limit the screen resolution range before generating this aggregation block index list. For example, the effective range is set to an abscissa from 0 to 3199 and an ordinate from 0 to 1799. Whenever the coordinate of an aggregation block exceeds this range, the coordinate value is excluded or corrected. After a set of boundary coordinates corresponding to each aggregation block is determined correctly, it is included in the same row of the aggregation block index list, and the number of chromatic aberration points is filled in synchronously. During the whole process, the previously determined chromatic aberration point record table will be continuously checked to identify which chromatic aberration points belong to the same aggregation block and combine them to calculate the boundary range and quantity. After all aggregation blocks are statistically completed, the corresponding aggregation area space coverage set can be obtained.

[0048] Formula: , The benefit of the formula is that it incorporates the spatial scale of the aggregation block, the number of chromatic aberration points, and the brightness range into a comprehensive evaluation structure at the same time. Through multiple combinations of taking the fourth root, taking the cube root, and logarithmic functions, the influences of the aggregation block size, the distribution of the number of chromatic aberration points, and the pixel brightness difference can be reflected in a single score value.

[0049] The steps to obtain the parameter are as follows: This parameter represents the abscissa of the upper left corner of the chromatic aberration aggregation area and is used to measure the value of the starting position of the aggregation block in the horizontal direction of the screen. Previously, according to the aggregation block statistical table, the set of pixel point values with the smallest abscissa in each aggregation block has been obtained, and this is used as the source. For example, for a 3200×1800 screen, the abscissa can be limited to 0 to 3199. Through statistics, it is known that the smallest abscissa in this aggregation block is 200, then .

[0050] The steps for obtaining the parameter are as follows: This parameter represents the ordinate of the upper left corner of the color difference aggregation area, and is the same as the minimum ordinate value of the color difference points in the aggregation block. If it is determined that the ordinates of all color difference points in an aggregation block are concentrated in the range of 120 to 150, then 120 is recognized as the minimum ordinate of this aggregation block. Therefore, .

[0051] The steps for obtaining the parameter are as follows: This parameter is used to identify the abscissa of the lower right corner of the color difference aggregation area, and is obtained by finding the maximum abscissa value of all color difference points in the same aggregation block, which is different from taking the minimum value, but taking the maximum value. If the maximum value of the abscissas of the color difference points in this aggregation block is found to be 250, then .

[0052] The steps for obtaining the parameter are as follows: This parameter is the ordinate of the lower right corner of the color difference aggregation area, and is similar to the above, and is obtained by extracting the maximum ordinate of the color difference points in the aggregation block. For example, if the ordinates of the color difference points in a certain area are distributed between 120 and 180, then the maximum value is 180. Furthermore, .

[0053] The steps for obtaining the parameter are as follows: This parameter represents the number of color difference points in the target aggregation area, and can be directly read from the total number of entries in the aggregation block statistical table generated previously for assignment. If 50 significant color difference points are counted in an aggregation block, then .

[0054] The steps for obtaining the parameter are as follows: This is one of the coordinates of the two farthest color difference points in this area, and is used to participate in the calculation of the maximum span of the area. After counting the coordinates of all pixel points inside the aggregation block, pairwise distance operations are performed and the two coordinate points corresponding to the global maximum distance are determined. The coordinate of one of the pairs is defined as , and the coordinate of the other point is defined as . When obtaining, a coordinate set is made for all color difference points in the aggregation block according to the row and column numbers, and then the distance values are compared one by one. After finding the pair with the maximum distance, their respective coordinates are respectively assigned to and in order to accurately reflect the maximum span of this area when calculating the risk score later. If the pair with the maximum distance in this area falls on (210, 130) and (290, 178), then .

[0055] The steps for obtaining the parameters are as follows: Corresponding to , this is the coordinate of another point of the two farthest color difference points in the same aggregation block. The obtaining method is the same. First, find the coordinate pair with the maximum distance through the comparison of the full-region coordinates, mark one of the points as , and mark the other point as . For example, the other point (290, 178) in the above example is .

