Visual MicroLED uniformity detection method for improving yield

Abstract

The invention relates to the technical field of image recognition, in particular to a visual MicroLED uniformity detection method for improving the yield, which comprises the following steps of: acquiring original image data of a MicroLED screen through a camera, acquiring the original image data, performing fine region division on an original image, capturing color change and brightness difference of a micro region on the surface of the display screen, and determining the uniformity of the MicroLED screen according to the color change and the brightness difference. The image segmentation precision is improved, and the positioning of the abnormal area of the screen is more accurate; in combination with deep color analysis of each area, color deviation can be accurately calculated, the position and color features of the abnormal area are tracked in real time, and the sensitivity and accuracy of anomaly detection are improved; and when an abnormal area is detected, the exposure condition can be dynamically adjusted to eliminate ambient light interference factors and ensure the accuracy of subsequent data acquisition.

Description

Technical Field

[0001] The present invention relates to the field of image recognition technology, and in particular to a visual MicroLED uniformity detection method for improving product yield. Background Art

[0002] The MicroLED uniformity detection method is to use image recognition technology to detect the MicroLED display screen, analyze and evaluate the brightness and color uniformity of the MicroLED devices on the display screen.

[0003] However, in actual operation, existing technologies usually rely on a rough analysis of the overall image without performing fine area division processing, ignoring the tiny but critical color and brightness differences between different areas of the screen. This leads to insufficient sensitivity to tiny abnormal areas on the screen surface, resulting in increased production costs and reduced product qualification rates, affecting the accuracy of overall product quality control. Summary of the Invention

[0004] The purpose of the present invention is to address the shortcomings of the prior art and propose a visual MicroLED uniformity detection method for improving the yield rate.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a visual MicroLED uniformity detection method for improving yield rate, comprising the following steps: Obtain the original image data of the MicroLED screen through the camera to obtain the original image data; Based on the grayscale value and coordinate information of the original image data, the image is divided into regions and then a watershed region growing strategy is used to generate segmented region data; color analysis is performed on each region in the segmented region data, and the color deviation of each region is calculated to obtain a regional color analysis result; Based on the regional color analysis result, identifying the region with abnormal color difference, recording the location and color characteristics of the region with abnormal color difference, and generating a color difference abnormality record; determining whether the region with abnormal color difference is caused by the ambient lighting conditions, and if so, adjusting the camera exposure time, capturing an image of the abnormal region again, and obtaining an adjusted image of the abnormal region; Applying an ESRGAN super-resolution model to the adjusted abnormal area image to perform resolution enhancement to obtain an enhanced detail image; and performing defect recognition on the enhanced detail image to generate a defect recognition result.

[0006] Preferably, the steps of obtaining the original image data are: The camera is configured to face the Micro LED screen, and the parameter settings of the camera, including focal length and exposure time, are adjusted to obtain adjusted camera settings; Based on the adjusted camera settings, the camera captures the light emitted by the Micro LED screen and converts the light into a digital image file to obtain raw image data that has not been processed. According to the raw image data that has not been processed, the integrity of the image and the quality of the captured image are checked to determine that the data file is not damaged, and the raw image data is obtained.

[0007] Preferably, the step of obtaining segmented region data is: The gray value and coordinate information of each pixel are extracted from the raw image data, and an initial connection graph is established between all pixels based on the eight-neighborhood rule. The connection edges with a gray value difference within a predetermined threshold are marked to obtain an initial pixel connected graph structure. According to the initial pixel connected graph structure, the local consistency scores of a plurality of candidate subgraph regions are obtained, and the calculation formula of the local consistency score of the candidate subgraph region is: ; Wherein, is the local consistency score of the th candidate subgraph region, is the set of all connected pixel pairs in the subgraph region, and are the gray values of pixel and pixel , respectively, is the number of connection edges in the region. According to the local consistency score of each candidate region, the pixel subgraph with a local consistency score higher than the median of the distribution is marked as an initial region mask. The initial region mask is subjected to a watershed region growing strategy and the boundary is calibrated to generate the segmented region data.

[0008] Preferably, the step of obtaining the region color analysis result is: Each region in the segmented region data is traversed in turn, and the RGB three-channel values of all pixels in each region are extracted and converted to CIELAB space color values. The value set of the L channel, a channel and b channel corresponding to each pixel in the region is counted to obtain a region color component dataset. Based on the region color component dataset, the color deviation index of each region is calculated, and the calculation formula of the color deviation index is: ; Wherein, represents the color deviation index of the Color deviation index of each area, and Respectively represent the maximum and minimum values ​​of the L channel in the area, Indicates the total number of pixels in the area. and Represent the mode values ​​of channel a and channel b in the area, represents the average value of the L channel, represents the variance of the L channel, represents the skewness coefficient of channel a, represents the average value of b channel; According to the color deviation index, all regions are numerically sorted according to the color deviation index and regions whose color deviation index is greater than the distribution mean are extracted as regional color analysis results.

[0009] Preferably, the steps for obtaining the color difference abnormality record are: Based on the regional color analysis results, the color abnormality intensity value of each region is calculated using the following formula: ; in, Indicates the The color abnormality intensity value of the area, Indicates the maximum value of the pixel of the R channel in the area, Indicates the number of pixels in the area that have a brightness higher than the average brightness of the entire image. Indicates the minimum value of the G channel pixel in the area, Indicates the number of color change points of all pixels in the area, represents the horizontal spatial span of pixels in the region, Indicates the maximum value of the pixel of the L channel in the area, Indicates the number of pixels in the area that are lower than the median of the a channel of the entire image. Indicates the minimum value of the B channel pixel in the area, Indicates the extreme saturation value of all pixels in the area. Indicates the vertical spatial span of pixels in the region; According to the color abnormality intensity value of each area, the intensity threshold is set to extract the abnormal area, and the abnormal area coordinate value and channel attribute combination are marked into the abnormal record file structure to generate the color difference abnormality record.

[0010] Preferably, the steps of acquiring the adjusted abnormal area image are: According to the color difference abnormality record, the camera exposure adjustment factor is calculated, and the calculation formula is: ; in, Indicates the The exposure adjustment factor for abnormal areas, Indicates the grayscale dynamic range of the image in this area, Indicates the lowest pixel brightness value in the area. represents the brightness variance within the image frame of the region, Indicates the ambient light intensity corresponding to the abnormal frame, Indicates the average brightness of the channels in the image frame of this area, Indicates the ratio of edge pixels in the color difference channel of the area, Indicates the number of overexposed pixels in the image frame of this area; According to the camera exposure adjustment factor, a camera exposure time reset rule is set, the exposure value is set to the time inversely calculated by the product of the current exposure time and the camera exposure adjustment factor, and the image frame corresponding to the abnormal area is recaptured to obtain an adjusted abnormal area image.

[0011] Preferably, the steps of acquiring the enhanced detail image are: Importing the adjusted abnormal area image, unifying the image data format, and adjusting the image size to the input resolution specified by the ESRGAN model according to the requirements of the ESRGAN model to obtain preprocessed image data; The ESRGAN model is loaded, and the preprocessed image data is used as input to perform feature extraction and amplification to generate an enhanced detail image.

[0012] Preferably, the steps for obtaining the defect identification result are: The convolutional neural network model is loaded, the enhanced detail image is input into the convolutional neural network model, and the multi-layer feature recognition capability of the convolutional neural network model is used for analysis, image features are extracted layer by layer, and potential defect areas are identified to obtain defect recognition results.

[0013] The present invention provides a visual MicroLED uniformity detection system for improving yield, comprising: a processor and a memory, wherein the memory is used to store a computer program, and the processor is used to call and run the computer program stored in the memory, so that the visual MicroLED uniformity detection system performs a visual MicroLED uniformity detection method.

