Light distribution measurement correction method and device

Abstract

The invention discloses a light distribution measurement correction method and device, and the method comprises the steps: carrying out the imaging of a light spot image, irradiated on a diffusion screen, of a first measured light source through imaging measurement equipment, obtaining the light distribution image data, and carrying out the scanning measurement of the relative movement of a stray light elimination photometric detector and the first measured light source, so as to obtain light distribution standard data, constructing a training sample data set consisting of a plurality of groups of light distribution image data and light distribution standard data; establishing a light distribution correction model based on deep learning, training the light distribution correction model by using the training sample data set, and calculating and optimizing model parameters through a loss function to obtain an optimized light distribution correction model; and measuring light distribution image data irradiated on the diffusion screen by the second measured light source by using imaging measurement equipment, and inputting the light distribution image data into the optimized light distribution correction model to obtain corrected light distribution data. The method is widely applied to various illuminant objects, realizes rapid self-adaptive correction processing, and improves the measurement precision and efficiency.

Description

Technical Field

[0001] This invention relates to the field of optical radiation measurement and image processing technology, specifically to a method and apparatus for optical distribution measurement and correction. Background Technology

[0002] The luminance or radiance distribution (hereinafter referred to as "light distribution") of a light source (such as a lamp or display device) is an important performance data point that plays a crucial role in the quality assessment and control of related products. A commonly used method in the industry is to measure the light distribution of a light source using an imaging measurement device paired with a diffuse reflection screen. This involves projecting a signal beam emitted by the light source onto a diffuse reflection screen and then using the imaging measurement device to measure and acquire the reflected light spot image on the screen. This method can quickly obtain the light distribution information of the light source under test. Provided there are sufficient pixels and good image quality, it can also achieve high resolution. Currently, it is widely used in scenarios such as automotive headlight light distribution testing and measurement of optical characteristics of information displays.

[0003] However, the accuracy of this method in measuring light distribution is limited. Besides the fact that the spectral matching and linearity of the device itself are inferior to those of a single-channel optical radiation detector, the main reason is stray light. Stray light sources include existing light sources in the environment (such as various signal lights, causing light leakage), light from the tested light source that illuminates the diffuse reflection screen and diffuses into the environment, and light from the tested light source that does not illuminate the diffuse reflection screen but directly enters the environment. This light undergoes one or more reflections in the test environment before being received by the imaging measurement device. Crosstalk also occurs on the sensor surface inside the imaging measurement device, producing a light distribution with stray light, which may differ significantly from the actual signal light distribution of the tested light source. Furthermore, the precise positioning or mapping relationship between the measured light distribution and the actual light spot image is often a factor to consider. Alignment between the sample and the system, operational factors, etc., during the actual measurement process can all introduce errors into the imaging measurement method.

[0004] To address the stray light problem, hardware modifications have been employed, such as constructing fully absorbing darkrooms and covering with low-reflectivity coatings. While these methods can reduce stray light generation to some extent, the stray light level is often still quite high. Without subtraction or correction, it cannot meet the requirements for high-precision measurements. Algorithm-based correction methods also exist, such as background subtraction, various filters, ray tracing models, and point spread function-based models. These methods largely rely on the analysis and modeling of the physical properties of the test environment. Establishing accurate models for various influencing factors is computationally complex and accuracy is difficult to guarantee. Furthermore, they are challenging to implement in actual testing and are difficult to apply to different types of light sources and test environments. Additionally, some methods superimpose an illuminance meter onto the imaging method to scan the sensitive area. However, this method is not suitable for correcting the entire light distribution image data, and for different light sources, measurement and correction need to be performed again for each test, resulting in low efficiency. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a light distribution measurement and correction method and apparatus. It aims to solve the problems that existing light distribution measurement and correction methods rely on the physical hardware conditions of the test environment, making it difficult to adapt to complex and ever-changing real-world test scenarios, thus limiting measurement accuracy. This new method falls under the category of artificial intelligence using deep learning neural network models and can be widely applied to different types of light sources under test and test environments, achieving fast and adaptive correction and improving measurement accuracy.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] This invention provides a method for measuring and correcting light distribution, specifically including the following steps:

[0008] S1: The light emitted by the first light source under test illuminates the diffuse screen and generates a light spot image. The imaging measurement device images the light spot image on the diffuse screen and obtains its light distribution image data. The stray light photometer detector moves relative to the first light source under test and scans to obtain the light distribution data emitted by the first light source under test, and uses this light distribution data as the light distribution standard data. By collecting measurement data from multiple different light sources, several sets of light distribution image data and corresponding light distribution standard data are obtained to construct a training sample dataset.

[0009] S2: Establish a deep learning-based light distribution correction model, and train the light distribution correction model using a training sample dataset. Specifically, this includes: inputting light distribution image data into the light distribution correction model, and the model outputting predicted light distribution correction data; using the predicted light distribution correction data and the corresponding standard light distribution data to participate in the calculation of the loss function, optimizing the model parameters with the goal of minimizing the calculated value of the loss function, and obtaining the optimized light distribution correction model.

[0010] S3: In the same or different measurement system as step S1, use an imaging measurement device to measure and obtain the light distribution image data of the light emitted by the second light source under test illuminating the diffuse screen. Input the light distribution image data into the optimized light distribution correction model to obtain the corrected light distribution data.

[0011] A schematic diagram of the light distribution measurement and correction method described in this invention is shown below. Figure 1As shown, step S1 obtains a training sample dataset based on measured light distribution image data and its standard light distribution data. Unlike traditional calibration methods, this does not require specific physical hardware attribute information for various test environments and avoids deviations caused by the difference between the idealized conditions of modeling analysis and the actual test conditions. In step S2, a light distribution calibration model based on deep learning is constructed. The light distribution calibration model is trained using the measured training sample dataset, which allows the light distribution measurement calibration model to learn and fit the complex mapping process for calibrating different light distribution image data, making it more in line with the high-precision requirements of actual measurement calibration. After training, in step S3, the optimized light distribution calibration model is used to apply calibration to the light distribution image data of the light source, directly outputting the calibrated light distribution data, realizing end-to-end measurement calibration application, and ensuring both measurement accuracy and calibration efficiency.

[0012] As a supplementary explanation, the light distribution refers to the spatial distribution of the luminance or radiance of the light source of interest, including but not limited to luminous intensity distribution, radiant intensity distribution, irradiance distribution of the illuminated surface, and brightness distribution of the light source or the illuminated surface. The light distribution image data is obtained by imaging and measuring the light spot image formed by the light source on a diffuser screen using an imaging measurement device with an array detector, and is expressed as Y = (y... a,b In the light distribution image data, Y is the symbol for luminance or radiance of interest, which can be specifically expressed as luminous intensity I, illuminance E, etc., and y represents the luminance or radiance value of interest. In the light distribution image data, (a,b) represents the spatial coordinates corresponding to the light source beam, specifically the spatial angle of the light emitted by the light source. Or the spatial position (x, y) on the illuminated surface; for luminous intensity distribution or radiant intensity distribution, spatial angles are typically used. The directional coordinates are taken as the origin of the luminance center of the light source. In practical applications, the luminance center is generally a practically marked position provided by the manufacturer based on the characteristics of the light source itself, called the reference center, denoted by the letter C. For the illuminance distribution or irradiance distribution of the illuminated surface, the spatial position coordinates (x, y) on the illuminated surface are usually used. Here, the illuminated surface refers to the plane where the diffuser screen receiving the light spot image of the light source is located. The light source mentioned is either the first tested light source or the second tested light source. In this article, the first tested light source refers to the light source used to construct the training sample dataset, which can be a standard light source or a non-standard light source; the second tested light source refers to various different types of tested light source objects.

[0013] In a measurement scenario where the imaging measurement equipment, diffuse screen, and light source pose are determined, (a,b) can be calculated by transforming the pixel coordinates (i,j) of the array detector with the positional relationship between the light source pose and the diffuse screen. Each pixel on the array detector corresponds one-to-one with a specified position point on the diffuse screen. After calibration, the coordinates (x,j) of each point on the illuminated surface, i.e., the diffuse screen, can be directly transformed from the pixel coordinates (i,j) to obtain the coordinates (x,j). i,j ,y i,j ); Position point on the diffuse screen (x i,j ,y i,j The direction of the corresponding luminous intensity is denoted as . The vector direction from the reference center C to that point on the diffuse screen and the orientation of the light source can be determined through coordinate vector transformation. Calculated.

[0014] Furthermore, the poses of the first light source under test and the diffuser screen are relatively fixed, and the imaging measurement device measures the light spot emitted by the first light source under test illuminating the diffuser screen under this relative pose to obtain light distribution image data; or the first light source under test and the diffuser screen undergo relative motion, and the imaging measurement device measures the light spot emitted by the first light source under test illuminating the diffuser screen under two or more relative poses to obtain two or more light distribution image data, and integrates them to obtain new light distribution image data.

