A small sample micro LED display chroma prediction method, medium and equipment

Abstract

This invention discloses a method, medium, and device for predicting the chromaticity of a MicroLED display using a small sample dataset, belonging to the field of digital display technology. The method includes: S1. Constructing a small sample benchmark dataset; S2. Performing physical constraint-based data fission processing on the small sample benchmark dataset to expand the data volume, obtaining an augmented dataset; S3. Training a Bayesian neural network using the augmented dataset, fitting the chromaticity mapping during training, and preventing overfitting through a loss function; S4. Collecting sample data to be predicted, inputting it into the trained Bayesian neural network for prediction, and obtaining the mean chromaticity coordinates and their uncertainty; S5. Evaluating the confidence level based on the prediction results and determining the color difference compensation decision. This invention can solve the technical problems of existing chromaticity prediction technologies, which are prone to severe overfitting with small sample calibration data and cannot quantify the confidence level of prediction results, leading to blindly confident erroneous predictions under extreme operating conditions.

Description

Technical Field

[0001] This invention belongs to the field of digital display technology, specifically relating to a method, medium, and device for predicting the chromaticity of a small sample MicroLED display based on a Bayesian neural network. Background Technology

[0002] After years of development, MicroLED displays have become the mainstream technology for next-generation high-end flat panel displays. Due to the light-emitting characteristics of MicroLEDs, severe nonlinear chromaticity drift occurs under different driving grayscale levels and operating temperatures. To meet high-quality display requirements, high-precision prediction and compensation of color differences must be performed at the system level. The accuracy of chromaticity prediction directly determines the effectiveness of subsequent compensation. Because the chromaticity drift law is extremely complex, existing physical compensation formulas are insufficient to meet the accuracy requirements, and the industry has begun to introduce machine learning models.

[0003] However, in actual engineering implementation, companies often find it difficult to spend months calibrating tens of thousands of data sets for network learning. When introducing existing traditional machine learning or shallow neural network models with small sample sizes, they face significant technical limitations: on the one hand, traditional probabilistic machine learning models or probabilistic networks based on approximate calculations either have limited feature representation capabilities and cannot meet the high-precision fitting requirements of MicroLED nonlinearity, or are prone to underfitting, leading to a cliff-like drop in accuracy, and the output prediction variance is often severely distorted; on the other hand, although conventional deterministic neural networks can achieve high fitting accuracy under normal distributions, they are prone to overfitting and lack the ability to quantify and recognize uncertainty. When faced with unknown blind spots (OOD) such as extreme abnormal temperatures, they will still give blindly confident and erroneous predictions, posing a significant risk of system runaway.

[0004] In summary, existing technologies have consistently failed to achieve both accurate color difference compensation and rigorous prediction of unknown blind zones under the constraint of small samples. Therefore, there is an urgent need for a new small-sample MicroLED display color prediction method based on Bayesian neural networks. Summary of the Invention

[0005] The present invention aims to at least partially solve one of the technical problems in the aforementioned related technologies.

[0006] Therefore, the purpose of this invention is to provide a small-sample MicroLED display chromaticity prediction method based on Bayesian neural networks, which can solve the technical problems of existing chromaticity prediction technology being prone to severe overfitting under small-sample calibration data and being unable to quantify the confidence of prediction results, leading to erroneous predictions with blind confidence under extreme working conditions.

[0007] To solve the above-mentioned technical problems, the present invention is implemented as follows: This invention provides a method for predicting the chromaticity of a small-sample MicroLED display, the method comprising: S1. Construct a small-sample benchmark dataset; S2. Perform physical constraint-based data fission processing on the small sample benchmark dataset to expand the data volume and obtain the augmented dataset; S3. Use an augmented dataset to train a Bayesian neural network, fit the chromaticity mapping during training, and prevent overfitting through a loss function; S4. Collect the sample data to be predicted, input it into the trained Bayesian neural network for prediction, and obtain the mean chromaticity coordinates and their uncertainty. S5. Based on the prediction results, conduct a confidence assessment and determine the color difference compensation decision.

