Machine learning-based mura feature extraction and recognition method

Abstract

The present invention relates to a machine learning-based mura feature extraction and recognition method, comprising the following steps: inputting a liquid crystal panel image, and carrying out adaptive histogram equalization on the image; carrying out high-dimensional feature extraction on each pixel of a full-field image; and entering a machine learning link, eliminating an outlier, implementing a clustering algorithm in the remaining samples, entering an optimal binarization segmentation link, and outputting a foreground point mask image. According to the machine learning-based mura feature extraction and recognition method of the present invention, by carrying out statistical analysis on the high-dimensional features of an image, a key position different from the common state of a liquid crystal panel is automatically recognized, and the key position is used as a suggestion for manual mura determination and is displayed on a software interface, so that mura defects of different types, different forms and different image features can be automatically recognized, and therefore, the operation is simple and convenient, the present invention has no high requirement for the hardware capability of a calculation unit, and the calculation unit can be quickly deployed at a plurality of stations.

Description

A mura feature extraction and identification method based on machine learning Technical Field

[0001] The present invention relates to the field of liquid crystal panel production, and in particular to a mura feature extraction and identification method based on machine learning. Background Art

[0002] In the LCD panel production process, identifying and locating mura defects is an essential and critical step in the entire defect detection process. Mura defects refer to uneven brightness on displays, resulting in various artifacts. The location accuracy and classification accuracy of mura defects are directly related to factors such as product yield, customer complaint compensation, and labor costs. Therefore, this process is extremely important.

[0003] Existing mura defect recognition technology can be divided into three major areas. The first relies on human recognition. A dedicated operator is placed in front of the camera screen. Each time the type and location of mura defects need to be identified, the operator uses software to stretch the image contrast, continuously adjusting the contrast until the mura defects are clearly visible on the screen. The second is an algorithm based on traditional template matching. This solution rotates and translates the LCD panel images actually captured on the production line against a pre-defined mura defect map, ultimately outputting the defect type and location. The third is an automatic recognition solution based on deep learning. Before model training begins, a large number of mura defect images are collected and manually annotated as a training set. After training, the model is able to automatically identify mura defects.

[0004] The disadvantage of manual inspection is that the identification of mura defects is significantly affected by human factors. Different inspectors often disagree on the location and presence of mura defects in the same LCD panel image. Current on-site manual inspections typically require an experienced inspector to lead the inspection team and conduct spot checks and review of manual inspection results, resulting in significant labor costs. Furthermore, errors can occur due to factors such as fatigue and poor positioning. For defects with inconsistent identification, yield issues arise, such as over-detection and missed detection.

[0005] The disadvantage of traditional visual methods is that sample images need to be collected for each form of mura defect. This process is relatively cumbersome. When launching a product, image algorithm personnel are required to perform multiple parameter adjustments based on the product characteristics, resulting in low efficiency in launching new products.

[0006] The disadvantages of deep learning methods include long model training cycles, large image collection volumes, and high requirements for computing units. For mura defects not found in the training set, multiple rounds of labeling and iterative training are required, further reducing deployment efficiency.

[0007] Summary of the Invention

[0008] To address the technical problems existing in the prior art, the present invention provides a machine learning-based mura feature extraction and identification method. This method statistically analyzes the high-dimensional features of an image to automatically identify key locations that differ from the normal state of the LCD panel. These key locations are then displayed on the software interface as suggestions for manual mura determination.

[0009] To achieve the above object, the technical solutions of the present invention are as follows:

[0010] A mura feature extraction and identification method based on machine learning includes the following steps:

[0011] Step 1: input the LCD panel image and perform adaptive histogram equalization on the image;

[0012] Step 2: extract high-dimensional features from each pixel of the full-field image;

[0013] Step three: Enter the machine learning phase, remove outliers, implement the clustering algorithm on the remaining samples, enter the optimal binary segmentation phase, and output the foreground point mask image.