[0015] As a technical solution, the relative motion between the first light source under test and the diffuse screen is achieved by mounting the first light source under test on a sample stage, with the sample stage driving the first light source under test to move. This movement includes rotation and / or translation, thereby changing the pose of the light source. At this time, the light distribution image generated on the fixed diffuse screen by the light emitted from the light source also changes accordingly. Two or more light distribution image data obtained by the imaging measurement device under two or more relative poses are respectively expressed in Y... 1, Y2……Y g (g≥2) indicates that when two light distribution images of light sources illuminating a diffuse screen at two adjacent relative poses have overlapping or closely adjacent portions, these overlapping or closely adjacent light distribution image portions can be used for image stitching and integration, such as... Figure 2 As shown, the new light distribution image data Y is finally obtained.

[0016] The rows and columns of the light distribution image data Y correspond to the light distribution spatial coordinates (a, b) of the light source, respectively. m,n ,b m,n () represents the spatial coordinates of the light distribution in the m-th row and n-th column of the light distribution image data. This represents the specific value corresponding to the m-th row and n-th column in the light distribution image data. For simplicity, when the light distribution spatial coordinates (a, b) are not involved and only the light distribution value at that location is referred to, it is abbreviated as y. m,nM and N are the total number of rows and columns in the light distribution image data, such as... Figure 3 As shown.

[0017] As another technical solution, the relative movement between the first light source under test and the diffuse screen is achieved by moving the diffuse screen to change the range of the light distribution image area generated by the first light source under test on the diffuse screen.

[0018] Furthermore, the process of obtaining the light distribution data emitted by the first light source through relative motion scanning measurement between the stray light elimination photometer and the first light source under test in step S1 includes: using one or more stray light elimination photometers to directly receive the light beam emitted by the first light source under test, causing relative motion between the stray light elimination photometer and the first light source under test, and the stray light elimination photometer scanning and measuring the light distribution data of the first light source under test as a function of spatial angle or spatial position. This light distribution data is used as standard light distribution data, denoted by y. s (a s ,b s ) indicates that, where the subscript s represents the standard value, and y s This represents the standard value of luminance or radiance of interest. The symbol 'y' here is a general term for the symbols of luminance or radiance of interest, specifically expressed as luminous intensity I, illuminance E, etc. (a s ,b s () represents the spatial angle of the emission from the first measured light source during scanning measurement using a stray light elimination photometer. Or in the spatial position (x) on the illuminated surface s ,y s To facilitate correspondence with data obtained from imaging measurement equipment, y is used as... s (a m,n ,b m,n ) represents the standard light distribution data corresponding to the light distribution image data in row m and column n.

[0019] Specifically, a stray light-eliminating photometer is used to receive the light beam emitted by the first light source under test, and the photometric or radiometric standard value y of interest for the first light source under test is measured. s At this time, the spatial angle or spatial position corresponding to the center of the detector receiving surface (a) s ,b s The light distribution data y of the first light source under test and the stray light elimination photometer can be calculated based on their poses. Through the relative motion between the stray light elimination photometer and the first light source under test, the poses of the photometer and / or the first light source under test change to achieve different spatial angles or positions, ultimately obtaining the light distribution data y of the first light source under test as a function of spatial angle or position. s (a s ,b sThe standard data for measuring light distribution using the stray light elimination photometer and the image data for measuring light distribution using the imaging measurement equipment can be obtained in the same or different measurement systems.

[0020] As one technical solution, the relative movement between the stray light elimination photometer and the first light source under test is achieved by fixing the stray light elimination photometer, mounting the first light source under test on the sample stage, and moving the first light source under test through the sample stage. The movement can be rotation or / and translation. Preferably, the stray light elimination photometer is used to scan and measure the first light source under test with respect to spatial angles. When the light distribution data changes, the reference center C of the first tested light source coincides with the rotation center of the sample stage, and the direction of the emitting surface (normal) of the first tested light source is... When the initial position of the sample stage rotation is taken as the starting position, and the normal of the emitting surface of the first measured light source is aligned with the center of the receiving surface of the photometer detector, then... It can be obtained directly from the rotation angle (ε,η) of the sample stage.

[0021] As another technical solution, the stray light elimination photometer moves relative to the first light source under test. This is achieved by keeping the first light source under test stationary while the stray light elimination photometer moves on the equivalent plane of the diffuse screen. The equivalent plane of the diffuse screen refers to a plane in which the distance from the first light source under test to this plane, its illumination posture on the plane, and the distance and illumination posture between the light source and the diffuse screen are consistent. The movement of the photometer on the equivalent plane of the diffuse screen is equivalent to directly measuring the standard light distribution data at different points on the diffuse screen.

[0022] As another technical solution, the stray light elimination photometer moves relative to the first light source under test, which is achieved by the sample stage driving the first light source under test to move and the photometer moving on the equivalent plane of the diffuse screen.

[0023] As a supplementary explanation, during the relative motion scanning measurement process between the stray light elimination photodetector and the first light source under test, different sampling step sizes can be used in different spatial regions. The scanning angle interval for different light distribution measurement regions is adjusted differentially according to the type of light source and / or its known light distribution characteristics. For example, for automotive headlight samples, the scanning angle interval is reduced in the transition region between light and dark cutoff lines to obtain more refined light distribution data; for light source samples with relatively uniform light distribution characteristics or relatively smooth transition angle regions in the measured light distribution image, the scanning angle interval is appropriately increased to reduce scanning time without affecting accuracy.

[0024] The aforementioned relative motion measurement can also be performed by replacing the first light source with the second light source.

[0025] Furthermore, based on the light distribution data measured by the stray light metering detector, new light distribution data is obtained using interpolation methods, and this new light distribution data is used as the standard light distribution data. The interpolation methods include, but are not limited to, linear interpolation or spline interpolation. The light distribution data measured using the stray light metering detector is usually discrete and finite in terms of spatial angle or spatial location, and may not cover the measurement data corresponding to the imaging measurement equipment in some spatial angles or locations. When measurement sampling is limited, interpolation methods can be used to further obtain light distribution data for other unmeasured spatial angles or locations.

[0026] Linear interpolation treats the data in the transition region between measured data points as a linear change to obtain the data corresponding to each point in the transition region. For interpolation of two-dimensional distributed data, bilinear interpolation is often used, which performs linear interpolation in two dimensional directions (in this invention, the two spatial coordinate directions of light distribution, a and b).

[0027] Spline interpolation is an interpolation method that fits discrete data points using piecewise low-order polynomials (called "spline functions"). It ensures that the interpolation result passes through all known points while achieving a smooth transition between adjacent segments. For spline function interpolation applications on two-dimensional light distribution data, bicubic spline interpolation is often used. This involves fitting a bivariate cubic polynomial to each rectangular region of the two-dimensional grid, while ensuring the continuity of the derivatives in both directions a and b. Bicubic spline interpolation yields relatively smooth interpolated data, which better matches the natural law of continuous variation in light distribution.

[0028] Furthermore, in step S1, the training sample dataset is expanded by changing any one or at least two combinations of different light source samples, different operating modes of the same light source, different alignment positions of the same light source relative to the diffuse screen, and different angles of the same light source relative to the diffuse screen. The different alignment positions and angles of the same light source relative to the diffuse screen are achieved by the sample stage moving the light source to change its pose. By collecting a large amount of different types of light distribution image data and standard light distribution data, a large amount of measured and effective training sample dataset can be provided for training the light distribution correction model. Preferably, the collected sample data is selected according to the actual measurement application requirements of light distribution correction. For example, if the type of the light source being measured is clearly defined, sample data collection of the same type of light source sample is added; if the light distribution measurement has clearly defined standard test conditions, sample data collection under those standard test conditions is added.

[0029] Furthermore, in step S1, the light distribution image data of the first tested light source undergoes data augmentation processing to generate a series of derived image data, which are then incorporated into the training sample dataset. The data augmentation processing methods include any one or a combination of at least two of blurring, noise addition, and distortion. The series of derived images obtained from the light distribution image data augmentation correspond to the same standard light distribution data as the original light distribution image data. By performing data augmentation processing on the light distribution image data, the training sample dataset is expanded, enhancing robustness.

[0030] The blurring process is achieved through convolution operations, with Gaussian blur being a common choice. The light distribution image data after Gaussian blurring is calculated as shown in formula (1):

[0031]

[0032] In the formula, y m,n G(u,v) represents the light distribution image data in the m-th row and n-th column; G(u,v) is the Gaussian kernel function, calculated by formula (2); u,v is the pixel coordinate offset with the center of the blur kernel as the origin; k is the side length of the Gaussian blur kernel.

[0033]

[0034] In the formula, σ is the standard deviation of the Gaussian kernel, which controls the degree of ambiguity. The larger the value of σ, the wider the ambiguity range.

[0035] The added noise refers to the addition of noise to the light distribution value y of the original light distribution image data. m,n Add a noise term ε to generate a new light distribution value y' m,n As shown in formula (3).

[0036] y' m,n =y m,n +ε (3)

[0037] Taking Gaussian noise, a common noise type, as an example, the added noise term ε has a mean of μ and a variance of σ. 2 Gaussian distribution N(μ,σ) 2 ) is obtained by random sampling, ε~N(μ,σ 2 ).