[0014] Compared with the prior art, the advantages and positive effects of the present invention are: The present invention directly collects MicroLED screen image data through a camera, performs detailed regional division on the original image, captures color changes and brightness differences in tiny areas on the display screen surface, improves image segmentation accuracy, and makes the positioning of abnormal areas on the screen more precise; combined with in-depth color analysis of each area, it can accurately calculate color deviations and track the position and color characteristics of abnormal areas in real time, improving the sensitivity and accuracy of abnormality detection; when an abnormal area is detected, the exposure conditions can be dynamically adjusted to eliminate ambient light interference factors and ensure the accuracy of subsequent data collection; in addition, image super-resolution enhancement technology is further used for local abnormal areas to reconstruct abnormal details with high quality, explore and amplify the detailed features of screen defects, and achieve higher-resolution defect feature extraction, ensuring that subtle defects can also be accurately identified, reducing the misjudgment rate and missed detection rate of MicroLED screens in the production inspection link, and improving the yield rate and overall inspection efficiency of MicroLED screens. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a schematic diagram of the steps of the present invention. DETAILED DESCRIPTION

[0016] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, 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 intended to limit the present invention.

[0017] See also Figure 1 The present invention provides a technical solution, a visual MicroLED uniformity detection method for improving product yield. From a program perspective, the execution body of the process can be a program installed on an application server or application terminal. It is understood that the method can be executed by any device, equipment, platform, or device cluster with computing and processing capabilities. In the embodiment of this specification, the visual MicroLED uniformity detection method for improving product yield includes the following steps: Obtain the original image data of the MicroLED screen through the camera to obtain the original image data; Based on the original image data, the image is divided into multiple regions to generate segmented regional data; each region in the segmented regional data is subjected to color analysis, the color deviation of each region is calculated, and the regional color analysis results are obtained; Based on the regional color analysis results, the region with color abnormality is identified, and the location and color characteristics of the region with color abnormality are recorded to generate a color abnormality record. Based on the color abnormality record, it is determined whether the color abnormality is caused by the ambient lighting conditions, the camera exposure time is adjusted, and an image of the abnormal region is captured again to obtain an adjusted image of the abnormal region. The ESRGAN super-resolution model is applied to the adjusted abnormal area image to perform resolution enhancement to obtain an enhanced detail image; defect recognition is performed on the enhanced detail image to generate a defect recognition result.

[0018] The steps for obtaining the original image data are: Configure the camera to face the MicroLED screen and adjust the camera parameters, including focal length and exposure time, to obtain the adjusted camera settings. Based on the adjusted camera settings, the camera captures the light emitted by the MicroLED screen and converts the light into a digital image file to obtain unprocessed raw image data; Based on the unprocessed original image data, the integrity of the image and the quality of the captured image are checked to determine that the data file is not damaged, thereby obtaining the original image data.

[0019] Specifically, based on the camera device positioned in front of the MicroLED display area, the actual distance from the camera to the screen is first measured and recorded as , and then measure the ambient light intensity with an external illuminance meter and record it as Then, based on the initial focal length range set by experience, the interval from 0.5 meters to 2.0 meters can be selected and adjusted multiple times in steps of 0.1 meters. During this period, the brightness distribution of the picture captured by the camera after each adjustment can be compared, and the complete interval of the picture grayscale value between 0 and 255 is regarded as the basic reference. If it is found that the grayscale histogram is concentrated in the high interval and the proportion exceeds 30%, it is judged to be overexposed. If the grayscale histogram is concentrated in the low interval and the proportion exceeds 30%, it is judged to be dark. On this basis, the exposure time is increased step by step from 5 milliseconds to 30 milliseconds at intervals of 5 milliseconds. For each level of exposure time, an imaging test is performed with the current focal length, and information such as the proportion of bright pixels, the proportion of dark pixels, and the overall grayscale variance value of the picture are recorded one by one. If it is monitored that the proportion of bright pixels is greater than 50% or the proportion of dark pixels is greater than 50%, it means that the current setting does not meet the requirements. At this time, the measured ambient light intensity can be used to determine the exposure time. and distance To simplify the calculation, let And it fluctuates by 1 to 3 milliseconds on this basis, It is an empirical coefficient selected based on the characteristics of the equipment and the results of past tests, such as The test shooting results in the current environment are combined to verify whether they are within an acceptable range. Through multiple sets of comparisons, the focal length range and exposure time range are gradually narrowed down, and finally the most appropriate adjustment plan is locked in, thereby obtaining executable camera parameter settings.

[0020] Based on the camera parameter settings obtained above, during the execution process, the light emitted by the MicroLED screen is collected by a shooting device with confirmed focal length and exposure time. The light signal is continuously captured at a sampling frequency of 30 frames per second and converted into a digital format. In this process, the grayscale resolution range of 8 bits or 10 bits can be pre-set according to the pixel bit depth, and the brightness value corresponding to each pixel within the resolution range is saved as an integer sequence from 0 to 255 or 0 to 1023. If more than 5% of the pixel grayscale values ​​in a single frame of data deviate from this range, it is considered an abnormality and needs to be re-collected. At the same time, by comparing the average brightness of each frame of the image and variance To judge the sampling stability, Between 120 and 200 and When the variance does not exceed 500, it can be regarded as a stable acquisition state. If the pixel variance of more than three consecutive frames exceeds this benchmark value, it is necessary to check for possible jitter or repeat the capture to ensure completeness. During this period, there is no need to introduce additional complex intermediate processing algorithms. It is only necessary to retain the original brightness and color information of each frame as the basis for subsequent analysis. After summarizing the above steps, the sequence formed by each frame is summarized into an image data set without any additional correction, and finally the unprocessed original image data is obtained.

[0021] According to the unprocessed raw image data obtained above, the effective pixel number of the image file is checked frame by frame during the execution process to see if it matches the expected resolution. For example, if the resolution is set to 1920×1080, the total number of pixels should be 2073600. After reading the pixel array of each frame, the actual number of pixels obtained is counted and the statistical result is ensured to fluctuate within 5 pixels above and below 2073600. If more than 5 pixels are missing or an unresolvable anomaly occurs, it is considered that the data integrity of this frame does not meet the standard. Then check whether the grayscale value and color value of the image fall within the defined value range. For example, For example, for 8-bit grayscale, the range can be set to 0 to 255. If the grayscale of any pixel is less than 0 or greater than 255, it means that the data is damaged or not recorded correctly, and the frame is marked for subsequent investigation. At the same time, the average brightness and color distribution of each frame are read and compared with the set range. The average brightness can be controlled between 100 and 200 based on the previous debugging results. If the brightness deviation of a frame exceeds 30, it is marked as an abnormal state. After completing all tests, if all frame data meet the above integrity and quality requirements, the data is considered to be intact, and the original image data that can be used later is finally obtained.

[0022] The steps to obtain the segmented region data are: Extract the grayscale value and coordinate information of each pixel from the original image data, establish an initial connection graph between all pixels based on the eight-neighborhood rule, mark the connection edges whose grayscale value differences are within a preset threshold, and obtain the initial pixel connectivity graph structure; According to the initial pixel connectivity graph structure, the local consistency score of the local connected subgraph is calculated using the following formula: ; in, For the The local consistency score of the candidate sub-image region, is the set of all pixel pairs with connected edges in the sub-graph area, and Pixels and pixels The gray value of is the number of connected edges in the region; According to the local consistency score of each candidate region, the pixel sub-image with a local consistency score higher than the distribution median is marked as the initial region mask, and the watershed region growing strategy is executed and the boundary is calibrated to generate the segmented region data.

