Backlight module light leakage measuring method and system based on machine learning

Abstract

The invention relates to the technical field of liquid crystal display backlight module detection, and discloses a backlight module light leakage measuring method and system based on machine learning. The backlight module light leakage determination method based on machine learning comprises the following steps: collecting optical characteristics of a backlight module to construct a matrix; suspicious light leakage areas are segmented and screened according to an environment light self-adaptive threshold value; local contrast enhancement is carried out on the weak light leakage signals, and multi-scale pyramid analysis is combined to enhance the weak light leakage signals; quantifying the complexity of the micro-area light leakage structure by means of box dimension calculation and multi-fractal spectrum analysis, and constructing a fractal feature vector; a correlation model is constructed based on machine learning to establish light leakage and defect mapping, and light leakage details and defect types are output through scoring. The fractal geometry theory, the image enhancement technology and the machine learning algorithm are combined, accurate quantification of the light leakage form and automatic identification of the defect type are achieved, the detection accuracy is improved, the false alarm rate is reduced, the omission ratio is reduced, and the detection sensitivity is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of liquid crystal display backlight module detection technology, and more specifically, to a backlight module light leakage measurement method and system based on machine learning. Background Art

[0002] As a key component of LCDs, the quality of the backlight unit (BLU) directly impacts the overall display quality. During BLU manufacturing, light leakage can occur due to material defects and poor assembly. These leaked areas appear as bright spots or lines on a black screen, severely impacting the user experience. Therefore, accurately detecting BLU light leakage is a crucial task in quality inspection during LCD manufacturing.

[0003] Existing backlight module light leakage detection technologies mainly include manual visual inspection and traditional machine vision methods. Manual visual inspection relies on the experience and judgment of the inspector, which not only has low detection efficiency, but also has strong subjectivity, inconsistent judgment standards, and is susceptible to fatigue. Traditional machine vision methods mainly identify light leakage areas through fixed threshold segmentation or simple edge detection algorithms. Although the effect is acceptable in standard laboratory environments, it has limitations in the complex lighting environment of actual production lines. Specifically, the existing technology faces the following major problems: Under complex ambient light conditions, the lighting is uneven, and the reflection and scattering interference are obvious, resulting in a high false alarm rate for the fixed threshold segmentation method; the slight light leakage signal in a small area is weak, and traditional detection algorithms are difficult to effectively extract, resulting in missed detections; the light leakage morphology is complex and changeable, and traditional geometric feature descriptions are insufficient, making it difficult to accurately associate it with the type of material defect; the lack of an adaptive detection parameter adjustment mechanism leads to poor adaptability to products of different specifications; the lack of the application of intelligent algorithms such as machine learning makes it impossible to achieve automated defect recognition and classification.

[0004] Therefore, there is an urgent need for a backlight module light leakage measurement method and system that can combine machine learning technology, have high detection accuracy, and automatically identify defect types to improve the quality control level of liquid crystal displays. Summary of the Invention

[0005] The present invention provides a backlight module light leakage measurement method and system based on machine learning, which solves the technical problems in related technologies such as interference from complex lighting environments, difficulty in identifying weak light leakage, accurate quantification of irregular shapes, and automatic identification of defect types.

[0006] The present invention discloses a backlight module light leakage measurement method based on machine learning, comprising the following steps:

[0007] Collect multi-dimensional optical feature information of the backlight module and construct the original feature matrix;

[0008] Based on the ambient light conditions, the adaptive threshold algorithm is used to perform preliminary segmentation of the backlight module and screen out suspicious light leakage areas;

[0009] Apply local contrast enhancement algorithm to the screened suspicious areas, combined with multi-scale pyramid analysis technology to enhance the characteristics of weak light leakage signals;

[0010] For the enhanced micro-area light leakage morphology, box dimension calculation and multifractal spectrum analysis are applied to quantify the structural complexity of the light leakage area and construct the fractal feature vector.

[0011] Based on the machine learning algorithm, a correlation model between fractal feature vectors and optical properties is constructed, and a mapping relationship between light leakage morphology and material defect type is established. The light leakage location, range, severity and defect type are calculated and output through the light leakage severity score.

[0012] Furthermore, the original feature matrix is ​​constructed as follows:

[0013] F(x,y)=[B(x,y),C(x,y),GLCM(x,y),E(x,y)];

[0014] Among them, B(x, y) represents brightness features, C(x, y) represents color features, GLCM(x, y) represents texture features, and E(x, y) represents edge features.

[0015] Furthermore, the calculation formula of the adaptive threshold algorithm is:

[0016] T(x,y)=μ(x,y)+k·σ(x,y);

[0017] Where (x, y) represents the coordinates, T(x, y) represents the dynamic threshold, μ(x, y) is the mean of the local area, σ(x, y) is the standard deviation of the local area, and k is the threshold coefficient, which is automatically adjusted according to the ambient light intensity.