[0056] The steps for obtaining the parameters are as follows: These two parameters respectively represent the maximum and minimum values of the brightness of the color difference points in this area, and need to be statistically analyzed in combination with the RGB brightness measurement results during acquisition. A brightness value will be recorded for each pixel during detection . If a certain pixel is a significant color difference point, its brightness value will be included in the statistical range of the aggregation block. By comparing the brightness data under this aggregation block one by one, the maximum value can be obtained, which is , and the minimum value is . For example, if it is measured that the maximum brightness in an aggregation block reaches 510 cd / ㎡ and the minimum brightness reaches 100 cd / ㎡, then .

[0057] Calculation process: Select an aggregation block, among which , the two color difference points with the farthest distance are respectively at (210, 130) and (290, 178). Therefore, . In addition, the number of color difference points in this area is . The maximum brightness and the minimum brightness are obtained through on-site brightness measurement. Substitute the above parameters into the formula in turn: First, calculate : ; Then take the fourth root: ; Then calculate : ; ; ; Take the cube root: ; Add the two terms: ; Then process : ; ; Final comprehensive operation: ; The result shows that the risk score value corresponding to the aggregation block is approximately 51.12. The higher the value, the larger the area of the aggregation block, the denser the number of color difference points, or the more significant the brightness difference. A threshold comparison can be set later to screen whether it is regarded as a high-risk area. For example, a threshold of 45 can be set. If then it is determined as a high-risk color difference area. Accordingly, the score value of 51.12 obtained in this example is significantly higher than 45. Therefore, this area is classified as a high-risk color difference area in the final evaluation.

[0058] The steps to obtain the color difference detection model are as follows: Collect color deviation images of past MicroLED displays, extract the color difference variance of the RGB channels, the number of abnormal pixels, the edge change amplitude, the discrete value of the gray channel, and the extreme value of the regional saturation of each frame of the image to form an image feature structure set; Based on the image feature structure set, calculate the image training effectiveness score value. The calculation formula is: ; where , , are the color difference variance values of the red, green, and blue channels respectively, is the discrete value of the gray channel, is the number of abnormal pixel points, is the maximum saturation value within the region, is the minimum saturation value within the region, is the number of continuous edge segments in the image, is the image training effectiveness score value; According to the image training effectiveness score value, screen the image frames for deep learning training to obtain the color difference detection model.

[0059] Specifically, after collecting the color deviation images of past MicroLED displays, it is necessary to centrally read the pixel information of each frame of the image. During the reading process, first confirm the resolution of each image file. For example, in an environment of 1920×1080, the abscissa index is defined as 0 to 1919, and the ordinate index is defined as 0 to 1079. Whenever a pixel point is read, its RGB value is obtained and scanned row by row in combination with the brightness data. In order to retrieve the color difference variance, the number of abnormal pixels, the edge change amplitude, the gray channel discrete value, and the regional saturation extreme difference value, it is necessary to first construct several data tables to record the coordinates, channel values, and edge position determination results of all pixels in the image. Then, by observing the channel fluctuation characteristics of some pixels, it is judged whether to include them in the abnormal pixel list. If any channel of a pixel has a significant jump compared with the reference pixel within the adjacent row and column range, it is counted in the abnormal pixel statistical table. Subsequently, the number of abnormal pixels in the entire image is counted and summarized to obtain the number of abnormal pixels. To observe the edge change amplitude, the gray gradient differences between adjacent pixels in the same row or the same column are compared, and the pixel position where the maximum gradient is located is recorded. Then, the maximum gradient value is included in the cumulative amount of the edge change amplitude. If there are more than 20 consecutive pixels of some similar pixels in the image and the brightness change is stable, special marking can be performed. The acquisition of the gray channel discrete value is based on the gray value of each pixel for variance or standard deviation calculation to obtain the distribution. For example, when reading each record and comparing it with the pre-set gray value range of 0 to 255, if it exceeds a large range, it is regarded as an invalid point and excluded. Finally, when collecting the regional saturation extreme difference value, it is necessary to first obtain the maximum and minimum saturation values of several block regions in the image, extract the saturation component in the color space such as HSI or HSV through per-pixel operation, and then subtract the maximum and minimum saturation values of each region. When these differences exceed the predetermined 60, it may indicate obvious local color deviation, so the corresponding extreme difference value is recorded and compared with other regions. All the above indicators are combined into a complete image feature record and stored in the image feature structure set. Repeat the operation until all the past collected image files are processed to obtain a feature structure set containing multiple frames of images.