[0014] As a preferred technical solution, in step 1, the adaptive histogram equalization calculation method is to count the histogram distribution of each grayscale value above 0-255, take the top 85% of the highest distribution as the upper and lower bounds of the histogram, and linearly pull the histogram to 0-255 with the upper and lower bounds, so that the mura defects are clearly displayed in the image for subsequent calculation and processing.

[0015] As a preferred technical solution, in step 2, the high-dimensional features include grayscale value, relative center coordinates, rectangular box variance, rectangular box maximum and minimum values, horizontal / vertical variance, horizontal / vertical maximum and minimum values, and envelope similarity, and each high-dimensional feature is assigned a corresponding weight.

[0016] As a preferred technical solution, the feature calculation method is: gray value I (x, y); relative center coordinates, that is, the normalized coordinates x of the point coordinate to the center of the liquid crystal panel rel =(xx center ) / wid,y rel =(yy center ) / het; variance of grayscale values ​​within the rectangular area The maximum and minimum values ​​of the grayscale value in the rectangular area min / max{I(x, y)|(x, y)∈Box}; the variance of the grayscale value in the horizontal and vertical directions The maximum and minimum values ​​of the grayscale values ​​in the horizontal and vertical directions min / max{I(x,y)|(x,y)∈col / row}; the envelope similarity is the sum of the absolute values ​​of the pixel-by-pixel differences between the grayscale values ​​in the rectangular area and the adjacent areas

[0017] As a preferred technical solution, in step three, the steps for removing outliers are as follows: calculating the geometric center of all samples in the high-dimensional feature space, and calculating the Euclidean distance of each sample to the geometric center; arranging the Euclidean distances from large to small, selecting and removing the largest 3%-5% of the data as outlier data; continuing to calculate the geometric center of the remaining samples after removal, and calculating the Euclidean distance of each sample of the remaining samples to the geometric center, and continuing to remove 3%-5% of the outlier data; repeating this step until about 70% of the data is retained before entering the next stage.

[0018] As a preferred technical solution, in step three, the clustering algorithm calculates the position of the point set with higher density in the space as the anchor point. According to the pre-set density threshold, the point above the density threshold is the anchor point, and several anchor points with similar distances are combined into the same anchor point. Finally, several anchor points are found in the entire high-dimensional sample space; with the anchor point as the core, the sample is segmented in the high-dimensional space; the distance from each individual sample to all the anchor points is calculated, and the point with the closest distance is assigned to it; the segmentation is repeated until the segmentation reaches a stable state.

[0019] As a preferred technical solution, in step three, the optimal binary segmentation link is to calculate the pixel grayscale mean for each clustered set, sort the pixel grayscale mean from high to low according to the average grayscale, and count the cumulative frequency. The point with the highest grayscale value and a cumulative frequency greater than 30% is taken as the foreground point of the mura defect.

[0020] Compared with the prior art, the present invention has the following beneficial effects:

[0021] (1) The machine learning-based mura feature extraction and identification method of the present invention automatically identifies key locations that are different from the normal state of the LCD panel by statistically analyzing the high-dimensional features of the image, and displays these key locations on the software interface as suggestions for manual mura determination.

[0022] (2) The main features of the mura feature extraction and identification method based on machine learning of the present invention are: it can automatically identify mura defects of different types, different forms, and different image features; it is simple to operate, does not require high hardware capabilities of the computing unit, and can be quickly deployed at multiple sites. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] FIG1 is a flow chart of a mura feature extraction and identification method based on machine learning of the present invention;

[0024] FIG2 is a comparison diagram before and after adaptive histogram equalization processing in a mura feature extraction and identification method based on machine learning of the present invention;

[0025] FIG3 is a mura detection image output by a mura feature extraction and recognition method based on machine learning of the present invention. DETAILED DESCRIPTION

[0026] The technical solution of the present invention is further described below in conjunction with specific embodiments:

[0027] As shown in Figure 1, a mura feature extraction and identification method based on machine learning includes the following steps:

[0028] Step 1: Input an LCD panel image and perform adaptive histogram equalization on the image. As shown in Figure 2, the adaptive histogram equalization calculation method is to count the histogram distribution of each grayscale value above 0-255, take the top 85% of the distribution as the upper and lower bounds of the histogram, and linearly pull the histogram to 0-255 using the upper and lower bounds, so that mura defects are clearly displayed in the image for subsequent processing.