[0018] Furthermore, the formula for enhancing the selected suspicious areas by the local contrast enhancement algorithm is:

[0019]

[0020] Among them, I′(x, y) is the enhanced image, I(x, y) is the original image, μ local (x,y) is the local mean, σ local (x, y) is the local standard deviation, α contrast is the enhancement coefficient, ε small A small constant to prevent division by zero, M(x, y) is a suspicious area marking map.

[0021] Furthermore, the multi-scale pyramid analysis technology includes constructing a Gaussian pyramid and a Laplacian pyramid of the enhanced image, and generating a final enhanced image by weighted fusion of multi-scale features, wherein the weight coefficients are adaptively adjusted according to the significance of light leakage features at different scales.

[0022] Furthermore, the box dimension calculation is achieved by the following steps:

[0023] Apply grids of different scales to the light leakage area, with the grid size decreasing exponentially;

[0024] Calculate the number of grids required to cover the light leak area at each grid size;

[0025] Draw a double logarithmic coordinate graph, with the horizontal axis being the logarithm of the grid size and the vertical axis being the logarithm of the number of grids;

[0026] The slope was obtained as the fractal dimension by linear regression.

[0027] Furthermore, the association model between fractal features and optical properties is implemented using a random forest classifier structure. The model includes an input layer, a decision tree set, and a voting integration layer, and is capable of distinguishing different types of defects.

[0028] Furthermore, the fractal feature vector is constructed as follows:

[0029] DF=[D,f(α min ), f(α max ), α0, Δα, CN];

[0030] Where D is the box dimension, f(α min ) and f(α max ) are the function values ​​corresponding to the minimum and maximum values ​​of the multifractal spectrum, α0 is the value of α when f(α) takes the maximum value, which represents the dominant scaling characteristics of the light leakage area, Δα is the value range of α, which represents the multifractal intensity of the light leakage area, and CN is the connectivity number, which represents the topological characteristics of the light leakage area.

[0031] Furthermore, the calculation formula for the light leakage severity score is:

[0032] S=w1·I norm +w2·A norm +w3·D norm +w4·L norm ;

[0033] Among them, S represents the light leakage severity score, I norm Represents the normalized light leakage intensity, A norm represents the normalized light leakage area, D norm represents the normalized fractal dimension, Lnorm It represents the normalized position factor, which is the distance between the light leakage area and the center of the screen. w1, w2, w3, and w4 are the first, second, third, and fourth weight coefficients, which are used to adjust the influence of different factors on the final score.

[0034] The present invention provides a backlight module light leakage measurement system based on machine learning, which is used to execute the above-mentioned backlight module light leakage measurement method based on machine learning, including:

[0035] Image acquisition and feature extraction module, used to collect multi-dimensional optical feature information of the backlight module and construct the original feature matrix;

[0036] Adaptive segmentation module, used to perform preliminary segmentation of the backlight module based on ambient light conditions using an adaptive threshold algorithm to screen suspected light leakage areas;

[0037] The contrast enhancement module is used to apply a local contrast enhancement algorithm to the screened suspicious areas, combined with multi-scale pyramid analysis technology to enhance the characteristics of weak light leakage signals;

[0038] The fractal feature analysis module is used to calculate the box dimension and perform multifractal spectrum analysis on the enhanced micro-area light leakage morphology, quantify the structural complexity of the light leakage area, and construct a fractal feature vector.

[0039] The machine learning defect assessment module is used to build an association model between fractal feature vectors and optical properties based on machine learning algorithms, establish a mapping relationship between light leakage morphology and material defect type, and calculate and output the light leakage location, range, severity and defect type through the light leakage severity score.

[0040] Beneficial effects of the present invention:

[0041] The present invention solves the difficult problem of backlight module light leakage detection under complex ambient light conditions by combining multi-dimensional optical feature analysis, dynamic threshold adaptive segmentation, local contrast enhancement, fractal dimension feature quantification and machine learning algorithm, and achieves the following beneficial effects: through multi-dimensional optical feature acquisition and dynamic threshold adaptive segmentation technology, adaptation to complex ambient light conditions is achieved, detection accuracy is improved, false alarm rate is reduced, and the robustness of the detection system is enhanced; the local contrast enhancement algorithm and multi-scale pyramid analysis technology are adopted to effectively enhance weak light leakage signals, improve detection sensitivity, and solve the problem that traditional methods are difficult to identify slight light leakage in small areas; fractal geometry theory is introduced to accurately quantify the light leakage morphology, overcome the limitations of traditional geometric feature description in the analysis of irregular light leakage morphology, and reduce the missed detection rate; the association model of fractal features and optical properties constructed based on the machine learning algorithm realizes the automatic mapping of light leakage morphology and material defect type, providing a direct basis for production process optimization; the entire detection process realizes the seamless integration of large-area detection and micro-area fine analysis, improves detection efficiency, and reduces computing resource requirements while ensuring high-precision detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is a flow chart of a backlight module light leakage measurement method based on machine learning of the present invention. DETAILED DESCRIPTION