[0008] In addition, the small-sample MicroLED display colorimetric prediction method according to the present invention may also have the following additional technical features: In some implementations, step S1 includes: converting the R, G, B driving grayscale parameters and ambient temperature to obtain X, Y, Z chromaticity coordinates, and organizing the X, Y, Z chromaticity coordinates into a small sample benchmark dataset.

[0009] In some implementations, step S2 includes: Establish a noise model: Define a normal distribution noise model based on the physical noise floor tolerance of the measuring instrument; Execute fission: Process each sample in a small sample benchmark dataset using a noise model to generate multiple noisy copies of the data; The noisy copy data and the original sample data are merged to form an augmented dataset.

[0010] In some of these implementations, the noise model includes the following: Add N(0,0.2) Gaussian noise to the R, G, B driving grayscale parameters in the input features; Add N(0,0.5) Gaussian noise to the temperature feature in the input; Add relative proportional noise to the X, Y, and Z chromaticity coordinates of the output labels.

[0011] In some implementations, the Bayesian neural network structure in step S3 includes: The input layer has a four-dimensional structure including R, G, B, and temperature T; The hidden layer is a multi-layer Bayesian weighted layer, and the weights and biases follow a prior Gaussian distribution; The activation function is a smooth nonlinear function to ensure that the gradient is continuous and stable; The output layer is a 3-dimensional Bayesian weighted layer.

[0012] In some implementations, the loss function in step S3 is: (1) in, To predict the linear normalized mean square error between the X, Y, Z chromaticity coordinates and the true chromaticity coordinates; To enhance the log-space mean square error of low-grayscale dark fields; The weighted mean square error mapped to a uniform color space. The divergence regularization loss is used to measure the difference between the posterior and prior distributions of a network. , , , These are the hyperparameter weights for the corresponding loss terms.

[0013] In some implementations, Monte Carlo sampling is performed during prediction in step S4, multiple sets of weight parameters are extracted from the posterior distribution of the weights, multiple prediction outputs are obtained for the same input, the mean of these prediction outputs is calculated as the final chromaticity coordinate prediction value, and then the variance of these prediction outputs with respect to the mean is calculated as the uncertainty index.

[0014] In some implementations, step S5 includes: Set a first threshold; If the variance is less than the first threshold, the current prediction is determined to be in the high confidence interval, the corresponding final chromaticity coordinate prediction value is output, and color difference compensation based on the prediction value is performed. If the variance is greater than or equal to the first threshold, the current prediction is determined to be in the low confidence range, the condition is marked as an extreme physical blind zone risk state, and the system is switched to the minimum guarantee drive mode or an early warning is issued to avoid blind compensation.

[0015] This invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the small-sample MicroLED display chromaticity prediction method as described in any of the preceding embodiments.

[0016] This invention also provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the small sample MicroLED display chromaticity prediction method as described in any of the preceding embodiments.

[0017] Compared with the prior art, the present invention has at least the following beneficial effects: In this embodiment of the invention, the small-sample MicroLED display colorimetric prediction method based on Bayesian neural networks is adapted to the pain point of small samples in the MicroLED industry: it solves the problem of excessively high optical calibration costs for display panels at the factory; only a small amount of basic calibration data (about 300 sets) is needed to complete reliable color difference modeling under a wide temperature range and full grayscale through the divergence regularization characteristics of Bayesian networks, which greatly shortens the hardware development and factory calibration cycle; In this embodiment of the invention, the small-sample MicroLED display chromaticity prediction method based on Bayesian neural networks has extremely strong anti-overfitting and blind zone prediction capabilities in small samples. It overcomes the shortcomings of traditional machine learning or neural network technologies, which are prone to severe overfitting under small sample data, lack cognitive boundary perception capabilities in the face of blind zones, and are prone to making erroneous predictions based on blind confidence. It effectively improves the reliability and authenticity of chromaticity prediction for MicroLEDs under wide temperature range and complex grayscale conditions, and provides a scientific basis with confidence labels for the robust decision-making and early warning of downstream display systems in complex environments.