[0023] Specifically, based on the grayscale value and coordinate information of each pixel read from the original image data, during the execution process, all pixels are first organized into a two-dimensional distribution table in row priority order, where the row mark represents the vertical order of the image, and the column mark represents the horizontal order of the image. Then, each pixel is traversed in a row-by-row scanning manner and its adjacent pixels are selected as comparison objects. The selection method of adjacent pixels can follow the adjacency relationship in the horizontal, vertical and diagonal directions. This means that for non-edge pixels inside the image, its surroundings usually contain eight possible adjacent pixels. Before determining whether there is a valid connection between the pixel and its adjacent pixels, it is necessary to compare them based on a pre-defined grayscale difference threshold. The threshold can be obtained from the grayscale statistical distribution of multiple sample images and can be corrected through multiple groups of tests. For example, if the threshold is set to 15 for images with rich details, it is because according to the 500 image samples collected in the early stage, it is detected that the main targets in the grayscale range of 0 to 255 usually have a grayscale change of less than 15, and when the threshold is exceeded. The brightness difference between pixels is large, so based on this, the effective connections between pixels can be made relatively uniform. During the traversal process, pixel pairs with a grayscale difference less than or equal to 15 can be marked as associated relationships, and a connection table is established in the memory to record a list of all pixel pairs with associated relationships. If a pixel point does not have any adjacent pixels that meet the difference condition, it is considered an isolated pixel and is not marked as a connected relationship. After completing all traversals, an initial connection graph is obtained. Each connecting edge in the graph represents that two pixels are similar in grayscale. In addition, in order to avoid including low-probability noise in the similarity category, the pixel mean, variance, and noise frequency can be comprehensively considered when selecting the threshold. For example, in the early statistics, it may be found that the noise accounts for 3% of the total pixels. If the grayscale difference of the noise points is concentrated between 30 and 50, the noise impact can be reduced by lowering the threshold by 5 to 10. After completing this step, all pixels and their connection relationships are concentrated into a data structure, and the initial pixel connectivity graph structure is finally obtained.

[0024] formula: The benefit of the formula is that it can directly quantify the grayscale difference between adjacent pixels into a score, and form an overall score for the local area by cumulatively summing all pixel pairs. In this process, the smaller the grayscale difference, the greater the weighted result, so that the pixel set with similar grayscale distribution can be evaluated higher. If some pixel pairs are too far apart in grayscale, the value of the corresponding item will be reduced and affect the final average score, thereby presenting the overall situation of pixel consistency in the area. It has auxiliary significance for identifying homogeneous areas in image processing and region division, and by introducing The adaptive measurement of the difference between pixels is realized, and the constant 1 is used in the molecular part so that any non-zero difference will affect the score accordingly. This can highlight the importance of the similarity of the gray scale distribution through simple calculation, and provide a quantitative basis for subsequent regional judgment.

[0025] The parameter acquisition step is to locate all possible pixel pairs in the first region in the initial pixel connected graph structure obtained in the previous step, and then integrate these pixel pairs into a set in the form of , which is called , and each represents that the pixels and have some adjacent relationship in this region. In order to determine which pixel pairs are included in the range, it is necessary to check each pixel in the first region to see if they meet the adjacent condition, if they meet, add to the set, if not, skip. When making the judgment, the gray scale difference threshold set in the previous step can be combined with the boundary information in the region to confirm, for example, in a certain processing process, if the total number of pixels in this region is 100 and the adjacent judgment finds that there are 56 pairs of pixels with connecting edges, then can be the set of 56 pixel pairs.

[0026] The parameter acquisition step is to extract the gray scale value of the pixel from the gray scale data obtained in the previous step and quantize it to an integer between 0 and 255. For some images with 10-bit or 12-bit depth acquisition, the original gray scale range of 0 to 1023 or 0 to 4095 can be mapped to 0 to 255 through a normalization formula, and the mapping method can be written as , where is the original gray scale value, 16 is the scaling factor, and the integer result is obtained through the floor operation. At this time can be used as the value in the formula for subsequent use. If 8-bit gray scale values are obtained directly during acquisition, then can be directly equivalent to the collected value.

[0027] The parameter acquisition step is to perform the same gray scale extraction and quantization operation on the pixel adjacent to the pixel , and finally obtain an integer value in the same range as . If the pixel with coordinates (10, 15) in the same image is With a grayscale value of 130, then in the formula Take 130, in order to facilitate the parallel processing of multiple images or multiple areas, all pixels can be sequentially The grayscale values ​​are put into a table and compared with the pixels The grayscale table is recorded accordingly, and then the combination of pixel coordinates is used to form a pixel pair set. When grayscale acquisition with higher bit depth is involved, the same scaling method as described above can be used to complete quantization. For example, under 12-bit acquisition conditions, if the original value range is 0 to 4095, it can be used The corresponding 8-bit grayscale value is obtained by .

[0028] The steps to obtain the parameters are as follows: The number of pixel pairs that meet the connection conditions can be counted in the set. For pixel pairs that satisfy the adjacent criteria, we can let .

[0029] Calculation process: The first step is to find all pixel pairs in the established connection relationship for a certain area that has been divided in the target image, and let the set of these pixel pairs constitute , calculate the cardinality of the set and get ; The second step is to Each of calculate , for example when , hour, ; The third step is to Add them one by one, if there are For pixel pairs, each pair of pixels can get a score in the specific calculation, and all the scores are accumulated to get ; The fourth step is to divide the above accumulated result by , thus obtaining For example, suppose that in a certain operation, 20 pairs of pixels are found in a region, and the cumulative result is 9.6, then , treating this value as And record.

[0030] The results show that when When the value is close to 1, it means that the similarity of pixel grayscale values ​​in the region is high. If the value is lower than 0.3, it may indicate that there is a large grayscale difference within the region. In this step, if multiple regions are calculated separately, The numerical values ​​of each area are compared. The area with the highest value may represent a pixel group with relatively high uniformity, while the area with lower value indicates that there may be a mutation in the grayscale distribution. This can be combined with the subsequent region growing strategy to complete more fine-grained image segmentation or anomaly detection.

[0031] According to the numerical list generated by the local consistency score of each candidate region, the score values ​​of all candidate regions are first sorted in ascending or descending order during the execution process, and the score in the middle position of the score distribution sequence is confirmed to be the median of the batch. If the number of candidate regions is odd, the value in the middle position of the sequence can be directly selected as the median after sorting. If the number of candidate regions is even, the average of the two middle scores can be defined as the median. After completing the statistics, the pixel sub-images with scores higher than the median are marked as the initial region mask in turn. At this time, the coordinate index record is required for each marked sub-image, and the region boundary and internal pixel range corresponding to the sub-image are explicitly identified in the actual image. The marked initial region mask is regarded as the seed point set of the subsequent watershed region growing strategy. Next, referring to the grayscale gradient partitioning method extracted from external data, a gradient matrix matching the image size is defined, in which each coordinate position stores a gradient value obtained by local difference operation, and then the watershed region growing is started. In the specific implementation, it can be performed around each initial mask position. Row pixel expansion: if the gradient value of the current pixel does not exceed the set range, the pixel is included in the growing region. For example, the gradient setting range is controlled between 0 and 50. This value refers to the actual measurement statistics of similar images in the early stage. If a block of pixels is detected to be still connected to the marked initial mask within the range of gradient value less than 50, the block of pixels is determined to belong to the same region. Otherwise, if the gradient value reaches 60 or above, the growth in that direction is terminated and the pixels at the boundary are treated as stagnant pixels. When multiple initial region masks overlap during the growth process, their scores can be compared and a decision can be made to retain or remove them. For example, if two regions have scores of 0.65 and 0.72 respectively and intersect at image coordinates (100, 200) to (110, 210), the pixels in the intersection can be assigned to the region with a score of 0.72 according to the rule of giving priority to the region with the higher score. After the expansion of all marked masks is completed, the boundary of all edge pixels is demarcated and recorded as the final region edge coordinate set. The aforementioned grown region pixel set is added to the edge coordinate sequence to obtain the segmented region data.