[0043] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that these embodiments are discussed solely to enable those skilled in the art to better understand and implement the subject matter described herein, and that the functions and arrangements of the elements discussed may be varied without departing from the scope of this specification. Various examples may omit, substitute, or add various processes or components as needed. Furthermore, features described in some examples may be combined in other examples.

[0044] At least one embodiment of the present invention discloses a backlight module light leakage measurement method based on machine learning, such as Figure 1 As shown, the following steps are included:

[0045] Step 1: Collect multi-dimensional optical feature information of the backlight module and construct an original feature matrix;

[0046] Step 1.1, image acquisition;

[0047] Using a high-precision CCD camera to capture the original image of the backlight module under controlled light conditions raw (x, y), where x and y represent the coordinate positions in the image.

[0048] Step 1.2, feature extraction;

[0049] Extract multi-dimensional optical features from raw images, including:

[0050] Brightness feature: extracts the brightness value B(x, y) of the pixel point to reflect the light intensity distribution of the backlight module;

[0051] Color features: Analyze the component values ​​C(x, y) of the RGB or HSV color space to capture the light leakage characteristics of different wavelengths of light;

[0052] Texture features: Calculate the gray level co-occurrence matrix GLCM(x, y) of the local area to characterize the texture characteristics of the light leakage area;

[0053] Edge features: Apply operators such as Sobel or Canny to extract edge gradient E(x, y) and identify light leakage boundaries.

[0054] Step 1.3, feature matrix construction;

[0055] The extracted features of each dimension are integrated into the original feature matrix F(x, y), and its calculation formula is:

[0056] F(x,y)=[B(x,y),C(x,y),GLCM(x,y),E(x,y)].

[0057] Among them, B(x,y) represents the brightness value at the coordinate (x,y), C(x,y) represents the color feature vector, GLCM(x,y) represents the texture feature vector, and E(x,y) represents the edge feature vector.

[0058] Step 2: Based on the ambient light conditions, the backlight module is preliminarily segmented using an adaptive threshold algorithm to screen out suspicious light leakage areas.

[0059] Step 2.1, calculation of local statistics;

[0060] Calculate the local mean and standard deviation for each region of the feature matrix F(x, y):

[0061]

[0062] Where μ(x, y) represents the mean of the local area centered at coordinate (x, y), w and h represent the width and height of the calculation window respectively, and F(i, j) represents the value of the feature matrix at coordinate (i, j);

[0063]

[0064] Among them, σ(x, y) represents the standard deviation of the local area centered on the coordinate (x, y), which is used to measure the discreteness of the eigenvalue.

[0065] Step 2.2, adaptive threshold generation;

[0066] According to the ambient light conditions, calculate the dynamic threshold T(x, y):

[0067] T(x,y)=μ(x,y)+k·σ(x,y);

[0068] Where T(x, y) represents the adaptive threshold at the coordinate (x, y), and k is the threshold coefficient, which is automatically adjusted according to the ambient light intensity and is usually between 0.5 and 2.0.

[0069] Step 2.3, preliminary segmentation of suspicious areas;

[0070] Apply the threshold to segment the feature matrix and generate a binary light leakage suspicious area marker map M(x, y):

[0071]

[0072] Wherein, M(x, y) represents the binarization result at the coordinate (x, y), a value of 1 indicates that the point is marked as a suspected light leakage area, and a value of 0 indicates that the point is a non-light leakage area.

[0073] Step 3: Apply a local contrast enhancement algorithm to the selected suspicious areas, combined with multi-scale pyramid analysis technology, to enhance the characteristics of weak light leakage signals;

[0074] Step 3.1, calculation of local statistical characteristics;

[0075] For each suspected light leakage area, calculate its local mean and local standard deviation:

[0076]

[0077] Among them, μ local (x, y) represents the mean value of the suspected light leakage area in the local window centered at coordinate (x, y), W(x, y) is the local window centered at (x, y), I(i, j) represents the value of the original image at coordinate (i, j), M(i, j) represents the value of the binary mark image at coordinate (i, j), and N is the number of pixels in the window marked as suspicious areas;

[0078]

[0079] Among them, σ local (x, y) represents the standard deviation of the suspected light leakage area in the local window centered at the coordinate (x, y), which is used to measure the degree of change of pixel values ​​in the local area.