[0018] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0019] Figure 1 This is a flowchart of a neural network-based colorimetric prediction method for MicroLED displays, as disclosed in one embodiment of the present invention. Figure 2 This is a flowchart for establishing a neural network model for training, as disclosed in one embodiment of the present invention; Figure 3 This is a prediction graph of the u' coordinates and the 95% confidence interval in a white field at 60°C, as disclosed in one embodiment of the present invention. Figure 4 This is a prediction plot of the v' coordinates and the 95% confidence interval in a white field at 60°C, as disclosed in one embodiment of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings and specific examples and application scenarios.

[0022] In some embodiments of the present invention, a small-sample MicroLED display chromaticity prediction method based on a Bayesian neural network is provided, comprising the following steps: S1. Place the MicroLED panel in a high and low temperature test chamber. By controlling the temperature of the high and low temperature test chamber, and using the ambient temperature to represent the panel temperature, obtain the driving grayscale parameters of the panel under limited calibration conditions, the ambient temperature, and the corresponding actual emission X, Y, Z chromaticity coordinates and other parameters to construct a small sample benchmark dataset. S2. Extract the physical noise floor tolerance of the measuring equipment, inject normally distributed random noise into the input features and output labels of the small sample benchmark dataset, and perform data fission processing to obtain an expanded dataset; S3. Input the amplified dataset into a neural network model configured with Bayesian weight layers for training. During the training process, KL divergence loss is introduced as a regularization term to suppress small sample overfitting. S4. Obtain the panel driving grayscale parameters and ambient temperature to be tested and input them into the trained neural network model, and output the predicted mean values ​​of X, Y, and Z chromaticity coordinates and the variance index that quantitatively represents the prediction confidence level. S5. Based on the quantitative variance index, assess the reliability level of the currently predicted mean values ​​of X, Y, and Z chromaticity coordinates, generate and output the chromaticity coordinate prediction results with confidence labels, and provide them to the downstream display control system for color difference compensation decisions.

[0023] In some embodiments of the present invention, the small sample benchmark dataset in step S1 contains approximately 300 sets of luminescent samples. Its input features include R, G, and B driving grayscale parameters and ambient temperature representing the panel temperature. The output labels include the X, Y, and Z chromaticity coordinates of the actual luminescence emitted by the panel.

[0024] In some embodiments of the present invention, the data fission processing described in step S2 specifically includes: establishing a normal distribution noise model based on the physical tolerance of the measuring instrument; introducing Gaussian noise with a mean of 0 and a set standard deviation to the input features, and introducing relative Gaussian noise with a set proportion to the output chromaticity coordinates; and combining and amplifying the data after superimposing noise with the original benchmark dataset multiple times to fill the feature sparsity region between small sample data.

[0025] In some embodiments of the present invention, the structure of the neural network model in step S3 includes: an input layer for receiving multi-dimensional input features; multiple hidden layers and an output layer, all configured as Bayesian weight layers that model their internal weights and biases as controlled probability distributions. The Bayesian weight layers no longer use fixed deterministic values, but instead set their learnable parameters to include the mean (…). μ ) and variational parameters used to control variance ( pIts advantage lies in the fact that by introducing a probability distribution, when the model faces unknown blind spots lacking calibration data, the randomness of the weight distribution can be transmitted to the output, naturally exposing a large output variance (cognitive uncertainty), thus providing an underlying mechanism for subsequent quantitative evaluation of prediction confidence and early warning of runaway; and a smooth nonlinear activation function connecting each hidden layer is used to ensure the gradient continuity and smooth update of the probability distribution during backpropagation.