[0038] The distortion refers to the transformation of the original light distribution image data in terms of geometric distribution or optical properties to generate new light distribution image data, including but not limited to affine transformation and gamma transformation. Affine transformation is any combination of basic transformations such as translation, rotation, scaling and shearing. Gamma transformation is shown in formula (4).

[0039]

[0040] In the formula, when γ < 1, the dark areas are brightened, and when γ > 1, the bright areas are darkened. This formula can be used to simulate the nonlinear response of the detector.

[0041] In step S2, the light distribution correction model is built based on a convolutional neural network so that the light distribution measurement correction model can learn and fit the complex mapping process for correcting image data with different light distributions.

[0042] A neural network is a widely parallel interconnected network composed of adaptive neurons, capable of simulating the interactive responses of biological nervous systems to real-world objects. A neural network generally consists of an input layer, at least one hidden layer, and an output layer. The input layer is the entry point for the neural network to interact with external data, receiving raw input data and passing it on. The hidden layer is the core area for feature extraction and data processing, and the output layer is the final output of the neural network. Neurons are connected by weights, the magnitude of which determines the degree to which the output of one neuron influences the input of another. After receiving input signals from other neurons, each neuron compares the received weighted sum with a threshold and processes it through an activation function to generate its output. Commonly used activation functions include the Sigmoid function (Sigmoid(x) = 1 / [1 + e^(-x))]) and the ReLU function (ReLU(x) = max(0,x)). The learning process of a neural network involves adjusting parameters such as weights between neurons (represented by k in this invention) using training data to minimize the error between the predicted and actual results, thus better simulating the ideal functional response. Generally, a quantification function, namely a loss function, is used to quantify and measure this error between the model's prediction and the actual result.

[0043] Convolutional neural networks (CNNs) are a type of neural network suitable for processing two-dimensional data objects such as images. They typically include convolutional layers and pooling layers. Convolutional layers are used to extract local features from the data, while pooling layers are used to downsample the features. Downsampling refers to the process of reducing the spatial size of the feature map while preserving image features. By using convolution and pooling operations, image data features can be effectively extracted, reducing the amount of data while retaining useful information.

[0044] As a technical solution, in step S2, the light distribution correction model is built based on a convolutional neural network. The convolutional neural network includes an encoder module and a decoder module. The encoder module includes several downsampling groups composed of convolutional layers and pooling layers, used to extract image features from the input light distribution image data. The decoder module includes several upsampling groups composed of upsampling layers and convolutional layers. The decoder module is used to process and restore the image features, outputting predicted light distribution correction data. The number of upsampling groups and downsampling groups is the same, and there are skip connections between them. (See schematic diagram below.)Figure 4 As shown. In contrast to downsampling, which compresses spatial dimensions to extract abstract features, upsampling refers to the process of reconstructing higher spatial resolution data based on abstract features. Skip connections are a type of cross-layer connection that directly connects the output of one layer in a deep learning network to subsequent layers. This stabilizes the optimization process of the entire network, allowing the model to simultaneously integrate and utilize multi-scale information, thereby improving task accuracy. In the technical solution of this invention, through skip connections, the results obtained by upsampling in the decoder module are concatenated with the features obtained during downsampling in the encoder module. This preserves richer feature information from the light distribution image data and optimizes the learning results of light distribution correction.

[0045] As another technical solution, in step S2, the light distribution correction model is built based on a convolutional neural network. The convolutional neural network includes a basic feature extraction module, a channel feature extraction module, and a reconstruction module. The basic feature extraction module includes a convolutional layer for extracting shallow image features. The channel feature extraction module includes several channel attention residual groups, a convolutional layer, and a residual connection. Each channel attention residual group contains several channel attention units with residual connections, a convolutional layer, and a residual connection, used to extract deep feature information from different channels of the image and adjust the feature weights of different channels. The reconstruction module includes a convolutional layer for integrating the feature information from each channel to output predicted light distribution correction data. (See schematic diagram below.) Figure 5 As shown.

[0046] In this technical solution, the use of residual connections can improve training efficiency and stability. Meanwhile, in the light distribution correction model, the extracted channel features may have varying importance for light distribution correction and do not need to be treated equally. The channel attention unit structure is as follows: Figure 6 As shown, through training, the model adaptively evaluates the differences in importance of each channel in the task, strengthens the features of important channels, and suppresses unimportant channels.

[0047] Furthermore, in step S2, the loss function calculation method includes, but is not limited to, mean squared error (MSE) or structural similarity index (SSIM). Mean squared error (MSE) and structural similarity index (SSIM) are two commonly used metrics for measuring image similarity. Let f(·) represent the light distribution correction model, and k represent the parameters of the light distribution correction model. The light distribution correction model represents the input light distribution image data. For the corresponding output light distribution correction data, LOSS(k) represents the loss function when the model parameter is k. The objective of optimizing the model parameters is to minimize the loss function, as shown in formula (5). The optimized parameters are represented as k. * .

[0048] k * =argmin k LOSS(k) (5)

[0049] In the formula, the argmin function represents the value of the variable that minimizes the objective function.

[0050] MSE measures the mean squared error between the predicted light distribution correction data and the standard light distribution data, reflecting pixel-level differences. A smaller value indicates that the prediction is closer to the true value. The loss function based on MSE is L... MSE (k) is calculated as shown in formula (6).

[0051]

[0052] In the formula, The light distribution correction model represents the input light distribution image data. The corresponding output light distribution correction data, where m and n represent the light distribution image data respectively. The row and column numbers, M and N are the total number of rows and columns in the light distribution image data, y s (a m,n ,b m,n () is light distribution image data The corresponding standard data for light distribution.

[0053] SSIM comprehensively considers the similarity of image data in terms of intensity, contrast, and structure. It not only considers the overall visual perception of the images before and after light distribution correction, but is also sensitive to changes in local structure. The calculation method of SSIM is shown in formula (7). The larger the SSIM value, the better the correction effect, that is, the more similar the two image data are. The maximum value is 1. The loss function calculation based on SSIM is based on the negative sign of SSIM, as shown in formula (8). That is, the larger the SSIM value, the smaller the loss function L. SSIM (k).

[0054]

[0055] In the formula, μ1 and μ2 represent the predicted light distribution correction data, respectively. Compared with standard light distribution data y s (a s ,b s The mean value of )

[0056] and These represent the variances of the predicted corrected light distribution data and the standard light distribution data, respectively.

[0057] σ 12 The covariance between the predicted corrected light distribution data and the standard light distribution data;

[0058] C1 and C2 are stability constants, which are used to prevent the denominator from being zero.

[0059] When evaluating SSIM across the entire image span, the mean and variance may vary drastically. The image data can be divided into different window regions for structural similarity index calculation, and then the average value can be taken as the overall SSIM to obtain the mean structural similarity SSIM, abbreviated as MSSIM. The corresponding loss function is shown in formula (9).

[0060]

[0061] In the formula, R is the number of window regions into which the image data is split.

[0062] Furthermore, in step S2, when calculating the loss function, the weight coefficients for different spatial angles or locations in the light distribution image data are different, and the weighted mean square error or weighted structural similarity index is used as the loss function. In practice, for certain types of measurement applications, only some feature points or regions in the light distribution image may be the focus of correction. For example, for conventional lamps, edge regions may not be important, while for vehicle lights, several important points may be more important, such as HV points and feature points specified in regulations for comparing light intensity limits. For the key feature points or regions, if the correction effect is not good, the loss penalty needs to be larger, making them a key component of the loss function, with higher weights than other non-key feature points or regions, thereby more specifically improving the training effect of the light distribution correction model. Weighted mean square error (LOSS) w,MSE (k) or the weighted structural similarity index LOSS w,SSIM The loss function calculation for (k) is shown in formulas (10) and (11), respectively. The methods for obtaining the weight coefficients include, but are not limited to: manual pre-marking, generation based on light distribution thresholds or gradients. For example, for regions with clearly known key spatial angles or spatial locations of light distribution, fixed weight coefficients can be manually designed, such as the HV point during vehicle headlight measurement or feature points specified in regulations for comparing light intensity limits; alternatively, weight coefficients can be determined and assigned based on the threshold or gradient of the light distribution value, assigning higher weights to key areas of interest such as dark areas and transitional regions between light and dark.

[0063]

[0064] In the formula, Light distribution image data The corresponding weighting coefficients.

[0065]

[0066] In the formula, R is the number of window regions into which the image data is split, and w r The weight coefficients are the weights corresponding to the r-th window region after splitting.

[0067] Furthermore, the characteristic functions of the measurement system or imaging measurement device are input into the light distribution correction model. These characteristic functions include a point spread function and / or a linearity function, serving as prior physical knowledge to assist in training the light distribution correction model. The aforementioned prior physical knowledge refers to known information, constraints, or regular conclusions based on physical laws, which can guide the model, helping it learn more efficiently and avoid distorted correction results or overfitting problems caused by purely data-driven approaches.