[0080] Step 3.2, contrast enhancement;

[0081] Apply local contrast enhancement algorithm to enhance suspicious areas:

[0082]

[0083] Among them, I′(x, y) represents the value of the enhanced image at the coordinate (x, y), α contrast is the enhancement coefficient, which controls the degree of contrast enhancement, ε small To prevent a small constant from dividing by zero, M(x, y) ensures that only the suspected light leakage areas are enhanced.

[0084] Step 3.3, multi-scale pyramid analysis;

[0085] Construct a Gaussian pyramid {G0, G1, ..., G n} and Laplace pyramid {L0, L1, ..., L n};

[0086] G i+1 =DownSample(G i );

[0087] L i =G i -UpSample(G i+1 );

[0088] Among them, G0 represents the original image of the enhanced Gaussian pyramid, G1, G i , G n Represent the lth, i, and nth layer Gaussian pyramid images respectively, L0 represents the original image after enhancement of the Laplacian pyramid, L1, L i , L n They represent the lth, i, and nth layer Laplacian pyramid images respectively, n is the number of layers of the Gaussian pyramid and Laplacian pyramid, DownSample and UpSample represent the downsampling and upsampling operations of the image respectively.

[0089] The specific implementation of multi-scale pyramid analysis technology is as follows:

[0090] The downsampling operation DownSample uses a 5×5 Gaussian kernel to filter and take one of every two pixels, reducing the image size to half of its original size:

[0091]

[0092] Where w(i, j) is the Gaussian weight matrix, G(x, y) represents the value of the downsampled image at the coordinate (x, y), and I(x+i, y+j) represents the value of the original image at the coordinate (x+i, y+j).

[0093] The upsampling operation UpSample doubles the size of the image by inserting zero values ​​into it, and then applies the same 5×5 Gaussian kernel for interpolation:

[0094]

[0095] Among them, U(x, y) represents the value of the upsampled image at the coordinate (x, y), G i+1 represents the next layer of Gaussian pyramid image, and Indicates the corresponding coordinates in the next layer image.

[0096] In practical applications, the number of pyramid layers, n, is typically set to 3 to 5, adaptively adjusted based on the size of the backlight unit and the expected range of light leakage features. For large LCD panels, a 5-layer pyramid is used to capture light leakage features at scales from millimeters to centimeters. For small and medium-sized panels, a 3-layer pyramid is sufficient.

[0097] Step 3.4, feature enhancement fusion;

[0098] By weighted fusion of multi-scale features, the final enhanced image I is generated enhanced :

[0099]

[0100] Among them, w i is the weight coefficient of the i-th layer Laplacian pyramid, which is adaptively adjusted according to the significance of light leakage features at different scales. i Represents the i-th level Laplacian pyramid image.

[0101] Step 4: Apply box dimension calculation and multifractal spectrum analysis to the enhanced micro-area light leakage morphology to quantify the structural complexity of the light leakage area and construct a fractal feature vector;

[0102] Step 4.1, segmentation and marking of light leakage areas;

[0103] For enhanced image I enhanced Apply adaptive threshold segmentation to obtain accurate binary image B of the light leakage area binary (x,y), and mark the connected areas.

[0104] Step 4.2, box dimension calculation;

[0105] The box counting method is applied to each marked light leakage area to calculate its fractal dimension D:

[0106]

[0107] Where D represents the fractal dimension of the light leakage area, ε represents the side length of the square covering the light leakage area, N(ε) represents the number of squares with a side length of ε required to cover the light leakage area, and lim ε→0 It represents the limit when the side length of the grid approaches zero. In actual calculation, it is approximated by the slope of linear regression:

[0108] logN(ε)≈D·log(1 / ε)+C;

[0109] The actual calculation of the box dimension is implemented as follows:

[0110] Binarized light leakage area image B binary , (x, y) is covered with a series of grids of different scales, with the grid size ε decreasing exponentially:

[0111] ε i =ε0·2 -i ;

[0112] Where ε0 represents the initial grid size, ε i represents the i-th grid size, i = 0, 1, 2, ..., k-1, k is the total number of grid sizes, usually 8 to 12;

[0113] For each grid size ε i , calculate the number of grids N(ε i );

[0114] Draw a double logarithmic coordinate graph with the horizontal axis being log(1 / ε i ), the vertical axis is logN(ε i );

[0115] The least squares method is used to fit a straight line, and the slope of the straight line is the fractal dimension D.

[0116] In the actual application of LCD backlight module light leakage detection, the fractal dimensions of different types of light leakage are different:

[0117] Light leakage caused by bubbles: The fractal dimension D is usually between 1.0 and 1.3, approaching the Euclidean geometry of a circle;

[0118] Light leakage caused by indentation: The fractal dimension D is usually between 1.3 and 1.6, appearing as a slightly irregular star-shaped structure;

[0119] Light leakage caused by material fracture: The fractal dimension D is usually between 1.6 and 1.8, showing a highly irregular branching structure;

[0120] Light leakage caused by uneven penetration of optical adhesive: The fractal dimension D is usually between 1.4 and 1.7, but its multifractal spectrum width Δα is significantly larger than other types.