[0032] The steps to obtain the regional color analysis results are: Traverse each region in the segmented regional data in turn, extract the RGB three-channel values ​​of all pixels from each region and convert them into CIELAB space color values, count the value sets of the L channel, a channel and b channel corresponding to each pixel in the region, and obtain the regional color component dataset; Based on the regional color component dataset, the color deviation index of each region is calculated using the following formula: ; in, Indicates the Color deviation index of each area, and Respectively represent the maximum and minimum values ​​of the L channel in the area, Indicates the total number of pixels in the area. and Represent the mode values ​​of channel a and channel b in the area, represents the average value of the L channel, represents the variance of the L channel, represents the skewness coefficient of channel a, represents the average value of b channel; According to the color deviation index, all areas are numerically sorted according to the color deviation index and the areas where the color deviation index is greater than the distribution mean are extracted as the regional color analysis results.

[0033] Specifically, based on the segmented area information obtained above, the coordinates and corresponding RGB three-channel values ​​of all pixels are selected from each area. After being extracted row by row, these RGB values ​​are processed according to the pre-established conversion standard. When converting, the reference white point D65 defined in the CIELAB color space is used and mathematical transformation is performed to map the RGB integer range to the floating-point range of the L, a, and b channels. In the statistical process, the red, green, and blue components of each pixel are first retrieved row by row and the coordinate identifier of the pixel is recorded. The RGB value of the pixel is converted into a CIEXYZ representation using a determined conversion matrix, and then the XYZ to L, a, and b calculation process is completed according to the CIELAB standard formula. If a value exceeding the RGB range of 0 to 255 is detected during the process, the abnormal information is recorded at the pixel entry and summarized separately. In order to ensure the accuracy of the value, the fixed-point operation method can be used to fix the number of decimal places to two. After the conversion is completed, all pixels in each area are scanned one by one and the L channel value sequence, the a channel value sequence and the b channel value sequence are collected. These sequences are classified and stored in the same data structure according to the pre-established format. When the traversal is completed and all pixels are converted, the collected L channel, a channel and b channel sequences are summarized into corresponding value sets based on the region granularity. They are managed accordingly in combination with the region index and the number of pixels, coordinate range and other information are recorded. When checking the validity of these data sets, when the number of pixels is less than ten, it can be regarded as an extremely small area and retained as usual. If multiple anomalies are detected in some areas during the conversion, the area is marked as a defect candidate area and further checked in the subsequent steps. After the summary is completed, the regional color component data set can be obtained. The regional color component data set includes: regional index or identification: records the unique number of each area for subsequent tracking and processing. Regional pixel coordinate range: lists the pixel coordinate interval or pixel set covered by the area to facilitate positioning in the image. L, a, and b channel value aggregation: The L, a, and b channel values ​​for all pixels in the region are collected one by one, forming a complete sequence or channel-by-channel array of the three-channel values. Internal statistics for the region, including the number of pixels, the maximum and minimum L channel values, the L channel mean and variance, and the mode, mean, and skewness coefficient of the a and b channels, are recorded to facilitate subsequent calculations of various color deviations and other analyses.

[0034] formula: The benefit of the formula is that it contains multiple channel dimension information. Reflects the relationship between the brightness range of the L channel and the number of pixels in the area, Establish a correlation between the most frequently occurring values ​​of channels a and b and the average value of channel L. The variance of the L channel and the skewness of the a channel are jointly calculated and combined with the average value of the b channel to form a comprehensive term. Multiple factors are aggregated into the same square root structure to achieve numerical fusion at the same scale.

[0035] The steps to obtain the parameters are as follows: first locate the L channel values ​​of all pixels in the area according to the regional color component data set in the previous step, find the maximum value and assign it For example, the maximum L value of the area collected in a certain time is 83.4, then Take 83.4.

[0036] The steps to obtain the parameters are: select the minimum value from the L value sequence of all pixels in the same area and assign it to , for example, the minimum L value detected in the same scene is 12.5, then Take 12.5.

[0037] The steps to obtain the parameters are: Count all pixels in the area and assign values , when the region boundary is fixed, the pixel coordinates within the region are accumulated row by row to obtain the total number of pixels in the region. If the region covers a pixel range of 200 in width and 100 in height, the corresponding number of pixels is 20,000. After the statistics are completed, this number is used as , for example, when the actual cumulative result is 19856, then =19856.

[0038] The parameter acquisition step is: select the value with the highest frequency in the a channel value sequence of all pixels in the area and take it as the mode. For example, in the a value sequence range of -128 to 127 in the area, the actual statistics show that the pixel with a value of 15 appears 642 times and has the highest proportion, then =15.

[0039] The steps to obtain the parameters are: process the b channel pixel values ​​in the same way as the a channel and find the b value with the highest number of occurrences, and assign it to When counting the b-axis values ​​in the region, if b=20 is found to occur the most and is counted as 1350 times, then =20.

[0040] The steps to obtain the parameters are: sum up the L values ​​of all pixels and divide by the total number of pixels , get the average value of L channel in this area, let this average value be given For example, when the sum of the L values ​​of the collected 19856 pixels is 1085920.8, let .

[0041] The parameter acquisition step is: for all pixel values of the L channel, first calculate the average value , then for each pixel L value, do operation and accumulate the sum, and then divide by to get the variance, if the sampled pixel L value sequence is , then set , this result is defined as , for example, when the sum of the squares of this batch of data reaches 159680.4 and is divided by 19856 to get 8.04, then =8.04.

[0042] The parameter acquisition step is: in the sequence of a values in this region, the skewness coefficient is calculated, first get the average of a values and get the deviation of each a value from the average, according to the definition of skewness coefficient , the calculation is completed, for example, the average value of this sequence is calculated to be 10.2 and the accumulated cubic deviation is 443215.6, the square deviation is 72150.4, etc., after being brought into the formula, the solution is about 0.37, then =0.37.

[0043] The parameter acquisition step is: add all b values in this region and divide by the total number of pixels , take the average and record it as , when the sum of the b value sequence collected in advance is divided by 19856 to get about 18.0, then =18.0.

[0044] Calculation process: First, sort the above parameters and bring them in one by one, let , , , , , , , , ; Second, calculate and do the square term, take , and square it to get ; Third, calculate , take , the absolute value is still 5.387; Fourth, calculate , take ; Fifth, add the above three items to get Finally, we get the square root of the value .

[0045] The results show that when When it is 2.354, it means that the color deviation comprehensive value obtained by superimposing the dynamic difference of L channel, the product of the mode of a channel and the mode of b channel, and the product of the variance of L channel and the skewness of a channel in this area is about 2.354. The larger the value, the more significant the deviation is. The smaller the value, the more uniform the color distribution is in this area. After the calculation is completed, it can be compared with other areas in the subsequent process. Perform a horizontal comparison to determine whether it falls into the high color difference area.

