Method and device for determining optical attenuation characteristics of display device

Abstract

The invention relates to a method and a device for determining optical attenuation characteristics of a display device. According to the embodiment provided by the invention, the total image brightness truth value data is generated through fusion of a plurality of sampling positions, a traditional simplified model is abandoned, the non-attenuation reference data is constructed directly based on physical measurement, the error of an attenuation tag is greatly reduced, and high-precision attenuation data support is provided for subsequent Demura correction. Through gray scale image acquisition under different gray scales, the attenuation prediction model can learn the attenuation law under each gray scale, the output optical attenuation characteristic data can adapt to a multi-gray-scale full scene from a low gray scale to a high gray scale and the like, the problem of gray scale drift type non-uniform brightness can be solved, and the accuracy of the optical attenuation prediction model is improved. Therefore, the brightness uniformity after gray scale correction is carried out on the optical attenuation characteristic data based on the scheme is obviously improved.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of computer vision and display device detection, and in particular to a display device optical attenuation characteristic determination method and device. BACKGROUND

[0002] In the manufacturing process of modern high-resolution micro display devices such as Micro-OLED, the micro lens array (MLA) is one of the core optical components, and its core role is to improve display brightness and improve viewing angle uniformity. As a key preprocessing link, pixel-level brightness uniformity correction (Demura, an image correction method for display panels such as OLED, MiniLED, etc., mainly through external compensation, the optical characteristics of the display device are corrected after the product) usually needs to collect panel brightness data through a high-precision camera to identify and compensate for inherent brightness non-uniformity defects (also known as Mura defects). However, the MLA has inherent "vignetting effect" (optical attenuation characteristic), which causes the real Mura defect information and the optical attenuation artifact introduced by the MLA to be mixed in the image data collected by the camera, which seriously interferes with the extraction of real Mura defects, directly leading to a decrease in Demura correction accuracy.

[0003] In some related technical solutions for solving optical attenuation, a single attenuation characteristic derived from a high gray scale is usually used to extend to all gray scales through a single high gray scale calibration mapping. The applicant found that the light transmittance and scattering characteristics of the MLA change nonlinearly with the driving gray scale, such as the significant enhancement of surface scattering effect at low gray scale, resulting in a steep attenuation curve and obvious differences in attenuation curve compared to high gray scale. The existing method introduces systematic errors, which further reduces the reliability of the measurement results.

[0004] Therefore, there is an urgent need for a solution that can more accurately measure the optical attenuation characteristics of a display device. SUMMARY

[0005] The present disclosure provides a plurality of embodiments of a display device optical attenuation characteristic determination method and device, at least one embodiment of which can obtain optical attenuation characteristic data with higher accuracy, solve the problem of brightness non-uniformity caused by gray scale drift, and significantly improve the brightness uniformity after gray scale correction based on the optical attenuation characteristic data of the present solution. The technical solution of the present disclosure includes the following: According to a first aspect of an embodiment of the present disclosure, a display device optical attenuation characteristic determination method is provided, comprising: obtaining a gray scale image of a display device to be measured at different gray scales, and determining brightness data of pixel nodes at each gray scale based on the gray scale image; inputting the luminance data into an attenuation prediction model to obtain optical attenuation characteristic data of the display device under test, wherein sample data used by the attenuation prediction model comprises: sample gray scale images of the display device under different gray scales at a plurality of sampling positions, attenuation prediction model input data generated based on luminance data of pixel nodes at each gray scale determined from the sample gray scale images, full-image luminance true value data at a corresponding gray scale generated according to luminance data at different gray scales of the plurality of sampling positions, and a comparison result of the full-image luminance true value data and measured luminance data of the display device as an attenuation label output by the attenuation prediction model.

[0006] In some embodiments of the method, the attenuation prediction model comprises a dual-path fusion feature network, the dual-path fusion feature network comprising a dual-path encoder, a cross-modal feature fusion module, a multi-gray scale linkage module, and a decoding output layer, wherein the dual-path encoder comprises a luminance encoding path and an angle encoding path for extracting gray scale luminance features and spatial angle features; the cross-modal feature fusion module fuses dual-path features of the luminance encoding path and the angle encoding path through an attention mechanism; the multi-gray scale linkage module models a gray scale nonlinear relationship through dynamic convolution and residual connection; and the decoding output layer restores spatial resolution through transposed convolution and skip connection, and outputs attenuation coefficients of each gray scale of a color channel.