[0121] Through the precise quantification of fractal dimension and its multifractal characteristics, automatic classification of light leakage morphology can be achieved without manual experience judgment, thereby improving the intelligence level and reliability of the detection system.

[0122] Step 4.3, multifractal spectrum analysis;

[0123] Calculate the multifractal spectrum f(α) of the light leakage area:

[0124] f(α)=τ*(q)=q·α(q)-τ(q)

[0125] Here, f(α) represents the multifractal spectrum, τ(q) is the quality index, which describes the scaling characteristics of the light leakage region at different q values, α(q) is the singularity index, which describes the local scaling characteristics of the light leakage region at different locations, and q is the moment well index, which is used to adjust the weighting of regions with different densities. The multifractal spectrum can more comprehensively characterize the complex structural characteristics of the light leakage region.

[0126] Step 4.4, fractal feature vector construction;

[0127] Construct the fractal feature vector DF, which includes the box dimension, multifractal spectrum feature points and topological features:

[0128] DF=[D,f(α min ), f(α max ), α0, Δα, CN];

[0129] Where D is the box dimension, f(α min ) and f(α max ) are the function values ​​corresponding to the minimum and maximum values ​​of the multifractal spectrum, α0 is the value of α when f(α) takes the maximum value, which represents the dominant scaling characteristics of the light leakage area, Δα is the value range of α, which represents the multifractal intensity of the light leakage area, and CN is the connectivity number, which represents the topological characteristics of the light leakage area.

[0130] Step 5: Based on the machine learning algorithm, a correlation model between fractal feature vectors and optical properties is constructed to establish a mapping relationship between light leakage morphology and material defect type. The light leakage severity score is used to calculate and output the light leakage location, range, severity, and defect type.

[0131] Step 5.1, feature fusion;

[0132] Combine the fractal feature vector DF with the optical feature matrix F(x, y) to construct the comprehensive description vector CF of the light leakage feature:

[0133] CF=[DF,F avg , F var , F max , A, P];

[0134] Among them, CF represents the comprehensive description vector of light leakage features, DF is the fractal feature vector, and F avg 、F var 、F max They represent the average value, variance and maximum value of the optical characteristics in the light leakage area, A represents the light leakage area, and P represents the light leakage perimeter.

[0135] Step 5.2, defect type mapping model construction;

[0136] Based on the pre-annotated typical defect sample library, a mapping relationship model M between light leakage characteristics and material defect types is constructed. defect , and its calculation formula is:

[0137] Type defect =M defect (CF);

[0138] Among them, Type defect Indicates the predicted defect type, M defect Represents a mapping model that maps the comprehensive description vector CF of light leakage features to the corresponding defect type.

[0139] Defect type mapping model M defect The random forest classifier structure is used for implementation. The model consists of the following components:

[0140] Input layer: receives the feature vector CF as input;

[0141] Decision tree set: consists of multiple decision trees {T1, T2, T i ,…,T K}, where T1, T2, T i 、T K They represent the 1st, 2nd, i, and Kth decision trees respectively, where K is the number of decision trees. Each decision tree is trained based on a different subset of the feature vectors.

[0142] Voting integration layer: Summarizes the classification results of all decision trees and uses the majority voting strategy to determine the final defect type.

[0143] The model training adopts the standard random forest training method and is constructed through the following steps:

[0144] For the training data set D train Perform random sampling to generate K subsample sets {D1, D2, D i ,…,D K}, where D1, D2, D i 、D K They are represented as the 1st, 2nd, i, and Kth subsample sets respectively, where K is the number of decision trees and subsample sets;

[0145] For each subsample set D i Train a decision tree T i ;

[0146] When each decision tree splits a node, it randomly selects m features (m<M) from the M features of the feature vector CF as split feature candidates;

[0147] Select the optimal splitting feature and threshold based on information gain or Gini coefficient;

[0148] Repeat steps 3 and 4 until all decision trees are constructed.

[0149] In practical applications, the model is specifically optimized for typical defect types of LCD backlight modules and can accurately distinguish the following defect types:

[0150] Bubble defects (characterized by regular circular light leakage with a low fractal dimension); indentation defects (characterized by irregular star-shaped light leakage with a medium fractal dimension); material fractures (characterized by linear distribution of light leakage with a high fractal dimension); uneven penetration of optical adhesives (characterized by fuzzy boundary light leakage with a wide multi-fractal spectrum).