[0046] According to the color deviation index distribution list obtained above, the deviation index values ​​of all areas are first sorted one by one and arranged in ascending order. After reading the sorting results, the median position is counted and the corresponding value is selected as the reference mean. When the total number of areas is odd, the middle item is selected as the mean. When the total number of areas is even, the two adjacent values ​​are added and divided by 2 as the mean. At this time, all areas with deviation indicators greater than the mean are listed and marked one by one at the corresponding area index. Then, the L channel, a channel, and b channel component information corresponding to these marked areas are checked, and the overview of each area at the mean and variance levels is listed respectively. A threshold range is used to distinguish between large and medium deviation phenomena. In the example, when A threshold of 3.0 indicates that any area with a deviation index exceeding 3.0 is considered to be at a high level of color difference. This threshold is derived from the statistical distribution of a large number of actual display screen tests. Preliminary analysis determined the deviation of approximately 3,000 areas and calculated their ninth-order quantiles, resulting in a range of approximately 2.95 to 3.05. Therefore, 3.0 was selected as the baseline. Subsequently, a list of areas with a deviation index exceeding 3.0 was compiled into key analysis items. The specific location and pixel distribution characteristics of these areas within the entire image were further determined by combining the corresponding coordinate intervals. After the retrieval and comparison of all deviation indicators were completed, a sequence of key areas of interest was formed in the final result. The relevant statistical information was combined to complete the index delivery, which is the regional color analysis result.

[0047] The steps for obtaining color difference abnormality records are as follows: Based on the regional color analysis results, the color abnormality intensity value of each region is calculated using the following formula: ; in, Indicates the The color abnormality intensity value of the area, Indicates the maximum value of the pixel of the R channel in the area, Indicates the number of pixels in the area that have a brightness higher than the average brightness of the entire image. Indicates the minimum value of the G channel pixel in the area, Indicates the number of color change points of all pixels in the area, represents the horizontal spatial span of pixels in the region, Indicates the maximum value of the pixel of the L channel in the area, Indicates the number of pixels in the area that are lower than the median of the a channel of the entire image. Indicates the minimum value of the B channel pixel in the area, Indicates the extreme saturation value of all pixels in the area. Indicates the vertical spatial span of pixels in the region; According to the color abnormality intensity value of each area, the intensity threshold is set to extract the abnormal area, and the abnormal area coordinate value and channel attribute combination are marked into the abnormal record file structure to generate the color difference abnormality record.

[0048] Specifically, the formula: The benefit of the formula is that it utilizes a combination of multiple color and brightness components such as R, G, B, L channels, and a channel, and uses different parameter items in the numerator and denominator to differentiate the brightness and darkness characteristics, channel extreme values, pixel distribution, saturation and other information within the region, and closely combines the color deviation and brightness correlation into a quantifiable indicator. At the same time, it uses the absolute value and square root form to emphasize the prominent difference in the numerical value. Once there is an extreme channel value or a large brightness difference in the result, the indicator will show a significant improvement, thereby quickly distinguishing potential abnormal color areas and forming a numerical judgment basis for color anomalies in actual detection.

[0049] The steps to obtain the parameters are: first Completely traverse the R channel values ​​of all pixels in the area, record the maximum integer value and define it as For example, when tens of thousands of pixels are accumulated at a resolution of 1920×1080, the R values ​​of these pixels are scanned and the maximum value is locked at 245. =245.

[0050] The steps to obtain the parameters are: In each area, the number of pixels whose brightness is higher than the average brightness of the whole image is counted and recorded as For example, in an area of ​​600×400, 240,000 pixels are actually detected, and the average brightness of the whole image is 120.3. After screening, it is found that there are 68,000 pixels in the area with a brightness exceeding 120.3. =68000.

[0051] The steps for obtaining the parameters are: collecting and counting the G channel values ​​of all pixels in the area, and selecting the minimum value in the G channel as For example, in an 8-bit channel, control the G value between 0 and 255, put the G values ​​of all pixels in the area into an array and find the minimum value. If the result is 8, then =8.

[0052] The steps to obtain the parameters are: count the number of color change points of all pixels in the area and set it as The color change point refers to the pixel adjacent boundary where the R, G, B channels (or CIELAB channels) in the pixel distribution in the area show a significant jump. The color difference of all pixels in this area can be compared one by one by using a fixed adjacency criterion (such as eight neighborhoods). When the difference exceeds a specific threshold, it is considered a color change point and recorded and accumulated to For example, in an area of ​​300×300, the color change frequency between pixels is counted, and if 352 positions meet the jump standard, the color change frequency between pixels is counted. =352.

[0053] The steps to obtain the parameters are: The pixel span of each region in the horizontal direction of the image is measured, and the difference between the leftmost and rightmost pixel coordinates is recorded as , during execution, you can first scan the horizontal coordinates of all pixels in the area, find the minimum and maximum values, and then subtract them to get the actual horizontal span. For example, when the leftmost coordinate in the area is at x=200 and the rightmost coordinate is at x=800, then =600.

[0054] The steps to obtain the parameters are: first collect the L channel values ​​of all pixels in the area and find the maximum value, which is defined as For example, when the upper limit of the L channel is detected to be 88.7 in a group of areas, =88.7.

[0055] The steps to obtain the parameters are as follows: count the number of pixels in the region whose a channel value is lower than the median of the a channel of the entire image and mark them as , you need to calculate the median of the a channel in the entire image range, you can sort all the pixels according to their a values ​​and take the midpoint of the sort, and then get the median of the whole image, and then The a value is compared pixel by pixel in each region, and the counts of all pixels below the median are increased by 1 and finally accumulated to For example, the median of the a channel in the entire image is 12. There are 36,000 pixels in the area, of which 18,000 pixels have a value less than 12. =18000.

[0056] The steps to obtain the parameters are: collect the minimum value of the B channel value in the area and use it as The specific operation is similar to that of the G channel or R channel. When traversing pixel by pixel, find the minimum item of all B values, record it and assign it to For example, if the minimum value of B is 5 in a certain test, then =5.

[0057] The steps to obtain the parameters are: calculate the extreme difference of the saturation of all pixels in the area and assign it to , you can first obtain the saturation S value of each pixel under the HSL or HSV color model, and then calculate the difference between the maximum S value and the minimum S value in the area, and define the difference as the saturation range. For example, in the range of 0 to 1, the maximum S value is 0.9 and the minimum S value is 0.2, then the range is 0.7.

[0058] The steps to obtain the parameters are: measure the pixel span of the area in the vertical direction of the image, compare the top and bottom pixel coordinates and take the difference as ,and The horizontal scanning process is the same as that of the image. First, scan the y coordinates of the pixels in this area and lock the minimum and maximum values. Subtract the two to get the vertical pixel span. For example, at a resolution of 1080, the minimum y=200 and the maximum y=520, then =320.

[0059] Calculation process: The first step is to calculate the parameters after the acquisition is completed. All are sorted out and given specific values, let , , , , , , , , , ; The second step is to calculate the left part of the molecule , insert the value into ; The third step is to calculate the right side of the numerator , insert the value into ; Step 4: Do the difference operation , and then take the absolute value is still 1414.53; Step 5: Take the square root , thus we get The results show that when When the value reaches 37.61, it indicates that there is a significant difference between the upward brightness component formed by the combination of the maximum value of the R channel and the minimum value of the G channel in this area and the indicators of the L channel and B channel. Overall, the degree of color anomaly is quite obvious. It can be compared with the subsequent intensity threshold. Once a certain threshold is exceeded, this area is determined to be a color anomaly area, and then enters the subsequent processing link and records the relevant coordinates.