[0026] When training the neural network model, the direct random sampling of network weights is mathematically non-differentiable, which can cause gradient propagation to break in the backpropagation algorithm. To ensure end-to-end optimization of the model, this embodiment employs a reparameterization technique during training. Specifically, this technique abandons the non-differentiable operation of directly sampling from the network weight distribution and instead introduces an independent standard Gaussian noise perturbation term that does not participate in gradient updates. Through this reparameterization operation, the originally non-differentiable random sampling black box is equivalently decomposed into a deterministic computation channel with differentiable parameters and a bypass-accessed pure random noise. Thus, during backpropagation, the gradient flow can directly bypass the random noise term and smoothly calculate the internal variational parameters (mean) to be optimized. μ The partial derivatives of the parameters used to control variance are used to ensure that the Bayesian neural network can be seamlessly integrated into the standard gradient descent framework to complete smooth and stable iterative updates.

[0027] Its joint loss function consists of a weighted sum of four parts: (1) These are the linearly normalized mean square errors of the predicted X, Y, and Z chromaticity coordinates compared to the true chromaticity coordinates. Enhance the log-space mean square error of low-grayscale dark fields Weighted mean square error mapped to uniform color space And the divergence regularization loss, which measures the difference between the posterior and prior distributions of the network. ;in, , , , These are the hyperparameter weights corresponding to the loss terms. The control system minimizes this loss function. The backpropagation algorithm is executed to iteratively update the weight probability distribution parameters of each layer of the Bayesian network until the model reaches the preset convergence condition.

[0028] Furthermore, in formula (1) Specifically: (2) in, It is the firsti Predicted chromatic tristimulus values ​​in each sample It is the true chromaticity tristimulus value in the i-th sample, using and The basic linear normalized mean square error is calculated. This refers to the number of samples in a single training iteration. This is the sample number.

[0029] In some embodiments of the present invention, since even small luminescence deviations can easily lead to severe color casts visible to the naked eye at low grayscale (dark field) conditions, the coordinates are inversely normalized to the actual physical range, and the logarithm is taken to calculate the logarithmic mean square error. This non-linearly amplifies the error gradient weights in the dark field region. (3) in, A minimal positive real constant set to prevent logarithmic zero overflow. .

[0030] In some embodiments of the present invention, in order to make the loss function of network training more closely resemble the actual perceptual characteristics of human vision regarding color difference and overcome the problem of uneven color difference evaluation in conventional physical space, this embodiment introduces the standard CIE 1976. Uniform chromaticity space. Specifically, using the following standard nonlinear transformation equation, the predicted tristimulus values ​​are mapped to a uniform chromaticity space, and their chromaticity coordinates are calculated. : (4) Based on this, and considering that the human eye's sensitivity to color difference is strongly modulated by background brightness, this invention extracts a value representing the true brightness of the panel. Stimulus values, constructing nonlinear visual weight distribution coefficients By coupling this weighting coefficient with the geometric distance in a uniform color space, a nonlinear weighted error that conforms to human visual perception is calculated. : (5).

[0031] Wherein, the variable dispersion regularization loss The calculation form is: (6).

[0032] The variable dispersion regularization loss Used to quantitatively measure the geometric distance between the variational posterior distribution and the prior distribution. The weights are predefined prior distributions (in this network, they are set to a standard Gaussian distribution); This represents the variational posterior distribution learned by the Bayesian weighted layer in the current iteration state. When training with a finite number of calibration data for MicroLED panels, the network is prone to over-memorizing data noise due to the sparse sample size, leading to excessive convergence and collapse of the posterior distribution into a traditional deterministic value. By introducing Equation (6) into the joint loss function, a strong regularization penalty can be mathematically applied to constrain the variational posterior distribution. Without deviating from reasonable prior boundaries, the risk of overfitting under small sample calibration is effectively suppressed. At the same time, the mechanism forcibly preserves the variance tolerance of network weights, ensuring that when the model faces unknown abnormal working conditions, it can transmit the uncertainty of the weights to the output end, and thus accurately output a high variance warning signal that represents cognitive uncertainty.