[0151] Step 5.3, light leakage severity scoring;

[0152] The light leakage severity score S is calculated based on the location, range, intensity and morphological characteristics of the light leakage:

[0153] S=w1·I norm +w2·A norm +w3·D norm +w4·L norm

[0154] Among them, S represents the light leakage severity score, I norm Represents the normalized light leakage intensity, A norm represents the normalized light leakage area, D norm represents the normalized fractal dimension, L norm It represents the normalized position factor, which is the distance between the light leakage area and the center of the screen. w1, w2, w3, and w4 are the first, second, third, and fourth weight coefficients, which are used to adjust the influence of different factors on the final score.

[0155] Step 5.4, output of evaluation results;

[0156] Generate a light leakage detection and assessment report containing the following information: light leakage location coordinates and distribution map; light leakage range and area statistics; severity score and classification; inferred possible defect type and confidence level; and fractal feature analysis results of light leakage morphology.

[0157] This embodiment achieves high-precision measurement of backlight module light leakage by organically combining the core technologies of multi-dimensional optical feature analysis, dynamic threshold adaptive segmentation, local contrast enhancement, fractal dimension quantification, and machine learning algorithms. In addition, the fractal dimension analysis method combined with the machine learning algorithm provides a precise means of quantifying the morphological characteristics of light leakage, overcoming the limitations of traditional geometric features in characterizing irregular light leakage morphology. At the same time, it realizes the correlation analysis between light leakage morphology and material defect type, and can directly infer the possible material defect type from the light leakage phenomenon, providing a direct basis for production process optimization. It is worth mentioning that this technology seamlessly integrates large-area detection and micro-area fine analysis in the same detection process, improving detection efficiency.

[0158] In one embodiment of the present invention, an example of the aforementioned backlight module light leakage measurement method based on machine learning is provided:

[0159] Application scenario description: A certain LCD display manufacturer faced the following problems during the quality inspection of 55-inch TV backlight modules: the lighting conditions on the production line were complex and changeable, and ambient light interfered with backlight leakage detection; traditional inspection methods had a low detection rate for tiny light leakage areas, resulting in some product quality issues being discovered only at the user end; manual inspection methods were highly subjective, with inconsistent test results and an inability to accurately determine the defect type.

[0160] The backlight module light leakage measurement method of this embodiment is used to solve the problem. The specific implementation process is as follows:

[0161] The experimental equipment and parameter configuration are shown in the following table:

[0162] Device / Parameter Category Specific configuration illustrate Image acquisition equipment 50-megapixel industrial camera Resolution 8192×5460 Light source system Uniform LED surface light source Color temperature 6500K±200K Processing Platform Industrial Computer Intel i7 processor, 32GB memory Dynamic threshold coefficient k Variable range 0.8-1.5 Automatically adjusts to ambient light intensity Contrast enhancement factor α 1.2 Experimental optimization parameters Multi-scale pyramid layers 4th floor Suitable for 55-inch backlight units

[0163] The size of the original backlight module image collected in this example is 8192×5460 pixels. The extracted feature dimensions and their data ranges are shown in the following table:

[0164] Feature Type Data Representation Numerical range Feature Dimension Brightness feature B(x, y) Single channel grayscale value 0-255 1 Color feature C(x, y) HSV color space H: 0-360, S: 0-1, V: 0-1 3 Texture feature GLCM(x,y) Energy, contrast, correlation, entropy Normalized to 0-1 4 Edge feature E(x, y) Gradient magnitude and direction Amplitude: 0-255, Direction: 0-360 2

[0165] In an actual test, the system detected a suspected light leakage signal in the lower right corner of the backlight module. The example values ​​of the brightness component of the local area (10×10 pixels) of the feature matrix F(x,y) are as follows:

[0166]

[0167] From this data, it can be seen that there is an obvious brightness anomaly in this area, and the brightness value of the central area is significantly higher than that of the surrounding area, which preliminarily conforms to the characteristic pattern of backlight leakage.

[0168] Applying the dynamic threshold adaptive segmentation algorithm to the above area, the local mean μ(x, y) is calculated to be 221.5, the local standard deviation σ(x, y) is 14.3, and the threshold coefficient k after automatic correction of the ambient brightness is 1.2. Therefore, the dynamic threshold T(x, y) is calculated to be 238.7.

[0169] Apply this threshold for segmentation and obtain the binary light leakage suspicious area marking matrix M(x, y) as follows:

[0170]

[0171] It can be seen that the system successfully marked the suspected light leakage areas and formed a preliminary light leakage area marking matrix.

[0172] Apply the local contrast enhancement algorithm to the marked suspicious area and calculate the local mean μ local (x, y) is 243.2 and the local standard deviation σ local (x, y) is 5.8, and the enhancement coefficient α contrast is 1.2, the small constant ε small is 0.001.