[0060] Formula: , The benefit of the formula is that it incorporates the color difference variance of the RGB three channels, the gray channel discrete value, the number of abnormal pixels, the saturation component difference, and the image edge information into the scoring structure at the same time. Through operations such as polynomial summation, logarithmic operation, and square root, the multi-dimensional indicators are comprehensively measured, thereby quickly distinguishing the effectiveness degree of different images in training at a numerical level.

[0061] The steps to obtain the parameter are as follows: This parameter represents the variance value of the color difference in the red channel of the image. After extracting the red values of all pixels in the entire frame of the image, the deviation from a certain reference mean needs to be calculated, and the results are squared and summed up and then divided by the total number of pixels to obtain the variance value. The specific process can be expressed as: , where is the red component value of the th pixel point, is the average value of the red components of this image, is the number of image pixels. For example, in an image with a resolution of 1920×1080, there are 2,073,600 pixels in total. After inductive statistics, the average value of the red channel is 85.2, and the sum of the squared deviations of all pixels from this mean is calculated, then divided by 2,073,600 and then square-rooted to obtain .

[0062] The steps to obtain the parameter are as follows: This parameter corresponds to the variance value of the color difference in the green channel. Similar to , it also needs to be obtained by sampling the green components of each pixel in the image and calculating the variance. It can be estimated in the way of to obtain the squared value of , is the average value of the green components, is the green component value of a single pixel. If the average value of the green components is approximately 90.1 and is substituted into it, is obtained.

[0063] The steps to obtain the parameter are as follows: This parameter is the variance value of the color difference in the blue channel. It is necessary to read the blue component values in the entire frame of the image and calculate the mean value , then according to , obtain the squared variance, and then square-root the result to get .

[0064] The steps to obtain the parameter are as follows: This parameter is the discrete value of the grayscale channel. It is necessary to first convert the image from the RGB space to the grayscale space, collect the grayscale brightness of each pixel point, and then calculate the degree of dispersion of the entire frame of the image in the following form: , and square-root the result to obtain , where represents the average value of the grayscale values of all pixels. When obtaining it, it is necessary to ensure that a unified weighting coefficient is used when performing grayscale calculations, , for example, in actual calculations, is obtained.

[0065] The steps to obtain the parameter are as follows: This parameter is the number of abnormal pixel points, which needs to be statistically calculated based on the abnormal pixel marks recorded in the sampling stage. When obtaining it, the number of rows of the abnormal pixel table generated in the early stage can be summarized. Abnormal pixels usually consist of some points with overly large edge gradients or color mutations. If 480 abnormal pixel points are statistically counted through full-frame scanning, then 。

[0066] The steps to obtain the parameter are as follows: This parameter is the maximum saturation value within the region of the image. The image needs to be divided into several blocks, and then the pixels of each block are converted using the HSV or HSL color model, and the saturation component is extracted , After retrieving the maximum saturation value of all blocks in the same image, the global maximum value of the entire image is taken as , If it is found through specific calculations and traversals that the highest saturation of the full-frame image is 0.95, then 。

[0067] The steps to obtain the parameter are as follows: This parameter represents the minimum saturation value within the region of the image, similar to , but the minimum value is taken. If the lowest value of the saturation component of some blocks is 0.12 after statistics, then 。

[0068] The steps to obtain the parameter are as follows: This parameter represents the number of continuous edge segments in the image. After extracting the edge information, the continuous edge segments need to be counted. In the same frame of image, the edge pixels that are connected to each other are determined as one continuous segment, and then the total number of independent edge segments is counted one by one. If it is found that there are 34 continuous edge segments after multi-channel detection and connectivity search, then 。