[0060] According to the color abnormality intensity value distribution list obtained above, all areas and their corresponding Then, we need to select an intensity threshold to distinguish normal and abnormal areas. We can use the experience of testing multiple batches of MicroLED display images to obtain a relatively stable range. For example, we can perform the same sampling test on 500 previous images and count all the abnormal areas. The intensity falls between 5 and 60, and the corresponding area is manually confirmed to see if there is a visual color cast problem. It is found that when the threshold is around 30, it meets the need for accurate identification of abnormal areas. Therefore, 30 can be set as the intensity threshold in this test, and then the intensity of all areas is measured. When a certain area is traversed, If it is greater than 30, the coordinate range and the specific extreme values ​​in the R, G, and B channels and the corresponding number of pixels are recorded. During the recording process, the starting coordinate and the ending coordinate of the area are first merged into a two-tuple, and then the above channel attributes and brightness statistics are written into the same entry. If there are multiple brightness segments or channel extreme difference segments in the area that meet similar abnormal performance, the corresponding tags can be added multiple times. For example, if a block with a horizontal coordinate of 100 to 200 and a vertical coordinate of 50 to 150 appears If the value of the R channel is greater than 30 and the maximum value of the R channel exceeds the average channel value range by more than 30, the color abnormality degree of the block is additionally described at this entry. After traversing all areas, a summary table of color abnormality entries is obtained. Finally, these entries are sequentially integrated to form a unified coordinate index list. Channel and brightness information is attached to each list entry. When cross-validation is required, multiple color features under the same coordinate can be compared again. After all the marked data are concentrated, a record of color difference abnormality is obtained.

[0061] The steps for obtaining the adjusted abnormal area image are: According to the abnormal color difference record, the camera exposure adjustment factor is calculated. The calculation formula is: ; in, Indicates the The exposure adjustment factor for abnormal areas, Indicates the grayscale dynamic range of the image in this area, Indicates the lowest pixel brightness value in the area. represents the brightness variance within the image frame of the region, Indicates the ambient light intensity corresponding to the abnormal frame, Indicates the average brightness of the channels in the image frame of this area, Indicates the ratio of edge pixels in the color difference channel of the area, Indicates the number of overexposed pixels in the image frame of this area; According to the camera exposure adjustment factor, the camera exposure time reset rule is set, the exposure value is set to the time inversely calculated by the product of the current exposure time and the camera exposure adjustment factor, and the image frame corresponding to the abnormal area is recaptured to obtain the adjusted abnormal area image.

[0062] Specifically, the formula: The benefit of the formula is that it introduces multiple quantitative indicators such as grayscale dynamic range, minimum pixel brightness value, brightness variance, ambient light intensity, channel average brightness, color difference channel edge pixel ratio and number of overexposed pixels, and combines these physical or statistical quantities related to the imaging effect of abnormal areas for calculation. Through two-stage multiplication and absolute value operation, the influence of regional lighting anomalies and edge overexposure on the final exposure adjustment factor is highlighted, thereby forming a calculable correction to the camera exposure time.

[0063] The parameters are obtained as follows: the grayscale dynamic range of the image in this area Refers to the difference between the grayscale values ​​of all pixels in the area (usually 0 to 255 or the range of 0 to 255 obtained after 10-bit equal depth mapping) from the minimum to the maximum. When performing statistics, you can first record the grayscale of each pixel in the area, sort these grayscale values ​​and take the difference between the maximum and minimum values, and set it as If the area contains many pixels, the efficiency can be improved by using segmented statistics. For example, when a 320×240 abnormal area has 76,800 pixels, after scanning, the minimum grayscale value is 3 and the maximum grayscale value is 180, then .

[0064] The steps to obtain the parameters are: the lowest pixel brightness value in the area Refers to the minimum brightness obtained after traversing the brightness values ​​of all pixels in the area. In order to obtain a consistent quantization standard, RGB needs to be converted to brightness or grayscale respectively. You can use Get the brightness of each pixel, traverse this area and record the minimum brightness, and then mark the value as If an area has 50,000 pixels and the minimum brightness in the statistical result is 15.6, then =15.6.

[0065] The steps to obtain the parameters are: the brightness variance of the image frame in the area It is used to measure the discrete degree of pixel brightness distribution in this area. The method of obtaining it is to first calculate the average brightness of all pixels in the area, set as , and then the brightness of each pixel and Do Operation, these values ​​are accumulated and divided by the number of pixels to get the variance, which is named , if the total number of pixels is , then the variance can be written as , and the final result is assigned to , for example, when the statistics Time, can be remembered The larger the value, the more uneven the pixel brightness distribution is, which will play a role in the numerator of the formula later. Example: The brightness variance of a certain area obtained through actual measurement is 180.5, then .

[0066] The steps to obtain the parameters are: the ambient light intensity corresponding to the area It can be obtained from the illuminance meter reading measured synchronously at the moment of exposure. The illuminance meter is placed on the periphery of the screen to capture the illumination value of the display surface in the current environment. The value is directly read according to the actual measurement, and then the value is matched with the shooting time corresponding to the area. If the illumination recorded at the moment of shooting is 380Lux, then .

[0067] The steps to obtain the parameters are: the average brightness of the channels in the image frame of the area It means that in this area, each pixel RGB is first converted to brightness, and then the brightness of all pixels is averaged, similar to the previous The calculation process does not calculate the degree of dispersion, only the simple average value is concerned. If there are pixels, add the brightness and then divide by That is, the average channel brightness of this area is obtained. For example, when the total number of pixels in the area is 35,000 and the accumulated brightness is 2,250,000, then .

[0068] The parameters are obtained as follows: the ratio of edge pixels in the color difference channel of the area Indicates the percentage of pixels with obvious color difference edges in this area. The acquisition process is to first identify all adjacent pixel boundaries that meet the color difference threshold in the area, such as the neighborhood where the difference between R, G or CIELAB is greater than the specified range, record the number of pixel sets belonging to "color difference edges", divide it by the total number of pixels in the area, and finally multiply it by 100 or convert it into a decimal form to obtain the percentage of edge pixels and assign it to For example, if the total number of pixels in a certain area is 40,000, and 1,200 pixels are detected to trigger edge conditions in the channel difference, then .

[0069] The steps to obtain the parameters are: the number of overexposed pixels in the image frame of the area It means that within this area, all pixels are tested to see if their brightness exceeds a preset limit. If the brightness value is higher than this limit, it is considered an overexposed pixel. After counting the total number of these pixels, the result is assigned to , usually refer to the saturation value of the display device or image sensor to make the setting, for example, 220 is used as the brightness overexposure threshold. When 500 pixels in a certain area exceed this brightness, .

[0070] Calculation process: The first step is to prepare the values ​​of various parameters: , , , , , , ; The second step is to calculate :Calculate first , take the reciprocal of about 0.075, and then , the reciprocal is about 0.089, and the sum of the two is 0.075+0.089=0.164; Step 3: Calculate :Calculate first , , so , then calculate , do subtraction 829.02-0.0000714 ; In the fourth step, the absolute value is still 829.01993, which is multiplied by the result of step 2, 0.164, which is ; Therefore .

[0071] The results show that when the area When reaching 136.018, it indicates that after considering the comprehensive factors such as gray dynamic range, minimum pixel brightness, brightness variance, ambient light intensity, channel average brightness, color difference edge pixel ratio, and overexposed pixel number, the region needs a larger exposure time correction factor. In the subsequent steps, the existing exposure time can be multiplied by this factor or calculated accordingly to set a new exposure time and capture the image again to obtain a more balanced picture.