[0173] The contrast enhancement algorithm was applied to enhance the original image, which effectively improved the contrast between the light leakage area and the background. The brightness value of the center point was enhanced from 250 to 258.3, with an enhancement effect of about 3.3%, making the weak light leakage features more obvious.

[0174] The system then constructed a four-layer Laplacian of Gaussian pyramid and performed multi-scale analysis on the enhanced image. The third layer of the pyramid image revealed faint light leakage areas that had not been detected in the original segmentation. After feature enhancement and fusion, the final marked light leakage area was expanded by 23%, capturing a more complete light leakage morphology.

[0175] The final light leakage area is analyzed by fractal dimension, and box counting is performed using 8 grids of different scales (from 4 pixels to 512 pixels, scaled by powers of 2). The counting results are shown in the following table:

[0176] Grid size ε (pixels) The number of grids required for coverage N(e) log(1 / ε) Log N(ε) 4 783 5.64 6.66 8 215 4.64 5.37 16 64 3.64 4.16 32 21 2.64 3.04 64 8 1.64 2.08 128 3 0.64 1.10 256 1 -0.36 0 512 1 -1.36 0

[0177] The data were linearly fitted by the least square method, and the slope D = 1.56 was obtained, that is, the fractal dimension of the light leakage area was 1.56. At the same time, the multi-fractal spectrum feature points were calculated and f(α min ) is 0.92, f(α max ) is 0.88, α0 is 1.62, Δ α is 0.7, and the connectivity number CN is 1.

[0178] According to the fractal eigenvector [1.56, 0.92, 0.88, 1.62, 0.7, 1] and the optical characteristics obtained above, a comprehensive analysis shows that the light leakage area is most likely caused by an indentation defect.

[0179] The system calculates the light leakage severity score S based on the fused feature vector CF as 68 (out of 100), which is classified as moderate light leakage. The light leakage is located in the lower right area, with an area of ​​about 12.6mm. 2 The light leakage defect type was determined to be an indentation defect with a confidence level of 94.2%. This determination was automatically generated by a machine learning model, without the need for manual intervention or empirical judgment.

[0180] The test results were compared with the actual disassembly inspection results, confirming that the backlight module did have a slight indentation defect, verifying the accuracy of the system's judgment.

[0181] Technical effect verification:

[0182] 1,000 backlight modules sampled from the production line were tested, and the performance indicators of this method were compared with those of traditional testing methods, as shown in the following table:

[0183] Detection method Accuracy (%) False alarm rate (%) Missed detection rate (%) <![CDATA[Sensitivity (cd / m 2 )]]> Traditional methods 82.3 12.4 18.6 0.05 This method 98.7 7.2 3.4 0.01 Improvement effect +16.4 -5.2 -15.2 +80％

[0184] It can be seen that this method improves the detection accuracy, reduces the false alarm rate and missed detection rate, and improves the detection sensitivity by 80%.

[0185] Defect type identification was performed on 100 light leakage samples of known defect types, and the identification accuracy of each defect was statistically analyzed, as shown in the following table:

[0186] Defect Type Sample size Correctly identified number Recognition accuracy (%) Class A: Bubble defects 28 27 96.4 Class B: Indentation defect 35 33 94.3 Category C: Material fracture 22 21 95.5 Category D: Uneven penetration of optical adhesive 15 14 93.3 total 100 95 95.0

[0187] The overall accuracy of defect type identification reached 95.0%, and the recognition accuracy of various types of defects was above 93%, verifying the defect classification capability of this method based on machine learning.

[0188] To further enhance system performance, transfer learning techniques can be employed, using a pre-trained deep learning model as a feature extractor. Combined with the fractal feature analysis method presented in this paper, this approach can build a more robust light leakage detection system. Experiments have shown that this hybrid approach can increase defect recognition accuracy to over 97.5%, demonstrating exceptional performance when the number of samples is limited.

[0189] The above examples verify that this method can effectively solve the technical problems faced by backlight module light leakage detection in actual production environments.

Claims

1. A backlight module light leakage measurement method based on machine learning, characterized in that: The following steps are involved: Collect multi-dimensional optical feature information of the backlight module and construct the original feature matrix; Based on the ambient light conditions, the adaptive threshold algorithm is used to perform preliminary segmentation of the backlight module and screen out suspicious light leakage areas; Apply local contrast enhancement algorithm to the screened suspicious areas, combined with multi-scale pyramid analysis technology to enhance the characteristics of weak light leakage signals; For the enhanced micro-area light leakage morphology, box dimension calculation and multifractal spectrum analysis are applied to quantify the structural complexity of the light leakage area and construct the fractal feature vector. Based on the machine learning algorithm, a correlation model between fractal feature vectors and optical properties is constructed, and a mapping relationship between light leakage morphology and material defect type is established. The light leakage location, range, severity and defect type are calculated and output through the light leakage severity score.