[0068] The point spread function (PSF) of an imaging measurement device refers to the response function for an ideal point light source. That is, when the input is a point light source, the light distribution image acquired will be degraded due to the influence of aberrations, diffraction, detector noise and other factors of the imaging optical system. The actual output is an expanded light spot, and the shape and intensity distribution of the light spot are described by the point spread function.

[0069] The image measurement result of the imaging measurement equipment is the result of convolving the original theoretical image with the point spread function and adding the influence of stray light and other noise, which can be expressed by formula (12):

[0070] Y M =Y T *PSF+N (12)

[0071] In the formula, * represents the convolution operation, Y M Y represents the measured light distribution image data. T This represents the original theoretical light distribution image data before degradation, where N is the noise term including stray light.

[0072] The linearity function is used to measure the light distribution value output Y of a measurement system or imaging measurement device. out With input Y in The degree of linear relationship between them is denoted by L, Y out =L(Y inWe typically expect a system to have a good linear response, but in practical measurement applications, measurement systems or imaging measurement devices often exhibit certain nonlinear characteristics. Considering nonlinear effects, the actual measured light distribution image data Y... M Compared with the original theoretical image data Y T The relationship is expressed as in formula (13).

[0073] Y M =L(Y T *PSF)+N (13)

[0074] The aforementioned assisted training scheme specifically includes: a physical prior processing module included in the light distribution correction model structure, and / or the introduction of a physical consistency loss term based on feature functions into the loss function. For example... Figure 7 As shown.

[0075] The physical prior processing module performs physical restoration and initial correction processing on the light distribution image data based on the input physical prior knowledge, including the point spread function and / or linearity function. For feature functions that include a linearity function, the physical prior processing module uses L... -1 For the input light distribution image data Y M Nonlinear response correction is performed; for features containing a point spread function, the physical prior processing module performs deconvolution processing on the nonlinear response corrected (if any) light distribution image data based on the known PSF to obtain the original theoretical light distribution image data Y before degradation. T .

[0076] A physical consistency loss term based on the feature function is introduced into the loss function, as shown in Equation (14). By adding the physical constraint term, the conformity between the light distribution correction data predicted by the light distribution correction model and the known physical feature function is improved.

[0077] LOSS(k) = LOSS data (k)+λ·LOSS phy (k) (14)

[0078] In the formula, LOSS data (k) represents the data reconstruction loss, which is a loss function calculated based on MSE or SSIM as described above;

[0079] LOSS phy (k) is the physical consistency loss term, which can be calculated according to formula (15);

[0080] λ is the weighting parameter for balancing the loss terms.

[0081]

[0082] In the formula, * represents the convolution operation, Y M This represents the measured light distribution image data, where m and n represent the light distribution image data respectively. The row and column numbers, M and N are the total number of rows and columns in the light distribution image data, f k The predicted light distribution correction data represents the output of the light distribution correction model. For cases involving only the point spread function or only the linearity function, the element L(f) in equation (15) k *PSF) The selection should be modified according to the specific situation. For example, for cases that only include the point spread function, this element should be modified to f. k *PSF; For cases that only include the linearity function, this element is modified to L(f) k ).

[0083] Furthermore, before training the light distribution correction model using the training sample dataset, data preprocessing is performed to unify the data resolution between the light distribution image data and the standard light distribution data. The light distribution correction model is applied to each light distribution image data set. Output predicted light distribution correction data When calculating the loss function, it is necessary to match the corresponding standard light distribution data y. s (a s ,b s )| m,n .

[0084] Typically, the array detectors of imaging measurement equipment have a high number of pixels, resulting in higher spatial resolution measurement data compared to stray light elimination photometers with a fixed receiving aperture.

[0085] As a technical solution, interpolation or super-resolution reconstruction is performed on the light distribution standard data corresponding to the light distribution image data to make the data resolution of the light distribution image data and the processed light distribution standard data the same. The interpolation methods include, but are not limited to, linear, spline, and polynomial function fitting. The super-resolution reconstruction refers to using algorithms to reconstruct the original light distribution standard data into reconstructed light distribution standard data with the same resolution as the light distribution image data.

[0086] As another technical solution, before training the light distribution correction model using the training sample dataset, the light distribution image data is cropped, sampled, or averaged to obtain a light distribution image data table with the same data resolution as the standard light distribution data.

[0087] Furthermore, before training the light distribution correction model using the training sample dataset, the spatial angle or spatial position coordinates of the light distribution image data are adjusted to correct the alignment error between the light distribution image data and the standard light distribution data. Due to the influence of systematic errors, operational factors, etc., for the same measurement beam, the spatial coordinates corresponding to the light distribution data recorded using both imaging measurement equipment and stray light metering detectors may deviate. For light source samples or measured areas with drastic changes in light distribution, even a small spatial coordinate deviation can lead to significant differences in the measured values, which adversely affects the learning process of the light distribution correction model. Alignment error refers to the light distribution data error caused by spatial coordinate deviation.

[0088] As a technical solution, preset spatial angle or spatial position coordinate adjustment options for spatial light distribution are provided. The alignment error between the light distribution image data after the spatial angle or spatial position coordinate adjustment and the corresponding light distribution standard data is calculated. The coordinate adjustment option with the smallest alignment error is selected, and the coordinate adjustment is applied to the light distribution image data. The alignment error calculation method includes, but is not limited to, mean square error and average absolute error. This invention uses the light distribution spatial coordinates (a...) in the light distribution standard data... s ,b s Based on this, the spatial coordinates of the light distribution in the light distribution image data are adjusted. Two or more coordinate adjustment options (Δa1, Δb1), (Δa2, Δb2), ..., (Δa...) are preset, either based on experience or directly. p ,Δb p (p is the number of coordinate adjustment options, p≥2), calculate the coordinate-adjusted light distribution image data. Corresponding standard light distribution data Alignment error E between q And select the coordinate adjustment pair with the smallest alignment error (Δa) w ,Δb w (w∈{1,2,……,p}), adjust the spatial coordinates of the light distribution image data from the original (a,b) to (a+Δa) w ,b+Δb w ).

[0089] As another technical solution, a pre-trained image registration module is introduced. The image registration module includes an image feature extraction function group and a registration function group. The image feature extraction function group is used to extract image features from the light distribution image data and the light distribution standard data. The registration function group is used to calculate a matrix for performing coordinate transformations such as translation and rotation on the light distribution image data based on the extracted image features, and to perform coordinate transformations on the light distribution image data based on the matrix.

[0090] For practical applications with a clearly defined purpose of light distribution measurement correction, a light distribution correction model can be trained by collecting or filtering training sample datasets, allowing the model to focus on learning specific aspects of light distribution correction effects. Stray light is a significant factor affecting the accuracy of imaging measurement light distribution data. As a technical solution, the training sample dataset includes several sets of light distribution dark area image data, strong contrast image data, and standard light distribution data of light sources. The training sample dataset is used to train a light distribution correction model for stray light error correction.

[0091] In the technical solution of this invention, step S3 involves using a light distribution correction model to correct the light distribution image data of the second light source under test (i.e., the actual measured light source). Specifically, in the same or different measurement system as step S1, an imaging measurement device is used to measure and obtain the light distribution image data of the light emitted by the second light source under test illuminating the diffuse screen. The light distribution image data of the second light source under test is then input into the optimized light distribution correction model to obtain the corrected light distribution data. As a supplementary explanation, the measurement system in the technical solution of this invention refers to the overall measurement environment including the laboratory and the measurement equipment. The measurement equipment includes an imaging measurement device, a diffuse screen, a sample stage, etc.

[0092] Furthermore, in step S3, the spatial light distribution image of the second measured light source is processed by region. Only a portion of the image data is corrected using the light distribution correction model, while other regions are not corrected. The regions for light distribution correction include dark areas with low luminance values ​​or areas with significant brightness contrast. In practical measurement applications, it is common for only some areas with severe stray light (such as dark areas with low luminance values ​​or areas with significant brightness contrast) to require correction. Performing light distribution correction only on these areas helps improve measurement correction efficiency.

[0093] Furthermore, after step S3, step S4 is also included: further using a stray light photometer detector to perform relative motion scanning measurement with the second light source under test in step S3 to obtain standard light distribution data of the second light source under test, and supplementing it to the training sample dataset together with the light distribution image data of the second light source under test; using the supplemented training sample dataset to iteratively train the model in S2 to obtain the iterative light distribution correction model; and inputting the light distribution image data of the second light source under test into the iterative light distribution correction model.

[0094] The deep learning-based light distribution correction model in this invention possesses self-training and iterative update capabilities. This self-training refers to the ability to retrain the model through the cumulative iteration of measurement data, continuously improving its performance for specific measurement systems or light sources. The light distribution correction model obtained through steps S1 and S2 is relatively basic and general, and can be considered as pre-training of the light distribution image data and its corresponding physical correction effects in the training sample dataset. For measurement systems and / or objects that differ from those acquired during data acquisition and model training, supplementary and updated light distribution image data and standard light distribution data are added to the training sample dataset. The model in S2 is then iteratively trained using this supplementary dataset to obtain an iteratively trained light distribution correction model. This allows the light distribution correction model to be fine-tuned for specific measurement systems and light sources, making it more suitable for the current light distribution correction task.