[0072] Based on the exposure adjustment factor obtained in the previous step, the current exposure time of the abnormal region is recorded during execution, and then the time is multiplied by the corresponding factor to obtain a new exposure time. Different display environments can use milliseconds or smaller units of precision to quantify the time. If the originally set time is within the range of 5 milliseconds to 20 milliseconds, the factors calculated for each abnormal region can be multiplied. If the result exceeds the allowed range, a predefined upper limit needs to be set to limit the adjustment range. For example, it is found in the previously collected samples that a factor greater than 100 will cause the image to be too bright. Therefore, a maximum value of 60 milliseconds can be set as a predetermined standard. When the exposure time corresponding to the factor calculation value is greater than 60 milliseconds, it is kept at 60 milliseconds and not increased further. Then, the image frame of the abnormal region is captured again according to the new exposure time. After shooting, the differences before and after are compared to determine the success of the adjustment by observing whether further high-light or dark-weak regions appear. If a large number of pixel brightness values still exceed the specified threshold of 220, the exposure time is adjusted downward. When the brightness of all abnormal regions is gradually balanced to the interval of 120 to 200, it is considered that the adjustment has reached the desired level. If necessary, these newly captured images can be analyzed again and compared with the previously saved ambient light intensity and brightness variance data. If the new image still has obvious brightness peaks, the camera parameters are adjusted and the process is repeated. Through multiple cycles, the final exposure parameters of the target abnormal region are obtained, and the adjusted abnormal region image is generated.

[0073] The steps for obtaining the enhanced detail image are as follows: Import the adjusted abnormal region image, unify the image data format, adjust the image size to the input resolution specified by the ESRGAN model according to the requirements of the ESRGAN model, and obtain the preprocessed image data. Load the ESRGAN model, input the preprocessed image data, perform feature extraction and enlargement, and generate an enhanced detail image.

[0074] Specifically, based on the adjusted abnormal area image that has been obtained, the pixel format of the image file is first unified, and a three-channel uncompressed RGB data structure is uniformly used. During this period, it is necessary to detect whether there are materials with different bit depths or color spaces that do not meet the requirements. If 8-bit and 10-bit are mixed, alignment is performed based on the predetermined benchmark format, and the pixel value range is scaled or quantized. The grayscale, brightness, and color components in all images are mapped to the range of 0 to 255. After the format is unified, the resolution and aspect ratio of each image are checked. If resolution differences are found, the ESRGAN model is used to adjust the resolution. The input resolution required by the model is adjusted. In the example, a fixed size of 512 in width and 512 in height can be set. The specific selection of this value is based on the previous statistical findings that the size of 512×512 can be smoothly operated on a graphics processing device with 8GB of video memory in actual operation and maintain an acceptable frame processing rate. If the original resolution of some images is very small, such as 100×100, they need to be expanded to the specified size through interpolation and then treated the same as other images. If the original resolution is much larger than 512×512, a segmented cropping strategy can be used, that is, first divide the oversized image into several blocks, and scale each block to 512×512 to ensure The subsequent model input will not be too large. If the remaining edge area is insufficient for the cropping size during the segmentation process, the edge area will be merged with the adjacent part and the pixels will be counted again to ensure that the divided images can be aligned to the predetermined size. During this period, the cropping coordinates corresponding to each image should also be recorded for subsequent restoration or comparison. If invalid pixels or missing pixels are detected at the edge of the image, they will be eliminated line by line or filled with surrounding pixel values. This threshold can be set based on the brightness distribution obtained previously. For example, no more than 200 missing or invalid pixels are allowed. If the preset threshold is exceeded, the image will be marked as unsuitable for further recognition. Attention should also be paid to the image. The color component integrity is checked. If the channel value of some pixels is abnormal, the pixel is recorded in the middle as a label to be observed. After all the above processing is completed, the same image is converted into a standard resolution and channel format data structure that meets the requirements of the ESRGAN model in memory. During the periodic summary, multiple comparisons can be performed based on the pre-established valid range, such as resolution comparison of 512×512 and channel value comparison of 0 to 255. If multiple errors exist, the original image is proofread again or the frame is discarded. Finally, all images that have passed the inspection and successfully converted to the specified input resolution are imported into the same task list for subsequent model calls to obtain the preprocessed image data.

[0075] Based on the image data that has completed format unification and size adjustment, first, the specific training process and inference use process of the ESRGAN model are determined, and whether incremental training is needed is judged in combination with the sample size during training. The ESRGAN model usually includes a generator and a discriminator of a generative adversarial network, and a batch of high-resolution and low-resolution image pairs need to be prepared for contrast training. If the generator and discriminator weights trained in advance are available, the corresponding weight files can be directly used and the fine-tuning process is executed. During training, the batch size can be set to between 16 and 32 to maintain stable operation on common graphics processing hardware. Each batch reads low-resolution images and corresponding high-resolution image labels for comparison. Through multiple ways such as adversarial loss, content loss, and perceptual loss, the generator weight is optimized so that the generator presents more real details and reduces noise in the enlarged image. The discriminator constantly tries to identify whether the enlarged image has texture features consistent with real images. After training is completed, a mature ESRGAN model parameter file is obtained. Then, in the inference use link, the abnormal area images that have been preprocessed and meet the 512x512 size requirement are sent one by one into the generator network. The generator extracts and enlarges the feature map through convolution and deconvolution operations, activates the convolution output of each layer, and stacks multiple skip connection structures to obtain more details in the deep layer. During enlargement, the image resolution is automatically interpolated and adjusted. Common enlargement ratios can be set to 2 times or 4 times. If 4 times enlargement is used, the output can reach a resolution of 2048x2048 when the input is 512x512. During this process, the output of each convolution layer is recorded in the local memory to analyze the feature extraction status of the image at different scales. If there are large-area edge artifacts after enlargement, retraining can be performed by changing the loss balance coefficient during training or appropriately adjusting the generator convolution layer depth. The specific values of the weight parameters are different for different sample set features. For example, if the sample set contains many noisy dark areas, the weight proportion of perceptual loss needs to be increased during training. In this scenario, the weight of perceptual loss can be set to 0.4, the weight of adversarial loss can be set to 0.3, and the remaining part of the content loss is 0.3. The convergence of the actual example is calculated to determine the reasonableness. When the output of the model reaches a balance between clarity and false edge phenomenon, the model can be frozen and put into use. The final ESRGAN model is applied to all abnormal area image enlargement steps, and the enlarged pictures are subjected to pixel inspection and the final generated high-resolution image sequence is recorded to generate enhanced detail images.

[0076] The defect recognition result acquisition step is: The convolutional neural network model is loaded, and the enhanced detail image is input into the convolutional neural network model. The multi-layer feature recognition capability of the convolutional neural network model is used for analysis, and image features are extracted layer by layer to identify potential defect areas to obtain defect recognition results.

[0077] Specifically, based on the enhanced detail image obtained previously, each layer of the convolutional neural network model consists of several convolution kernels, activation functions, and pooling operations. If different types of image defects need to be distinguished, a fully connected layer can be added at the end of the network for classification output. The training data of the model can come from the image samples with defect labels and non-defect labels collected in the early stage. These samples are shuffled and divided into training set and validation set. The loss function is set to cross entropy or mean square error, and the selection is made according to the difference between classification tasks and detection tasks. The batch size is set to 32 or 64 during training. Data enhancement operations such as random flipping or random cropping can be performed on the input image to avoid the network only remembering specific fixed sample patterns. An optimization algorithm is used when updating the weights, and the training error and validation error are calculated after each round of iteration. The ratio of correctly identified defects in each round is recorded. In this stage, if it is found that the recognition rate of a certain type of defect is significantly lower than that of other categories, more similar samples can be collected or synthesized for enhanced training, or the network depth can be further increased, such as expanding two to three layers of convolution and adjusting the feature map size. Adjustments are made until the validation error converges and the recognition rate reaches a pre-set range of 80% to 95%. The trained model incorporates multi-layer feature extraction capabilities, extracting local texture and edge distribution layer by layer when a new image to be inspected is input. Finally, the output layer determines the presence or location of a defect. If pixel-level defect localization is required, the network architecture can be designed as a fully convolutional model and output a probability heatmap with the same resolution as the input. Alternatively, existing object detection architectures can be used for coordinate calibration. A confidence threshold can also be set during training to distinguish true defects from detection noise. For example, a threshold of 0.5 can be set and the optimal solution obtained through practical calculations. After network deployment, the previously generated enhanced detail image is input into the convolutional neural network, and features such as suspicious textures, fractures, or brightness anomalies within the image are analyzed frame by frame. The network scores at the final classification or detection layer are recorded, and pixels above the threshold are marked as defective areas. After all batch inferences are completed, the coordinates of the identified feature points or boundary distribution are derived from the network output to obtain the defect recognition results.