[0033] In some embodiments of the present invention, the specific steps of evaluating the reliability level of the current predicted X, Y, Z chromaticity coordinate mean based on the quantitative variance index in step S5 include: when the variance index is less than a preset first threshold, determining that the current chromaticity prediction result is in a high confidence interval, and outputting a high-confidence predicted mean of X, Y, Z chromaticity coordinates; when the variance index is greater than or equal to the preset first threshold, determining that the current prediction confidence has decreased, and marking the prediction result under the current ambient temperature and driving grayscale parameters as a low-confidence state with extreme physical blind zone risk.

[0034] Example 1: Please see Figure 1 and Figure 2 As shown, this embodiment of a small-sample MicroLED display chromaticity prediction method is deployed in the logic control unit at the bottom of the panel. The system receives the current RGB driving grayscale value and the real-time ambient temperature signal; after the signal enters the Bayesian network, the network calculates the compensation mean of the target chromaticity coordinates and the variance distribution of the current prediction confidence through Monte Carlo sampling; the control unit adaptively adjusts the distribution of compensation weights or switches to a robust safety-of-base driving mode by determining the level of the current confidence.

[0035] The preparation and execution steps of the small-sample MicroLED display chromaticity prediction method based on physical constraint data fission and Bayesian networks are as follows: Step 1: Set physical tolerance boundaries and perform compact data fission 1.1 Obtain a limited benchmark dataset for factory testing of MicroLED panels, with an effective number of light-emitting samples of approximately 300 sets. The data features include the input R, G, and B grayscale parameters, the test temperature, and the corresponding X, Y, and Z chromaticity coordinates.

[0036] 1.2 Extract the physical noise floor tolerance of the optical testing instrument and establish a normal distribution noise model accordingly. Gaussian noise with a mean of 0 and a standard deviation of 0.2 is introduced into the input features, and Gaussian noise with a standard deviation of 0.5 is introduced into the temperature features; relative Gaussian noise with a ratio of 0.002 is introduced into the output chromaticity coordinates.

[0037] 1.3. The above physical noise is superimposed on a limited benchmark dataset and the data is amplified by a set multiple to fill the feature sparsity area between small sample data to form a training set.

[0038] Step 2: Construct a Bayesian neural network to suppress overfitting 2.1 Establish a multi-layer feedforward network structure with an input dimension of 4 (R,G,B,T) and an output dimension of 3 (X,Y,Z).

[0039] 2.2 Configure the hidden and output layers of the network as Bayesian weighted layers so that the weights and biases within the network follow a prior Gaussian distribution.

[0040] 2.3. A smooth nonlinear activation function is used in the hidden layer of the network. The smooth nonlinear activation function includes, but is not limited to, Softplus, GELU or Swish function, to ensure the gradient continuity and smooth update of the probability distribution during backpropagation under small sample conditions.

[0041] Step 3: Calculate the joint loss function of physical constraints and perform iterative optimization. 3.1 During the small sample training process, extract the predicted chromaticity coordinates and the real target coordinates from the network output, and calculate the basic linear mean square error.

[0042] 3.2 After inversely normalizing the coordinates to the real physical range, take the logarithm to calculate the logarithmic mean square error, thereby amplifying the error gradient weight in the low grayscale (dark field) range.

[0043] 3.3. Using formulas, the physical coordinate system is mapped to a uniform color space, and a nonlinear weight distribution is constructed in combination with brightness to calculate the human visual perception error.

[0044] 3.4. Calculate the KL divergence between the posterior and prior distributions of all Bayesian layers in the network using the variable dispersion formula, and use this regularization term to suppress the network's memory of small sample noise.

[0045] 3.5. Sum the above errors with weights and update the network weight distribution based on the weighted summation result.

[0046] Step 4: Quantification and Output of Prediction Confidence 4.1. Real-time acquisition of the driving grayscale to be compensated and the ambient temperature.

[0047] 4.2 The network performs probabilistic inference and outputs the cognitive variance distribution of the chroma compensation mean and the quantitative representation of the prediction confidence.

[0048] 4.3 Output the calculated cognitive variance to the downstream control unit.

[0049] For the parts of this invention not described in detail, please refer to the prior art or the art known to those skilled in the art. This embodiment does not limit these aspects and will not describe them in detail here.