[0095] The present invention also provides a light distribution measurement and correction device based on the light distribution measurement and correction method, comprising: a sample stage (2), a diffuse screen (3), an imaging measurement device (4), and a control processing module (6); the diffuse screen (3) is set at a certain distance from the light source (1), the light emitted by the light source (1) illuminates the diffuse screen (3) to generate a light spot image, the imaging measurement device (4) is aligned with the diffuse screen (3) and obtains the light distribution image, the control processing module (6) is communicatively connected to the imaging measurement device (4), the control processing module (6) is used to acquire the light distribution image data obtained by the imaging measurement device (4), the control processing module (6) further includes a light distribution correction processing unit (61) based on deep learning, the light distribution correction processing unit (61) is used to calculate and output the light distribution data of the light source according to the light distribution image data. The light source is a first light source under test or a second light source under test.

[0096] As a technical solution, the light distribution measurement and correction device further includes one or more stray light elimination photometers (5). The stray light elimination photometer (5) includes a photometric detector, a radiometric detector, or a spectroradiometer and its sampling device, used to scan and measure the light distribution data emitted by the light source, and to use this light distribution data as the light distribution standard data. The control processing module (6) also includes a self-training standard data unit (62), which is communicatively connected to the stray light elimination photometer (5). The self-training standard data unit (62) is used to receive the light distribution standard data acquired by the stray light elimination photometer for the self-training of the light distribution correction processing unit (61). Preferably, the stray light elimination photometer is equipped with an extinction tube to minimize stray light interference. When the stray light elimination photometer is a photometric detector, the photometric detector is precisely matched to the V(λ) curve.

[0097] As a technical solution, the sample stage (2) includes a rotation mechanism and / or a translation mechanism. Driven by the sample stage (2), the light source (1) and the diffuse screen (3) move relative to each other. The sample stage (2) is connected to the control and processing module (6) for communication.

[0098] As a technical solution, a stray light elimination aperture (7) is set between the light source and the diffuser screen. Using the stray light elimination aperture blocks a large number of light beams from areas outside the measurement angle, reducing stray light interference generated by reflected light on the diffuser screen and inside the imaging measurement device, improving the quality of the acquired light distribution image data, and further enhancing the learning efficiency of the light distribution correction model. In addition, by adjusting the number and placement of the stray light elimination apertures, light distribution image data with different degrees of stray light interference can be acquired, enriching the training sample database and further used for self-training of the deep learning-based light distribution correction processing unit.

[0099] The beneficial effects of this invention are as follows: The technical solution of this invention adopts the method of constructing and training a light distribution correction model. By introducing high-precision standard light distribution data obtained from actual measurements by a stray light metering detector, the light distribution correction model can perform deep learning to fit the correction process of light distribution image data. After training, it does not need to rely on various fine numerical information of physical properties of various test environments. It can directly output light distribution correction data for input light distribution image data. It is widely applicable to various types of light source objects under test and test environments, realizes fast and adaptive correction processing, and effectively improves measurement accuracy and efficiency. Attached Figure Description

[0100] Appendix Figure 1 This is a schematic diagram illustrating the principle of the optical distribution measurement and correction technology of the present invention;

[0101] Appendix Figure 2 This is a schematic diagram of the light distribution image data mosaic integration of the present invention;

[0102] Appendix Figure 3 This is a schematic diagram of the light distribution image data table of the present invention;

[0103] Appendix Figure 4 This is a schematic diagram of a light distribution correction model structure built on a convolutional neural network;

[0104] Appendix Figure 5 A schematic diagram of another optical distribution correction model structure based on a convolutional neural network;

[0105] Appendix Figure 6 This is a schematic diagram of the channel attention unit.

[0106] Appendix Figure 7 A schematic diagram illustrating the use of characteristic functions as prior physical knowledge to assist in training a light distribution correction model;

[0107] Appendix Figure 8 This is a schematic diagram of a light distribution measurement and correction device provided in Embodiment 1 of the present invention;

[0108] Appendix Figure 9 This is a schematic diagram of a light distribution correction model structure provided in Embodiment 1 of the present invention;

[0109] Appendix Figure 10 This is a schematic diagram of a light distribution measurement and correction device provided in Embodiment 2 of the present invention;

[0110] Appendix Figure 11 This is a schematic diagram of a light distribution measurement and correction device provided in Embodiment 3 of the present invention.

[0111] In the figure, 1-light source, 2-sample stage, 3-diffusing screen, 4-imaging measurement device, 5-stray light elimination photodetector, 6-control processing module, 61-deep learning-based light distribution correction processing unit, 62-self-trained standard data unit, 7-stray light elimination aperture. Detailed Implementation

[0112] The specific embodiments of the present invention are described below with reference to the accompanying drawings. However, those skilled in the art should understand that the following embodiments are for illustrative purposes only and are not intended to limit the scope of the present invention. Those skilled in the art should understand that modifications can be made to the following embodiments without departing from the scope and spirit of the present invention. The scope of protection of the present invention is defined by the appended claims.

[0113] Example 1:

[0114] This embodiment discloses a light distribution measurement and correction method and apparatus, specifically relating to the measurement and correction of luminous intensity distribution. Image data is measured and corrected.

[0115] A schematic diagram of a light distribution measurement and correction device disclosed in this embodiment is shown below. Figure 8 As shown, the system includes a sample stage 2, a diffused screen 3 positioned opposite the sample stage 2, an imaging measurement device 4, a stray light elimination photometer 5, and a control and processing module 6. The diffused screen 3 and the stray light elimination photometer 5 are positioned at a certain distance relative to the light source. A stray light elimination aperture 7 is positioned between the sample stage 2 and the diffused screen 3. The aperture of the stray light elimination aperture 7 is slightly larger than the angle between the line connecting the edge of the largest measurable luminous surface and the edge of the measurement area of ​​the diffused screen 3.

[0116] Sample stage 2 includes a rotation mechanism and a translation mechanism, which can drive the light source to rotate and translate around the horizontal and vertical axes. The rotation center of the sample stage is denoted as point O. The horizontal rotation axis of sample stage 2 coincides with the center normal of the diffuse screen 3, and the rotation center O of sample stage 2 coincides with the center O of the diffuse screen. MThe distance between them is known. In this embodiment, the imaging measurement device 4 is set in the space between the diffuse screen 3 and the sample stage 2, and the imaging measurement device 4 is aligned with the diffuse screen 3, always remaining fixed, and its spatial coordinates relative to the rotation center O of the sample 2 are known.

[0117] The stray light elimination photometer 5 is located centered below the edge of the diffuse screen. It is a photometric detector with an extinction tube. After precise V(λ) curve matching and calibration by placing it at the rotation center O of the sample stage via the reference center of the light intensity standard lamp, it is used to measure the luminous intensity.

[0118] The control processing module 6 includes a deep learning-based light distribution correction processing unit 61, which is communicatively connected to the imaging measurement device 4. The light distribution correction processing unit receives light distribution image data acquired by the imaging measurement device and outputs corrected light distribution data. The control processing module 6 also includes a self-training standard data unit 62, which is communicatively connected to the stray light metering detector 5 and receives light distribution standard data acquired by the stray light metering detector for self-training of the light distribution correction processing unit.

[0119] In this device, the imaging measurement equipment 4 has been calibrated and can directly measure the light intensity distribution image data. The correspondence between the coordinates of each pixel and the coordinates (x, y) of each point on the diffuse screen has also been calibrated; and the point spread function (PSF) is known. The light source 1 installed on the sample stage 2 can be either the first light source to be measured or the second light source to be measured. The first light source to be measured refers to the light source used to construct the training sample dataset and establish the light distribution correction model; it can be a standard light source or a non-standard light source. The second light source to be measured refers to various types of light source objects, which are the actual light sources that need to be measured. The acquisition of the training sample dataset and the measurement and calibration of the light source are both performed in this device. The above device is used to implement spatial light intensity distribution. The specific steps for image data measurement and correction are as follows:

[0120] A1: Acquisition of light distribution image data samples:

[0121] A1a: In the light distribution measurement and correction device, the first light source under test is mounted on the sample stage and lit. The light emitted by the first light source under test illuminates the diffuse screen to produce a light spot image. The imaging measurement device is aligned with the diffuse screen, and the light source is positioned at position 1. Below, the light intensity distribution image data of the first measured light source is obtained. (x C1 ,y C1 ,z C1 () indicates the coordinate position of the reference center C of the first measured light source in pose 1. This represents the deviation of the direction of the emitting surface (normal) of the first measured light source from the reference direction at pose 1. The emission angle of the first light source under test corresponds to the data in the m-th row and n-th column of the light intensity distribution image data I1. This is based on the coordinates of the diffuse screen points corresponding to the pixels of the array detector and the position coordinates (x, y) of the reference center C of the first light source under test. C1 ,y C1 ,z C1 ) and the attitude of the first measured light source Calculated;

[0122] A1b: Adjusting the sample stage moves the first light source under test to change its pose. The i-th pose of the first light source under test is denoted as... Different poses Repeat the measurement operation of the imaging measurement device in A1a to obtain the light intensity distribution image data I of the first light source under test in different poses. i ;

[0123] A1c: Change the position of the aperture and repeat the light intensity distribution image data measurement operation in A1a-A1b to obtain the light intensity distribution image data of the first light source under test under different degrees of stray light interference.