[0029] Step 2: Extract high-dimensional features from each pixel of the full-field image. High-dimensional features include grayscale value, relative center coordinates, rectangular box variance, rectangular box maximum and minimum values, horizontal / vertical variance, horizontal / vertical maximum and minimum values, and envelope similarity. Each high-dimensional feature is assigned a corresponding weight. The feature calculation method is as follows:

[0030] Gray value: I (x, y); relative center coordinates, that is, the normalized coordinates of the point coordinates to the center of the LCD panel: x rel =(xx center ) / wid,y rel =(yy center ) / het; Variance of grayscale values ​​within the rectangular area: You can select multiple rectangular areas of different sizes to expand the feature dimension. The maximum and minimum values ​​of the grayscale value in the rectangular area: min / max{I(x,y)|(x,y)∈col / row}, you can select multiple rectangular areas of different sizes to expand the feature dimension. The variance of the grayscale value in the horizontal and vertical directions:

[0031] Multiple horizontal and vertical regions of different ranges can be selected to expand the feature dimension. The maximum and minimum values ​​of the grayscale values ​​in the horizontal and vertical directions are: min / max{I(x,y)|(x,y)∈col / row}. Multiple horizontal and vertical regions of different ranges can be selected to expand the feature dimension. Envelope similarity, that is, the sum of the absolute values ​​of the pixel-by-pixel differences between the grayscale values ​​in the rectangular region and the adjacent regions: Multiple rectangular regions of different sizes and multiple adjacent regions can be selected to expand the feature dimension.

[0032] Step three, enter the machine learning stage. The first step is to eliminate outliers. Calculate the geometric center of all samples in the high-dimensional feature space, and calculate the Euclidean distance of each sample to the geometric center. Arrange the Euclidean distances from large to small, and select and eliminate the largest 3%-5% of the data as outlier data. After elimination, continue to calculate the geometric center of the remaining samples, and calculate the Euclidean distance of each sample to the geometric center, and continue to eliminate 3%-5% of the outlier data. Repeat this step until about 70% of the data is retained before entering the next stage.

[0033] A clustering algorithm is applied to the remaining samples. Essentially, the clustering algorithm calculates the locations of densely populated point clusters in space as anchor points. Points above a pre-set density threshold are designated as anchor points. Several anchor points with similar distances are combined into a single anchor point, ultimately finding a number of anchor points in the entire high-dimensional sample space. Using these anchor points as the core, the samples are segmented in the high-dimensional space. The segmentation algorithm calculates the distance from each individual sample to all anchor points and assigns the closest point to it. For example, considering a sample distribution with N anchor points, the total distance from a given sample to each of the N anchor points is calculated, recorded as D(n). The anchor point D(m) with the minimum value in D(n) is assigned to the mth anchor point. After the initial anchor point segmentation, N groups of points are generated, centered around the anchor point. For each point set, the geometric center of all samples is calculated and used as the new anchor point for that point set. Based on these N new anchor points, the next anchor point segmentation is performed on all point samples. Repeat the segmentation pattern described above for multiple iterations until the difference between the anchor point position generated in the kth iteration and the anchor point position generated in the k-1th iteration is less than the set threshold. This means that the segmentation has reached a stable state. This completes the clustering analysis task.

[0034] Finally, the optimal binarization segmentation phase begins. For each clustered set, the mean grayscale value of each pixel is calculated, sorted from high to low by average grayscale value, and the cumulative frequency is calculated. Points with the highest grayscale value and a cumulative frequency greater than 30% are selected as foreground mura defects.

[0035] As shown in Figure 3, the foreground point mask image is finally output to the software front end, and the algorithm task is completed in one go.