[0078] The above content details a visual MicroLED uniformity inspection method for improving product yield. Accordingly, this specification also provides a visual MicroLED uniformity inspection system for improving product yield. The system includes a processor and a memory, the memory being configured to store a computer program. The processor is configured to call and execute the computer program stored in the memory, causing the visual MicroLED uniformity inspection system to perform the visual MicroLED uniformity inspection method.

[0079] The above are merely preferred embodiments of the present invention and do not limit the present invention in any other form. Any technician familiar with the profession may use the technical content disclosed above to change or modify it into an equivalent embodiment with equivalent changes and apply it to other fields. However, any simple modification, equivalent change and modification made to the above embodiment based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall still fall within the scope of protection of the technical solution of the present invention.

Claims

1. A visual MicroLED uniformity detection method for improving yield, characterized in that: The following steps are involved: Obtain the original image data of the MicroLED screen through the camera to obtain the original image data; Based on the grayscale value and coordinate information of the original image data, the image is divided into regions and then a watershed region growing strategy is used to generate segmented region data; color analysis is performed on each region in the segmented region data, and the color deviation of each region is calculated to obtain a regional color analysis result; Based on the regional color analysis result, identifying the region with abnormal color difference, recording the location and color characteristics of the region with abnormal color difference, and generating a color difference abnormality record; determining whether the region with abnormal color difference is caused by the ambient lighting conditions, and if so, adjusting the camera exposure time, capturing an image of the abnormal region again, and obtaining an adjusted image of the abnormal region; Applying an ESRGAN super-resolution model to the adjusted abnormal area image to perform resolution enhancement to obtain an enhanced detail image; and performing defect recognition on the enhanced detail image to generate a defect recognition result.

2. The visual MicroLED uniformity detection method for improving yield rate according to claim 1, characterized in that: The steps for obtaining the original image data are: Configure the camera to face the MicroLED screen and adjust the camera parameters, including focal length and exposure time, to obtain the adjusted camera settings. Based on the adjusted camera settings, the camera captures light emitted by the MicroLED screen and converts the light into a digital image file to obtain unprocessed raw image data; Based on the unprocessed original image data, the integrity of the image and the quality of the captured image are checked to determine whether the data file is damaged, thereby obtaining the original image data.

3. The visual MicroLED uniformity detection method for improving yield rate according to claim 1, characterized in that: The steps for obtaining the segmented region data are as follows: Extracting the grayscale value and coordinate information of each pixel from the original image data, establishing an initial connection graph between all pixels based on the eight-neighborhood rule, marking the connection edges whose grayscale value differences are within a preset threshold, and obtaining an initial pixel connectivity graph structure; According to the initial pixel connectivity graph structure, local consistency scores of several candidate sub-image regions are obtained. The calculation formula of the local consistency scores of the candidate sub-image regions is: ； in, For the The local consistency score of the candidate sub-image region, is the set of all pixel pairs with connected edges in the sub-graph area, and Pixels and pixels The gray value of is the number of connected edges in the region; According to the local consistency score of each candidate region, the pixel sub-image with a local consistency score higher than the distribution median is marked as an initial region mask, and the watershed region growing strategy is performed on the initial region mask and the boundary is calibrated to generate the segmented region data.

4. The visual MicroLED uniformity detection method for improving product yield according to claim 1, wherein: The steps for obtaining the regional color analysis results are: Traversing each region in the segmented regional data in turn, extracting the RGB three-channel values ​​of all pixels from each region and converting them into CIELAB space color values, and counting the value sets of the L channel, a channel, and b channel corresponding to each pixel in the region to obtain a regional color component dataset; Based on the regional color component data set, the color deviation index of each region is calculated. The calculation formula of the color deviation index is: ； in, Indicates the Color deviation index of each area, and Respectively represent the maximum and minimum values ​​of the L channel in the area, Indicates the total number of pixels in the area. and Represent the mode values ​​of channel a and channel b in the area, represents the average value of the L channel, represents the variance of the L channel, represents the skewness coefficient of channel a, represents the average value of b channel; According to the color deviation index, all regions are numerically sorted according to the color deviation index and regions whose color deviation index is greater than the distribution mean are extracted as regional color analysis results.

5. The visual MicroLED uniformity detection method for improving yield rate according to claim 1, characterized in that: The steps for obtaining the color difference abnormality record are: Based on the regional color analysis results, the color abnormality intensity value of each region is calculated using the following formula: ； in, Indicates the The color abnormality intensity value of the area, Indicates the maximum value of the pixel of the R channel in the area, Indicates the number of pixels in the area that have a brightness higher than the average brightness of the entire image. Indicates the minimum value of the G channel pixel in the area, Indicates the number of color change points of all pixels in the area, represents the horizontal spatial span of pixels in the region, Indicates the maximum pixel value of the L channel in the area, Indicates the number of pixels in the area that are lower than the median of the a channel of the entire image. Indicates the minimum value of the B channel pixel in the area, Indicates the extreme saturation value of all pixels in the area. Indicates the vertical spatial span of pixels in the region; According to the color abnormality intensity value of each area, the intensity threshold is set to extract the abnormal area, and the abnormal area coordinate value and channel attribute combination are marked into the abnormal record file structure to generate the color difference abnormality record.

6. The visual MicroLED uniformity detection method for improving product yield according to claim 1, wherein: The steps of acquiring the adjusted abnormal area image are as follows: According to the color difference abnormality record, the camera exposure adjustment factor is calculated, and the calculation formula is: ； in, Indicates the The exposure adjustment factor for abnormal areas, Indicates the grayscale dynamic range of the image in this area, Indicates the lowest pixel brightness value in the area. represents the brightness variance within the image frame of the region, Indicates the ambient light intensity corresponding to the abnormal frame, Indicates the average brightness of the channels in the image frame of this area, Indicates the ratio of edge pixels in the color difference channel of the area, Indicates the number of overexposed pixels in the image frame of this area; According to the camera exposure adjustment factor, a camera exposure time reset rule is set, the exposure value is set to the time inversely calculated by the product of the current exposure time and the camera exposure adjustment factor, and the image frame corresponding to the abnormal area is recaptured to obtain an adjusted abnormal area image.

7. The visual MicroLED uniformity detection method for improving product yield according to claim 1, wherein: The steps of acquiring the enhanced detail image are: Importing the adjusted abnormal area image, unifying the image data format, and adjusting the image size to the input resolution specified by the ESRGAN model according to the requirements of the ESRGAN model to obtain preprocessed image data; The ESRGAN model is loaded, and the preprocessed image data is used as input to perform feature extraction and amplification to generate an enhanced detail image.

8. The visual MicroLED uniformity detection method for improving product yield according to claim 1, wherein: The steps for obtaining the defect identification result are: The convolutional neural network model is loaded, the enhanced detail image is input into the convolutional neural network model, and the multi-layer feature recognition capability of the convolutional neural network model is used for analysis, image features are extracted layer by layer, and potential defect areas are identified to obtain defect recognition results.

9. A visual MicroLED uniformity inspection system for improving yield, characterized in that: include: A processor and a memory, the memory being used to store a computer program, the processor being used to call and run the computer program stored in the memory, so that the visual MicroLED uniformity detection system performs the method according to any one of claims 1 to 8.

Images