[0124] A1d: Change the working mode of the first light source under test, repeat the light intensity distribution image data measurement operation in A1a-A1c, and obtain the light intensity distribution image data of the first light source under test under different working modes;

[0125] A1e: Change the light source sample and repeat the light intensity distribution image data measurement operation in A1a-A1c to obtain light intensity distribution image data of different light source samples;

[0126] A1f: All light intensity distribution image data samples obtained previously The data is input and stored in the light distribution correction processing unit based on deep learning. The distribution characteristics of the collected light intensity distribution image data samples are initially recognized and analyzed to divide the angular regions with different light distribution gradient change characteristics.

[0127] A2: Light distribution image data enhancement:

[0128] Sample of measured light intensity distribution image data (with I m,n The light intensity distribution image data in the m-th row and n-th column is used to generate a series of derived light intensity distribution image data I' using Gaussian blur and gamma transform, as shown in formulas (16) and (18), and supplemented into the light intensity distribution image data sample.

[0129]

[0130] In the formula, I m,nG(u,v) represents the light intensity distribution image data in the m-th row and n-th column; G(u,v) is the Gaussian kernel function, calculated by formula (18); u,v is the pixel coordinate offset with the center of the blur kernel as the origin; k is the side length of the Gaussian blur kernel.

[0131]

[0132] In the formula, σ is the standard deviation of the Gaussian kernel.

[0133]

[0134] In the formula, when γ < 1, the dark areas are brightened, and when γ > 1, the bright areas are darkened, which is used to simulate the nonlinear response of the detector.

[0135] A3: Standard data acquisition for light distribution correction:

[0136] A3a: In the light distribution measurement and correction device, the first light source under test is mounted on the sample stage, and the reference center C of the first light source under test coincides with the rotation center O of the sample stage. The light source is lit, and the rotation angle of the sample stage is adjusted so that the center of the first light source under test and the direction of the emitting surface (normal) are aligned with the center of the receiving surface of the stray light elimination photometer, which is taken as the initial rotation position of the sample stage. The rotation of the sample stage drives the first light source under test to rotate, and the rotation angle (ε, η) of the sample stage is taken as the spatial angular direction of the emitted beam of the first light source under test. The luminous intensity value of the first light source under test in this direction is obtained by scanning with a stray light-eliminating photometer. In sampling regions where the light distribution gradient changes drastically, use smaller sample stage rotation angle intervals;

[0137] A3b: For each operating mode of each light source sample, repeat the above operation in A3a and use a stray light elimination photometer to measure and obtain the standard data of spatial light intensity distribution for each light source sample under each operating mode. The data is input and stored in the self-trained standard data unit.

[0138] A4: Construction of the training sample dataset:

[0139] For each set of light intensity distribution image data, the spatial angle intervals in the standard light intensity distribution data are used as a benchmark to average the light intensity distribution image data, resulting in a light intensity distribution image data table with the same data resolution as the standard light intensity distribution data; multiple spatial angle coordinate adjustment options for the light intensity image data are set based on empirical values. Based on the mean square error, the alignment error E between the coordinate-adjusted light intensity distribution image data and the standard light intensity distribution data under each option is calculated according to formula (19). q Select the coordinate adjustment option with the smallest alignment error. The coordinates of the light intensity distribution image data are adjusted as shown in formula (20); several sets of light intensity distribution image data and corresponding standard light intensity distribution data are obtained, and a training sample dataset is constructed.

[0140]

[0141] In the formula, m and n represent the row and column numbers of the light intensity distribution image data, respectively, and M and N are the total number of rows and columns in the light intensity distribution image data. This is the light intensity distribution image data in the m-th row and n-th column under the q-th coordinate adjustment option, corresponding to the luminous intensity angle direction. Is with The corresponding standard data for light intensity distribution.

[0142]

[0143] In the formula, This represents the light intensity distribution image data at row m and column n after adjusting the spatial coordinates. Compared with the original light intensity distribution image data, the light intensity distribution values ​​remain unchanged, but the corresponding emission angle changes from... Become

[0144] A5: Establishment of a deep learning-based light distribution correction model:

[0145] In the deep learning-based light distribution correction processing unit, a light distribution correction model is built based on a convolutional neural network, including a physical prior processing module, an encoder module, and a decoder module, such as... Figure 9 As shown, the physical prior processing module performs deconvolution processing on the light intensity distribution image data based on the known PSF to obtain the original theoretical light intensity distribution image data before degradation. The encoder module is used to extract light intensity distribution image features. The light intensity distribution image features extracted by the encoder module are input into the decoder. The decoder module is used to calculate and restore the light intensity distribution image features and output the predicted light intensity distribution correction data. The internal structure of the encoder module and the decoder module is as follows: Figure 4 As shown.

[0146] A6: Training of a Deep Learning-Based Light Distribution Correction Model:

[0147] Light intensity distribution image data from the training sample dataset Input light distribution correction model, light distribution correction model f k Output the predicted light intensity distribution correction data under parameter k. The loss function LOSS(k) is calculated as shown in formula (21):

[0148]

[0149] In the formula, m and n represent the light intensity distribution image data, respectively. The row and column numbers, M and N, represent the total number of rows and columns in the light intensity distribution image data. It is light intensity distribution image data The corresponding standard data for light intensity distribution.

[0150] The model parameters k are optimized by minimizing the calculated loss function value. This completes the supervised learning of the model's correction process for light intensity distribution image data, resulting in the optimized light distribution correction model with parameters k. * See formula (5).

[0151] A7: Measurement and correction of measured light intensity distribution image:

[0152] In the light distribution measurement and correction device, the second light source under test is mounted on the sample stage, and the light intensity distribution image data of the light emitted by the second light source under test illuminating the diffuse screen is obtained by using an imaging measurement device. The light distribution correction processing unit based on deep learning receives the light intensity distribution image data obtained by the imaging measurement device and outputs the corrected light intensity distribution data of the second light source under test.

[0153] A8: Deep Learning-Based Self-Training for Light Distribution Correction

[0154] In the light distribution measurement and correction device, the standard data of light intensity distribution emitted by the second light source under test in step A8 is further obtained by scanning and measuring using a stray light photometer detector. This data, together with the light intensity distribution image data of the second light source under test in step A8, is added to the training sample dataset. Step A6 is repeated using the supplemented training sample dataset to iteratively train the model and obtain the iterative light distribution correction model.

[0155] Example 2:

[0156] This embodiment discloses a light distribution measurement and correction method and apparatus, specifically involving the measurement and correction of light distribution E(x,y) image data.

[0157] A schematic diagram of a light distribution measurement and correction device disclosed in this embodiment is shown below. Figure 10 As shown, the system includes a sample stage 2, a movable diffuser screen 3 positioned opposite the sample stage 2, an imaging measurement device 4, a stray light elimination photometer 5, and a control and processing module 6. The diffuser screen 3 and the stray light elimination photometer 5 are positioned at a distance from the light source 1. A stray light elimination aperture 7 is positioned between the sample stage 2 and the diffuser screen 3. The aperture of the stray light elimination aperture 7 is slightly larger than the angle between the line connecting the edge of the largest measurable luminous surface and the edge of the measurement area of ​​the diffuser screen 3.

[0158] Sample stage 2 includes a rotation mechanism and a translation mechanism, which can drive the light source to rotate and translate around the horizontal and vertical axes. The rotation center of the sample stage is denoted as point O. The horizontal rotation axis of sample stage 2 coincides with the center normal of the diffuse screen 3, and the rotation center O of sample stage 2 coincides with the center O of the diffuse screen. M The distance between them is known. In this embodiment, the imaging measurement device 4 is set in the space between the diffuse screen 3 and the sample stage 2, and the imaging measurement device 4 is aligned with the diffuse screen 3, always remaining fixed, and its spatial coordinates relative to the rotation center O of the sample 2 are known.

[0159] The stray light elimination photometer 5 is a photometric detector with an extinction tube. After precise V(λ) curve matching, it is used to measure illuminance values. It can move on a plane (denoted as plane p) parallel to the diffuse screen behind it. The distance Δd between the position of the diffuse screen during measurement and plane p is known. By adjusting the light source pose using a turntable, the difference in positional coordinates between plane p and the diffuse screen in the direction perpendicular to the diffuse screen is compensated. This plane p can be used as the equivalent plane of the diffuse screen. The center of the detector's light-receiving surface (denoted as O) d The correspondence between the position of a point moving on the equivalent plane of the diffuse screen and its position coordinates on the diffuse screen is known.