[0069] Calculation process: Taking a frame of detected image as an example, assume , then first calculate: ; Square each value and then accumulate: , The sum is , Then take the square root: , Then add , Then add , The sum of the three: ; Then calculate , When , then , Multiply by 44.188 obtained previously: ; The result shows that the effectiveness score of this image in training is approximately 26.78. Generally, the higher the score, the more the multi-dimensional indicators such as color difference information, edge structure, and saturation in the image can reflect the differences. If a score threshold of 25 is set for the deep learning training process, the current image significantly exceeds 25, and it can be included in the training set to strengthen the support of diverse training data for the color difference detection task.

[0070] According to the image training effectiveness score value obtained previously, it is necessary to compare each item by item among all the collected color deviation images and select the image frames that meet the screening conditions. During the execution process, first list the score values extracted from each frame of the image in descending order, and remove the images with score values lower than a certain threshold from the candidate set, so as to obtain a list of candidate images. To further complete the deep learning training, it is necessary to perform labeling processing on these selected images, including confirming the boundary of the possible deviation area in the image and its corresponding RGB value distribution. Then, before training, batch archiving operations will be performed on the image frames, and the acquisition time, acquisition channel, and the device batch code matched will be noted in the information record of each frame of the image. If pixel information loss is found in the image, it will be decided whether to fill it in according to the recorded coordinate position. Finally, after all the selected images are ready, the image frames together with their supporting labels will be loaded into the artificial neural network training process. The training process will read all the packaged image frame data at one time at the input end, and update the weights of each pixel feature according to the previously defined order of convolutional layer, pooling layer, and fully connected layer. After processing a certain number of image frames, the error of the current model will be backpropagated and corrected, and several rounds of iterations will be gradually accumulated, so that the model converges to a better state in the discrimination process of the color difference distribution characteristics. The design of the convolutional structure will be based on the size of the input image. For example, a 7×7 convolutional kernel is selected to extract local features, and then the negative values are suppressed to 0 through the ReLU activation function to form a more representative feature map inside the network. At the same time, during the middle and late stages of training, the accuracy and loss values at different training rounds will be recorded, and the training will be stopped when the accuracy is higher than 95%. After multiple rounds of operations and discriminations, the final color difference detection model can be obtained. If you want to use this model in actual production later, it is necessary to input the target image into the network during the inference stage and let the network output the corresponding color difference determination result. Thus, a complete training model for MicroLED color deviation detection can be obtained.

[0071] The steps to obtain the color difference recognition result are as follows: Deploy the color difference detection model to the production line; Obtain the image data of the high-risk color difference area and input the image data of the high-risk color difference area into the color difference detection model; The color difference detection model processes the image data of the input high-risk color difference areas, analyzes and identifies the types and degrees of color differences, and generates color difference recognition results.

[0072] Specifically, obtain the image data of the high-risk color difference areas, input the image data of the high-risk color difference areas into the color difference detection model. For the list of high-risk color difference areas formed in the previous detection process, it is necessary to locate the image content on the screen according to the coordinate range of each area. If the abscissa of the area is between 100 and 200 and the ordinate is between 300 and 360, then read the pixel data with row and column indices in the corresponding interval and capture them at the same interval as the global image acquisition frequency. Each high-risk area may contain about hundreds to thousands of pixel points. In order to completely cover the image features, this area can be continuously collected in multiple frames to generate a batch of local images with timestamps. Then, when processing the data, first check whether there is a situation where the brightness saturation threshold or the pixel channel deviation threshold is exceeded in the acquisition record. For example, the previously set brightness saturation threshold can be 500 cd / ㎡, and the channel deviation threshold can be 50%. When specifically obtaining, it is necessary to compare each registered pixel brightness or channel offset value item by item, and record all the exceeded limits in the items to be analyzed. Next, send all the local images that exceed the threshold to the model inference link in sequence. If there are high-risk phenomena in multiple areas at the same time, they will be queued and input according to a fixed priority. By this way of first positioning and then screening, the model can be concentrated on detecting key areas, thus saving overall computing resources. Subsequently, convert these area images into the form of data matrices in sequence and place them at the input end of the convolutional network to view the discriminant output of the model for channel fluctuations and pixel anomalies. When necessary, record the sampling frequency and the number of sampling items by editing a dynamic monitoring list, and at the same time archive the collected data into an area image library for subsequent further comparison and evaluation. After completing the above process, obtain the area image data obtained this time and input it into the model to complete the detection of the color difference situation.