2. The backlight module light leakage measurement method based on machine learning according to claim 1, characterized in that: The original feature matrix is ​​constructed as follows: F(x,y)=[B(x,y),C(x,y),GLCM(x,y),E(x,y)]; Among them, B(x, y) represents brightness features, C(x, y) represents color features, GLCM(x, y) represents texture features, and E(x, y) represents edge features.

3. The method according to claim 1, characterized in that The calculation formula of the adaptive threshold algorithm is: T(x,y)=μ(x,y)+k·σ(x,y); Where (x, y) represents the coordinates, T(x, y) represents the dynamic threshold, μ(x, y) is the mean of the local area, σ(x, y) is the standard deviation of the local area, and k is the threshold coefficient, which is automatically adjusted according to the ambient light intensity.

4. The backlight module light leakage measurement method based on machine learning according to claim 1, characterized in that: The formula for enhancing the screened suspicious areas by the local contrast enhancement algorithm is: Among them, I′(x, y) is the enhanced image, I(x, y) is the original image, μ local (x, y) is the local mean, σ local (x, y) is the local standard deviation, α contrast is the enhancement coefficient, ε small A small constant to prevent division by zero, M(x, y) is a suspicious area marking map.

5. The backlight module light leakage measurement method based on machine learning according to claim 1, characterized in that: The multi-scale pyramid analysis technology includes constructing a Gaussian pyramid and a Laplacian pyramid for the enhanced image, and generating a final enhanced image by weighted fusion of multi-scale features, wherein the weight coefficients are adaptively adjusted according to the significance of light leakage features at different scales.

6. The backlight module light leakage measurement method based on machine learning according to claim 1, characterized in that: The box dimension calculation is achieved by the following steps: Apply grids of different scales to the light leakage area, with the grid size decreasing exponentially; Calculate the number of grids required to cover the light leak area at each grid size; Draw a double logarithmic coordinate graph, with the horizontal axis being the logarithm of the grid size and the vertical axis being the logarithm of the number of grids; The slope was obtained as the fractal dimension by linear regression.

7. The backlight module light leakage measurement method based on machine learning according to claim 1, characterized in that: The association model between fractal features and optical properties is implemented using a random forest classifier structure. The model includes an input layer, a decision tree set, and a voting integration layer, and is capable of distinguishing different types of defects.

8. The backlight module light leakage measurement method based on machine learning according to claim 1, characterized in that: The fractal feature vector is constructed as follows: DF=[D,f(α min ),f(α max ), α0, Δα, CN]; Where D is the box dimension, f(α min ) and f(α max ) are the function values ​​corresponding to the minimum and maximum values ​​of the multifractal spectrum, α0 is the value of α when f(α) takes the maximum value, which represents the dominant scaling characteristics of the light leakage area, Δα is the value range of α, which represents the multifractal intensity of the light leakage area, and CN is the connectivity number, which represents the topological characteristics of the light leakage area.

9. The backlight module light leakage measurement method based on machine learning according to claim 1, characterized in that: The calculation formula for the light leakage severity score is: S=w1·I norm +w2·A norm +w3·D nrom +w4·L norm ; Among them, S represents the light leakage severity score, I norm Represents the normalized light leakage intensity, A norm represents the normalized light leakage area, D norm represents the normalized fractal dimension, L norm It represents the normalized position factor, which is the distance between the light leakage area and the center of the screen. w1, w2, w3, and w4 are the first, second, third, and fourth weight coefficients, which are used to adjust the influence of different factors on the final score.

10. A backlight module light leakage measurement system based on machine learning, used to execute the backlight module light leakage measurement method based on machine learning according to any one of claims 1 to 9, characterized in that: include: Image acquisition and feature extraction module, used to collect multi-dimensional optical feature information of the backlight module and construct the original feature matrix; Adaptive segmentation module, used to perform preliminary segmentation of the backlight module based on ambient light conditions using an adaptive threshold algorithm to screen suspected light leakage areas; The contrast enhancement module is used to apply a local contrast enhancement algorithm to the screened suspicious areas, combined with multi-scale pyramid analysis technology to enhance the characteristics of weak light leakage signals; The fractal feature analysis module is used to calculate the box dimension and perform multifractal spectrum analysis on the enhanced micro-area light leakage morphology, quantify the structural complexity of the light leakage area, and construct a fractal feature vector. The machine learning defect assessment module is used to build an association model between fractal feature vectors and optical properties based on machine learning algorithms, establish a mapping relationship between light leakage morphology and material defect type, and calculate and output the light leakage location, range, severity and defect type through the light leakage severity score.

Images