[0160] The control processing module 6 includes a deep learning-based light distribution correction processing unit 61, which is communicatively connected to the imaging measurement device 4. The light distribution correction processing unit receives light distribution image data acquired by the imaging measurement device and outputs corrected light distribution data. The control processing module 6 also includes a self-training standard data unit 62, which is communicatively connected to the stray light metering detector 5 and receives light distribution standard data acquired by the stray light metering detector for self-training of the light distribution correction processing unit.

[0161] In this device, the imaging measurement device 4 has been calibrated and can directly measure the illuminance distribution image data. The correspondence between the coordinates of each pixel and the coordinates (x, y) of each point on the diffuse screen has also been calibrated. The linearity function L of the imaging measurement device 4 is known, and the output illuminance measurement value is related to the theoretically input true illuminance signal: E out =L(E in The light source installed on sample stage 2 can be either a first test light source or a second test light source. The first test light source refers to the light source used to construct the training sample dataset, which can be a standard light source or a non-standard light source. The second test light source refers to various types of test light source objects, which are the actual light sources that need to be measured. The acquisition of the training sample dataset and the measurement and calibration of the light source are both performed within this device. The specific steps for implementing the measurement and calibration of the illuminance distribution E(x,y) image data using the above device are as follows:

[0162] B1: Acquisition of light distribution image data samples:

[0163] B1a: In the light distribution measurement and correction device, the first light source to be measured is mounted on the sample stage, the light source is lit, and the light emitted by the light source illuminates the diffuse screen to produce a light spot image. The imaging measurement device is aligned with the diffuse screen, and the position and pose conditions of the light source are determined. Below, the illuminance distribution image data of the first measured light source is obtained. (x C1 ,y C1 ,z C1 () indicates the coordinate position of the reference center C of the first measured light source in pose 1. This represents the deviation of the direction of the emitting surface (normal) of the first measured light source from the reference direction at pose 1. The coordinate position of the illuminated surface corresponding to the m-th row and n-th column in the illumination distribution image data E1 is obtained by the correspondence between the pixel of the array detector and the point coordinate of the diffuse screen.

[0164] B1b: Adjusting the sample stage moves the first light source under test to change its pose. The i-th pose of the first light source under test is denoted as... Different poses Repeat the measurement operation of the imaging measurement device in B1a to obtain illuminance distribution image data E on the diffuse screen under different poses of the first light source under test. i ;

[0165] B1c: Change the position of the aperture and repeat the illuminance distribution image data measurement operation in B1a-B1b to obtain the illuminance distribution image data of the first light source under test with different degrees of stray light interference;

[0166] B1d: Change the working mode of the first light source under test, repeat the illuminance distribution image data measurement operation in B1a-B1c, and obtain the illuminance distribution image data of the first light source under test in different working modes;

[0167] B1e: Change the light source sample and repeat the illuminance distribution image data measurement operations in B1a-B1d to obtain illuminance distribution image data for different light source samples;

[0168] B1f: Samples of all the illumination distribution image data collected above The data is input and stored in the deep learning-based light distribution correction processing unit.

[0169] B2: Light distribution image data enhancement:

[0170] Sample image data of measured illuminance distribution in B1 (E) m,n(Representing the illuminance distribution image data in the m-th row and n-th column), Gaussian noise is added to generate a series of derived illuminance distribution image data E', as shown in formula (22), and these are supplemented into the illuminance distribution image data sample set.

[0171] E' m,n =E m,n +ε (22)

[0172] In the formula, the added noise term ε has a mean of μ and a variance of σ. 2 Gaussian distribution N(μ,σ) 2 ) is obtained by random sampling, ε~N(μ,σ 2 ).

[0173] B3: Acquisition of standard data for light distribution correction:

[0174] In the light distribution measurement and correction device, the diffuse screen is removed, and the first light source under test is mounted on the sample stage. The light emitted by the first light source under test illuminates the equivalent plane p of the diffuse screen. Under the measurement conditions corresponding to each measured illuminance distribution image data, the pose of the first light source under test on the sample stage is adjusted according to the distance Δd between the position of the diffuse screen used for measurement and the plane p, so that the relative pose of the first light source under test when illuminating the equivalent plane p of the diffuse screen is consistent with the relative pose when illuminating the diffuse screen when measuring the illuminance distribution image data. The stray light elimination photodetector 5 moves on the equivalent plane p of the diffuse screen, measures the illuminance value at each point on the equivalent plane p of the diffuse screen, scans and measures the illuminance distribution data, and uses it as the illuminance distribution standard data. The collected illuminance distribution standard data sample {E s (x s ,y s The data is input and stored in the self-training standard data unit.

[0175] B4: Construction of the training sample dataset:

[0176] For each set of illuminance distribution image data, using the spatial angle interval in the illuminance distribution image data as a reference, spline interpolation is performed on the corresponding illuminance distribution standard data to make the spatial coordinate interval in the illuminance distribution image data and the processed illuminance distribution standard data the same; several sets of illuminance distribution image data and corresponding illuminance distribution standard data are obtained, and a training sample dataset is constructed.

[0177] B5: Establishment of a deep learning-based light distribution correction model:

[0178] In the deep learning-based light distribution correction processing unit, a light distribution correction model is built based on a convolutional neural network, with the structure as follows: Figure 5 , Figure 6As shown, it includes a basic feature extraction module, a channel feature extraction module, and a reconstruction module. The basic feature extraction module extracts shallow feature information of the image, the channel feature extraction module extracts deep feature information of different channels of the image and adjusts the feature weights of different channels, and the reconstruction module integrates the feature information of each channel and outputs the predicted illumination distribution correction data.

[0179] B6: Training of a Deep Learning-Based Optical Distribution Correction Model:

[0180] Illumination distribution image data from the training sample dataset Input light distribution correction model, light distribution correction model f k Output the predicted illuminance distribution correction data under parameter k. The loss function LOSS(k) is calculated as shown in formula (23):

[0181] LOSS(k) = LOSS data (k)+λ·LOSS phy (k) (23)

[0182] In the formula, LOSS data (k) represents the data reconstruction loss, calculated using formula (26); LOSS phy (k) is the physical consistency loss term, which can be calculated according to formula (25); λ is the weight parameter for balancing each loss term.

[0183]

[0184] In the formula, μ1 and μ2 represent the predicted illuminance distribution correction data, respectively. Compared with the standard data of illuminance distribution E s (x s ,y s The mean value of ) and σ represents the variance of the predicted illuminance distribution corrected data and the illuminance distribution standard data, respectively; 12 C1 and C2 represent the covariance between the predicted illuminance distribution correction data and the standard illuminance distribution data; C1 and C2 are stability constants.

[0185]

[0186] In the formula, Ex m,n ,y m,n This represents the measured illuminance distribution image data, where m and n represent the illuminance distribution image data respectively. The row and column numbers, M and N are the total number of rows and columns in the light distribution image data, f k The light distribution correction model represents the input Ex. m,n ,ym,n The predicted correction data, where L is the known linearity function of the imaging measurement device.

[0187] The model parameter k is optimized by minimizing the loss function calculation value, as shown in formula (5). This completes the supervised learning of the light distribution image data correction process, resulting in the optimized light distribution correction model with the parameter k. * .

[0188] B7: Measurement and correction of measured illuminance distribution image:

[0189] In the light distribution measurement and correction device, the second light source under test is mounted on the sample stage, and the illuminance distribution image data of the light emitted by the second light source under test illuminating the diffuse screen is obtained by using an imaging measurement device. The light distribution correction processing unit based on deep learning receives the illuminance distribution image data obtained by the imaging measurement device and outputs the corrected illuminance distribution data.

[0190] In the above embodiments, the measurement correction of the measured light distribution image in step S3 (corresponding to steps A8-A9 in embodiment one and step B7 in embodiment two) can also be performed in a different measurement system than when the training sample dataset was collected. The different measurement system refers to a change in the overall measurement environment, including the laboratory and measurement equipment. The measurement equipment includes imaging measurement equipment, a diffuse screen, a sample stage, etc.

[0191] Example 3

[0192] The following discloses a measurement device different from that used in step S1 for acquiring the training sample dataset, used for measurement correction of the measured light distribution image (corresponding to steps A8-A9 in Embodiment 1 and step B7 in Embodiment 2), as shown in the schematic diagram below. Figure 11 As shown, the system includes a sample stage 2, a diffused screen 3 positioned opposite the sample stage 2, an imaging measurement device 4, and a control and processing module 6. The diffused screen 3 is positioned at a distance from the light source. A stray light removal aperture 7 is positioned between the sample stage 2 and the diffused screen 3. The imaging measurement device 4 is positioned between the diffused screen 3 and the sample stage 2, and is aligned with the diffused screen 3, remaining fixed in a stationary state. The control and processing module 6 includes a deep learning-based light distribution correction processing unit 61, which is communicatively connected to the imaging measurement device 4. The light distribution correction processing unit receives the light distribution image data acquired by the imaging measurement device and outputs the corrected light distribution data.

[0193] In the above embodiments, the light distribution standard data of the light source does not need to be obtained in the same measurement system as the light distribution image data of the light source. The light distribution data of the light source from the traditional distribution photometric measurement standard laboratory can also be used as the light distribution standard data.