[0073] The color difference detection model processes the image data of the input high-risk color difference regions, analyzes and identifies the types and degrees of color differences, and generates color difference recognition results. When starting to perform inference on these images, it first reads the pixel matrix of the regional images and performs calculations according to the convolution kernels of the model. The red, green, and blue channel components of each pixel are multiplied by the weights stored during training respectively, then the bias term is added and a non-linear transformation is performed. If it is found within a certain kernel window that the channel gradient between adjacent pixels increases to 50% of the previously set channel deviation threshold, a relatively high output activation value will be triggered after convolution, and a prominent mark will be formed in the feature map of the network. The fully connected layer of the model will synthesize the local responses of each kernel to obtain the confidence value for the color difference judgment. Subsequently, the color difference types are classified into categories such as brightness drift, gray scale anomaly, or color channel imbalance through the classification layer. Each pixel point is marked with a specific type label, and further a quantitative score is provided for the detected degree of color difference at the end of the network. For example, it is set that 0 to 10 is mild, 10 to 30 is moderate, 30 to 60 is relatively severe, and above 60 is high intensity. If it is subsequently detected that the scores of some pixels are higher than 30, they will be uniformly included in the severe level region, and the row and column positions of these pixels will be recorded in the list. The entire inference process will be carried out in a batch inference manner on the GPU or other computing hardware. After processing a batch of image frames each time, the corresponding color difference recognition results, including the type determination and color offset amplitude of each pixel, will be written into the result set. If it is detected that the channel differences in the image are concentrated in the green and blue channels, it is determined as a cold color offset; otherwise, if they are concentrated in the red and green channels, it is determined as a warm color offset. Finally, according to the statistical results of each frame of image, a result list is output to the operator for verification when the processing is completed, and the necessary data support is provided for subsequent automated screening and product grading. After completing the recognition of all high-risk color difference region images, the final color difference recognition results are obtained.

[0074] The above is only the preferred embodiment of the present invention, and it does not limit the present invention in other forms. Any person skilled in the art may use the technical content disclosed above to make changes or modifications into equivalent embodiments with equivalent changes and apply them to other fields. However, as long as it does not depart from the technical solution content of the present invention, any simple modification, equivalent change, and modification made to the above embodiments based on the technical essence of the present invention still fall within the protection scope of the technical solution of the present invention.

Claims

1. A visual MicroLED color deviation recognition method for improving the yield rate, characterized in that, Including the following steps: Collect color data of the MicroLED display screen, continuously read the RGB color values of each point on the display screen, establish a color distribution database, calculate the color difference by comparing the color values of each point with the average color distribution, and obtain preliminary color difference data; for the preliminary color difference data, screen the data points with deviations exceeding the set threshold, and generate a list of significant color difference points; Perform spatial analysis on the list of significant color difference points, analyze the aggregation trend of color difference points by calculating the distance between each color difference point and its nearest neighbor point, and obtain the color difference aggregation analysis result; According to the color difference aggregation analysis result, identify the estimated color difference area and obtain the high-risk color difference area; Collect color deviation images of past MicroLED display screens as input, perform deep learning training, and generate a color difference detection model; Apply the color difference detection model to perform color difference detection on the high-risk color difference area and generate a color difference identification result.