[0050] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of the present invention.

Claims

1. A method for predicting the chromaticity of a small-sample MicroLED display, characterized in that, The method includes: S1. Construct a small-sample benchmark dataset; S2. Perform physical constraint-based data fission processing on the small sample benchmark dataset to expand the data volume and obtain the augmented dataset; S3. Use an augmented dataset to train a Bayesian neural network, fit the chromaticity mapping during training, and prevent overfitting through a loss function; S4. Collect the sample data to be predicted, input it into the trained Bayesian neural network for prediction, and obtain the mean value of chromaticity coordinates and its uncertainty. S5. Based on the prediction results, conduct a confidence assessment and determine the color difference compensation decision.

2. The method for predicting the chromaticity of a small-sample MicroLED display according to claim 1, characterized in that, Step S1 includes: converting the R, G, B driving grayscale parameters and ambient temperature to obtain X, Y, Z chromaticity coordinates, and organizing the X, Y, Z chromaticity coordinates into a small sample benchmark dataset.

3. The method for predicting the chromaticity of a small-sample MicroLED display according to claim 1, characterized in that, Step S2 includes: Establish a noise model: Define a normal distribution noise model based on the physical noise floor tolerance of the measuring instrument; Execute fission: Process each sample in a small sample benchmark dataset using a noise model to generate multiple noisy copies of the data; The noisy copy data and the original sample data are merged to form an augmented dataset.

4. The method for predicting the chromaticity of a small-sample MicroLED display according to claim 3, characterized in that, The content of establishing the noise model includes: Add N(0,0.2) Gaussian noise to the R, G, B driving grayscale parameters in the input features; Add N(0,0.5) Gaussian noise to the temperature feature in the input; Add relative proportional noise to the X, Y, and Z chromaticity coordinates of the output labels.

5. The method for predicting the chromaticity of a small-sample MicroLED display according to claim 1, characterized in that, The Bayesian neural network structure in step S3 includes: The input layer has a four-dimensional structure including R, G, B, and temperature T; The hidden layer is a multi-layer Bayesian weighted layer, and the weights and biases follow a prior Gaussian distribution; The activation function is a smooth nonlinear function to ensure that the gradient is continuous and stable; The output layer is a 3-dimensional Bayesian weighted layer.

6. The method for predicting the chromaticity of a small-sample MicroLED display according to claim 1, characterized in that, The loss function in step S3 is: (1) in, To predict the linear normalized mean square error between the X, Y, Z chromaticity coordinates and the true chromaticity coordinates; To enhance the log-space mean square error of low-grayscale dark fields; The weighted mean square error mapped to a uniform color space. The divergence regularization loss is used to measure the difference between the posterior and prior distributions of a network. , , , These are the hyperparameter weights for the corresponding loss terms.

7. The method for predicting the chromaticity of a small-sample MicroLED display according to claim 1, characterized in that, In step S4, Monte Carlo sampling is performed during prediction. Multiple sets of weight parameters are extracted from the posterior distribution of the weights. Multiple prediction outputs are obtained for the same input. The mean of these prediction outputs is calculated as the final chromaticity coordinate prediction value. The variance of these predicted outputs from the mean is then calculated as an indicator of uncertainty.

8. The method for predicting the chromaticity of a small-sample MicroLED display according to claim 7, characterized in that, Step S5 includes: Set a first threshold; If the variance is less than the first threshold, the current prediction is determined to be in the high confidence interval, the corresponding final chromaticity coordinate prediction value is output, and color difference compensation based on the prediction value is performed. If the variance is greater than or equal to the first threshold, the current prediction is determined to be in the low confidence range, the condition is marked as an extreme physical blind zone risk state, and the system is switched to the minimum guarantee drive mode or an early warning is issued to avoid blind compensation.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the small sample MicroLED display chromaticity prediction method according to any one of claims 1-8.

10. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the steps of the small sample MicroLED display chromaticity prediction method according to any one of claims 1-8.

Images