[0036] This embodiment is only a further explanation of the present invention and is not a limitation of the present invention. After reading this specification, those skilled in the art may make non-creative modifications to this embodiment as needed, but as long as they are within the scope of the claims of the present invention, they are protected by patent law.

Claims

1. A mura feature extraction and identification method based on machine learning, characterized in that: The method comprises the following steps: Step 1, inputting a liquid crystal panel image, and performing adaptive histogram equalization on the image; Step 2: extract high-dimensional features from each pixel of the full-field image; Step three, enter the machine learning phase, remove outliers, implement the clustering algorithm on the remaining samples, enter the optimal binary segmentation phase, and output the foreground point mask image.

2. The mura feature extraction and identification method based on machine learning according to claim 1, characterized in that: In the step 1, the calculation method of the adaptive histogram equalization is to count the histogram distribution of each gray value above 0-255, take the top 85% of the highest distribution as the upper and lower bounds of the histogram, and linearly pull the histogram to 0-255 with the upper and lower bounds, so that the mura defects are clearly displayed in the image for subsequent calculation and processing.

3. The mura feature extraction and identification method based on machine learning according to claim 1, characterized in that: In the step 2, the high-dimensional features include grayscale value, relative center coordinates, rectangular box variance, rectangular box maximum and minimum values, horizontal / vertical variance, horizontal / vertical maximum and minimum values, and envelope similarity, and each high-dimensional feature is assigned a corresponding weight.

4. The mura feature extraction and identification method based on machine learning according to claim 3 is characterized in that: The characteristic calculation method is: gray value I (x, y); relative center coordinates, that is, the normalized coordinates x of the point coordinates to the center of the LCD panel rel =(xx center ) / wid,y rel =(yy center ) / het; the variance of the grayscale values ​​in the rectangular area dev[{I(x, y)|(x, y)∈Box}]; the maximum and minimum values ​​of the grayscale values ​​in the rectangular area min / max{I(x, y)|(x, y)∈col / row}; the variance of the grayscale values ​​in the horizontal and vertical directions dev[{I(x, y)|(x, y)∈col / row}]; the maximum and minimum values ​​of the grayscale values ​​in the horizontal and vertical directions min / max{I(x, y)|(x, y)∈col / row}; the envelope similarity is the sum of the absolute values ​​of the pixel-by-pixel differences between the grayscale values ​​in the rectangular area and the adjacent areas 5. The mura feature extraction and identification method based on machine learning according to claim 1, characterized in that: In step three, the steps of removing outliers are as follows: calculating the geometric center of all samples in the high-dimensional feature space, and calculating the Euclidean distance from each sample to the geometric center; arranging the Euclidean distances from large to small, selecting and removing the largest 3%-5% of the data as outlier data; continuing to calculate the geometric center of the remaining samples after removal, and calculating the Euclidean distance from each sample of the remaining samples to the geometric center, and continuing to remove 3%-5% of the outlier data; repeating this step until about 70% of the data is retained before entering the next stage.

6. The mura feature extraction and identification method based on machine learning according to claim 1, characterized in that: In the step 3, the clustering algorithm calculates the position of the point set with higher density in the space as the anchor point. According to the pre-set density threshold, the point above the density threshold is the anchor point, and several anchor points with similar distances are combined into the same anchor point. Finally, several anchor points are found in the entire high-dimensional sample space; with the anchor point as the core, the sample is segmented in the high-dimensional space; Calculate the distance between each individual sample and all anchor points, and assign the point with the closest distance to it as its affiliation; The segmentation is repeated until the segmentation reaches a stable state.

7. The mura feature extraction and identification method based on machine learning according to claim 1, characterized in that: In step three, the optimal binary segmentation link is to calculate the pixel grayscale mean for each clustered set, sort the pixel grayscale mean from high to low according to the average grayscale and count the cumulative frequency, and take the point with the highest grayscale value and cumulative frequency greater than 30% as the mura defect foreground point.

Images