2. The method for identifying color deviation of visual MicroLED for improving the yield rate according to claim 1, wherein The steps for obtaining the preliminary color difference data are as follows: Collect the RGB color values of each pixel point on the MicroLED display screen, record the color values, and form a color value data set; Based on the color value data set, calculate the average value of the RGB color values and generate a color difference data set; Based on the color difference data set, calculate the color difference influence factor of each point, and summarize the color difference influence factors to obtain the preliminary color difference data. The calculation formula is: ； Among them, represents the color difference influence factor of the i-th pixel point, represents the value of the i-th pixel in the j-th color of the RGB color spectrum, represents the average value of the average color distribution in the j-th color, represents the standard deviation of the j-th color.

3. The method for visually identifying color deviation of MicroLEDs for improving the yield rate according to claim 1, wherein The steps for obtaining the list of significant color difference points are as follows: Classify and sort the color difference influence factors of each pixel point in the preliminary color difference data, extract the color difference interval values between adjacent pixels according to the arrangement order, and generate a color difference interval sequence; Based on the color difference interval sequence, calculate the dynamic threshold. The calculation formula is: ； wherein, is the k-th color difference interval value, is the maximum value in the color difference interval sequence, is the minimum value in the color difference interval sequence, is the median of the color difference interval sequence, is the total number of color difference interval values, is the red component value of the k-th pixel, is the green component value of the k-th pixel, is the dynamic threshold; According to the dynamic threshold, screen the pixel points in the preliminary color difference data with color difference influence factors greater than the dynamic threshold to form a list of significant color difference points.

4. The method for identifying color deviation of visual MicroLED for improving the yield rate according to claim 1, wherein The steps for obtaining the color difference aggregation analysis result are as follows: Extract the two-dimensional coordinates of each point in the list of significant color difference points, calculate the Euclidean distance of each point to its nearest neighbor point, and obtain a distance data set; Based on the distance data set, calculate the spatial aggregation degree of color difference points. The calculation formula is: ； Among them, represents the distance between the i-th color difference point and its nearest neighbor, is the number of points in the distance dataset, is the spatial aggregation degree; According to the spatial aggregation degree, determine the spatial aggregation trend of color difference points and generate a color difference aggregation analysis result.

5. The method for identifying the color deviation of visual MicroLED for improving the yield rate according to claim 1, characterized in that The steps for obtaining the high-risk color difference area are as follows: Extract the boundary coordinate ranges of each aggregation block from the color difference aggregation analysis result, and count the number of color difference points within each boundary range to generate an aggregation area spatial coverage set; Based on the aggregation area spatial coverage set, calculate the risk score value of each area. The calculation formula is: ； Among them, and are the coordinates of the upper left corner and the lower right corner of the color difference aggregation area respectively, is the number of color difference points in this area, and are the coordinates of the two farthest color difference points in this area, and are the maximum and minimum values of the brightness of the color difference points in this area, is the risk score value of this area; According to the risk score value, screen the spatial areas with risk score values higher than the set threshold to form a high-risk color difference area.

6. The method for identifying the color deviation of visual MicroLED for improving the yield rate according to claim 1, wherein, The steps for obtaining the color difference detection model are as follows: Collect color deviation images of past MicroLED display screens, extract the RGB channel color difference variance, abnormal pixel quantity, edge change amplitude, gray channel discrete value, and regional saturation extreme difference of each frame of image to form an image feature structure set; Based on the image feature structure set, calculate the image training effectiveness score value. The calculation formula is: ； Among them, , , are the color difference variance values of the red, green, and blue channels respectively, is the grayscale channel discrete value, is the number of abnormal pixel points, is the maximum saturation value within the region, is the minimum saturation value within the region, is the number of continuous edge segments in the image, is the image training effectiveness score value; Filter image frames for deep learning training according to the training effectiveness score value of the image to obtain a color difference detection model.

7. The method for visually identifying the color deviation of MicroLEDs for improving the yield rate according to claim 1, wherein, The steps for obtaining the color difference recognition result are as follows: Deploy the color difference detection model to the production line; Obtain the image data of the high-risk color difference area, and input the image data of the high-risk color difference area into the color difference detection model; The color difference detection model processes the input image data of the high-risk color difference area, analyzes and identifies the type and degree of color difference, and generates a color difference recognition result.

Images