Claims

1. A method for measuring and correcting light distribution, characterized in that, Specifically, the following steps are included: S1: The light emitted by the first light source under test illuminates the diffuse screen, generating a light spot image. The imaging measurement device images the light spot image on the diffuse screen and obtains its light distribution image data. The stray light photometer detector moves relative to the first light source under test and scans to obtain the light distribution data emitted by the first light source under test, and uses this light distribution data as the light distribution standard data. By collecting measurement data from multiple different light sources, multiple sets of light distribution image data and corresponding light distribution standard data are obtained to construct a training sample dataset. S2: Establish a deep learning-based light distribution correction model, and train the light distribution correction model using a training sample dataset. Specifically, this includes: inputting light distribution image data into the light distribution correction model, and the model outputting predicted light distribution correction data; using the predicted light distribution correction data and the corresponding standard light distribution data to participate in the calculation of the loss function, optimizing the model parameters with the goal of minimizing the calculated value of the loss function, and obtaining the optimized light distribution correction model. S3: In the same or different measurement system as step S1, use an imaging measurement device to measure and obtain the light distribution image data of the light emitted by the second light source under test illuminating the diffuse screen. Input the light distribution image data into the optimized light distribution correction model to obtain the corrected light distribution data.

2. The light distribution measurement and correction method according to claim 1, characterized in that, In step S1, the poses of the first light source under test and the diffuse screen are relatively fixed. The imaging measurement device measures the light spot emitted by the first light source under test and illuminated by the diffuse screen under this relative pose to obtain light distribution image data; or, the first light source under test and the diffuse screen move relative to each other. The imaging measurement device measures the light spot emitted by the first light source under test and illuminated by the diffuse screen under two or more relative poses to obtain two or more light distribution image data, and integrates them to obtain new light distribution image data.

3. The light distribution measurement and correction method according to claim 1, characterized in that, In step S1, the process of obtaining the light distribution data emitted by the first light source under test by scanning the relative motion between the stray light elimination photometer and the first light source under test includes: using one or more stray light elimination photometers to directly receive the light beam emitted by the first light source under test, causing the stray light elimination photometer to move relative to the first light source under test, and the stray light elimination photometer scanning and measuring the light distribution data of the first light source under test as the spatial angle or spatial position changes.

4. The light distribution measurement and correction method according to claim 3, characterized in that, Based on the light distribution data scanned by the stray light photometer detector, new light distribution data is obtained by interpolation methods, and this light distribution data is used as the standard light distribution data. The interpolation methods include linear interpolation or spline interpolation.

5. The light distribution measurement and correction method according to claim 1, characterized in that, In step S1, the training sample dataset is expanded by changing any one or at least two of the following: different light source samples, different working modes of the same light source, different alignment positions of the same light source relative to the diffuse screen, and different angles of the same light source relative to the diffuse screen.

6. The light distribution measurement and correction method according to claim 1, characterized in that, In step S1, the light distribution image data of the first tested light source is subjected to data augmentation processing to generate a series of derived image data, which are then included in the training sample dataset. The data augmentation processing methods include any one or at least a combination of two of the following: blurring, adding noise, and distortion.

7. The light distribution measurement and correction method according to claim 1, characterized in that, In step S2, the light distribution correction model is built based on a convolutional neural network, which includes an encoder module and a decoder module. The encoder module includes several downsampling groups composed of convolutional layers and pooling layers, used to extract image features from the input light distribution image data. The decoder module includes several upsampling groups composed of upsampling layers and convolutional layers, used to process and restore the image features and output the predicted light distribution correction data. The number of upsampling groups and downsampling groups is the same, and there are skip connections between them.

8. The light distribution measurement and correction method according to claim 1, characterized in that, In step S2, the light distribution correction model is built based on a convolutional neural network, which includes a basic feature extraction module, a channel feature extraction module, and a reconstruction module. The basic feature extraction module includes a convolutional layer for extracting shallow feature information of the image. The channel feature extraction module includes several channel attention residual groups, a convolutional layer, and a residual connection. Each channel attention residual group contains several channel attention units with residual connections, a convolutional layer, and a residual connection, used to extract deep feature information of different channels of the image and adjust the feature weights of different channels. The reconstruction module includes a convolutional layer for integrating the feature information of each channel to output predicted light distribution correction data.

9. The light distribution measurement and correction method according to claim 1, characterized in that, In step S2, the loss function is calculated using either mean squared error or structural similarity index.

10. The light distribution measurement and correction method according to claim 9, characterized in that, In step S2, when calculating the loss function, the weight coefficients of different spatial angles or different spatial locations in the light distribution image data are different. The weighted mean square error or the weighted structural similarity index is used as the loss function. The weight coefficients are obtained by manual preset annotation or generation based on the light distribution threshold or gradient.

11. A light distribution measurement and correction method according to claim 1, 7, or 8, characterized in that, The characteristic functions of the measurement system or imaging measurement device are input into the light distribution correction model. These characteristic functions include a point spread function and / or a linearity function, which serve as prior physical knowledge to assist in training the light distribution correction model.

12. The light distribution measurement and correction method according to claim 1, characterized in that, Before training the light distribution correction model using the training sample dataset, data preprocessing is performed to unify the data resolution between the light distribution image data and the light distribution standard data. The method for unifying the data resolution includes: interpolating or super-resolution reconstructing the light distribution standard data, or cropping, sampling, or averaging the light distribution image data.

13. The light distribution measurement and correction method according to claim 1, characterized in that, Before training the light distribution correction model using the training sample dataset, the spatial angle or spatial position coordinates of the light distribution image data are adjusted to correct the alignment error between the light distribution image data and the standard light distribution data. Specific methods include: presetting spatial angle or spatial position coordinate adjustment options; calculating the alignment error between the adjusted light distribution image data and the corresponding standard light distribution data; selecting the adjustment option with the smallest alignment error to perform coordinate adjustment on the light distribution image data; the alignment error calculation method includes mean square error or average absolute error; or introducing a pre-trained image registration module, which includes an image feature extraction function group and a registration function group. The image feature extraction function group is used to extract image features from the light distribution image data and the standard light distribution data; the registration function group is used to calculate a matrix for coordinate transformation of the light distribution image data based on the extracted image features, and to perform coordinate transformation on the light distribution image data based on this matrix.

14. The light distribution measurement and correction method according to claim 1, characterized in that, The training sample dataset contains several sets of light distribution dark area image data, strong light-dark contrast image data, and their light distribution standard data. The training sample dataset is used to train a light distribution correction model for stray light error correction.

15. A light distribution measurement and correction method according to claim 1 or 14, characterized in that, In step S3, the spatial light distribution image of the second light source under test is processed by region. Only some regions of the image data are corrected using the light distribution correction model, while other regions are not corrected. The regions for which light distribution correction is performed include dark areas with low luminance values ​​or regions with significant light-dark contrast.

16. The light distribution measurement and correction method according to claim 1, characterized in that, Step S3 is followed by step S4: the light distribution standard data of the second light source under test is obtained by relative motion scanning measurement between the stray light photometer detector and the second light source under test in step S3, and together with the light distribution image data of the second light source under test, it is added to the training sample dataset; the model in S2 is iteratively trained using the supplemented training sample dataset to obtain the iterative light distribution correction model.

17. A light distribution measurement and correction device based on the light distribution measurement and correction method according to claim 1, characterized in that, include: The system includes a sample stage (2), a diffuse screen (3), an imaging measurement device (4), and a control processing module (6). The diffuse screen (3) is positioned at a certain distance from the light source (1). The light emitted by the light source (1) illuminates the diffuse screen (3) to generate a light spot image. The imaging measurement device (4) aligns with the diffuse screen (3) and obtains a light distribution image. The control processing module (6) receives the light distribution image data obtained by the imaging measurement device (4). The control processing module (6) also includes a light distribution correction processing unit (61) based on deep learning. The light distribution correction processing unit (61) is used to calculate and output the light distribution data of the light source based on the light distribution image data. The light source is either the first light source under test or the second light source under test.

18. The light distribution measurement and correction device according to claim 17, characterized in that, The system includes one or more stray light metering detectors (5), each stray light metering detector (5) including a photometric probe, a radiometric probe or a spectroradiometer and its sampling device, used to scan and measure the light distribution data emitted by the light source, and use this light distribution data as the light distribution standard data. The control processing module (6) also includes a self-training standard data unit (62), which is communicatively connected to the stray light metering detector (5). The self-training standard data unit (62) is used to receive the light distribution standard data acquired by the stray light metering detector for the self-training of the light distribution correction processing unit (61).

19. The light distribution measurement and correction device according to claim 17, characterized in that, The sample stage (2) includes a rotation mechanism and / or a translation mechanism. Driven by the sample stage (2), the light source (1) and the diffuse screen (3) move relative to each other. The sample stage (2) is connected to the control and processing module (6) in communication.

20. The light distribution measurement and correction device according to claim 17, characterized in that, An anti-stray light aperture (7) is provided between the light source (1) and the diffuser screen (3).

Images