[0007] In some embodiments of the method, the multi-gray scale linkage module extracts global statistical features of each gray scale channel, uses a multi-layer perceptron to learn the global statistical features to generate a dynamic convolution kernel related to input features, performs multi-group convolution processing on the input features using the dynamic convolution kernel to generate a plurality of groups of feature data reflecting linkage relationships between different gray scales, extracts spatial adaptive weights from the input features through an attention mechanism, and uses the spatial adaptive weights to perform weighted fusion on the plurality of groups of feature data.

[0008] In some embodiments of the method, in the training phase of the attenuation prediction model, the method further comprises: obtaining a shooting angle of the pixel node; generating full-image luminance true value data at a corresponding gray scale according to luminance data at different gray scales of the plurality of sampling positions, comprising: selecting luminance data in a region with a shooting angle less than a preset angle threshold as candidate effective true values, and generating full-image luminance true value data based on the candidate effective true values.

[0009] In some embodiments of the method, after obtaining the luminance data of the pixel nodes at each gray scale determined based on the sample gray scale images and the comparison result, the method further comprises: matching the luminance data corresponding to each color channel and the corresponding comparison result to generate initial sample training data of each color channel; The initial sample training data is down-sampled respectively to generate independent sample training data of each color channel.

[0010] In some embodiments of the method, the different gray scales include gray scale values selected according to equal division or equal proportion within a specified gray scale range.

[0011] In some embodiments of the method, during the training process of the attenuation prediction model, a staged training method is used, and a constraint condition of training is gradually increased in each training stage. The training stage includes a physical constraint stage, and the physical constraint stage uses a loss weight exponential warm-up function to dynamically adjust the weight of the physical constraint loss term. The expression of the loss weight exponential warm-up function is wherein beta_epoch is the weight coefficient of the physical constraint loss term in the epoch round of training, beta is the maximum weight of the physical constraint loss term, k is a preset attenuation coefficient for controlling the weight growth rate, and epoch is the training round.

[0012] In some embodiments of the method, the staged training includes: a basic reconstruction stage, in which the training round epoch is valued from 1 to E_1, an initial loss function selected is used as a loss function, and a first learning rate is used for training; a physical constraint introduction stage, in which the training round epoch is valued from E_1+2 to E_toal, a physical constraint function is added to construct a loss function, the weight coefficient of the physical constraint function is controlled using a loss weight exponential warm-up function, and a second learning rate is used for training; a fine tuning stage, in which the training round epoch is valued from E_2+1 to E_toal, E_toal is the total training round, a complete composite loss function is used, and a third learning rate is used for training; wherein E_1<E_2, the first learning rate<the second learning rate<the third learning rate.

[0013] In some embodiments of the method, the method further includes correcting the measured luminance data of the display device to be measured based on the optical attenuation feature data to obtain attenuation-corrected luminance data.

[0014] According to a second aspect of the embodiments of the present disclosure, a display device optical attenuation feature determination apparatus is provided, including: a collection module configured to acquire a gray scale image of a display device to be measured under different gray scales, and determine luminance data of a pixel node under each gray scale based on the gray scale image; The model processing module is configured to input the luminance data into an attenuation prediction model to obtain optical attenuation characteristic data of the display device under test, wherein sample data used by the attenuation prediction model comprises: sample grayscale images of the display device under different gray scales at a plurality of sampling positions, attenuation prediction model input data generated based on luminance data of pixel nodes at each gray scale determined according to the sample grayscale images, and full-image luminance true value data corresponding to different gray scales generated according to luminance data at the different gray scales of the plurality of sampling positions, and a comparison result of the full-image luminance true value data and measured luminance data of the display device is used as an attenuation label output by the attenuation prediction model.

[0015] The embodiments provided by the present disclosure fuse full-image luminance true value data generated by a plurality of sampling positions, discard traditional simplified models, directly construct reference data without attenuation based on physical measurement, greatly reduce the error of the attenuation label, and provide high-precision attenuation data support for subsequent Demura correction. And through grayscale image acquisition at different gray scales, the attenuation prediction model can learn the attenuation law at each gray scale (such as a high attenuation coefficient caused by low gray scale scattering enhancement and a low attenuation coefficient caused by high gray scale saturation), and the output optical attenuation characteristic data can adapt to multiple gray scales of full scenes from low gray scale to high gray scale, and can solve the problem of gray scale drift type luminance non-uniformity, so that the luminance uniformity after gray scale correction based on the optical attenuation characteristic data of the present solution is obviously improved.

[0016] It should be understood that the foregoing general description and the following detailed description are only exemplary and explanatory, and cannot limit the present disclosure. BRIEF DESCRIPTION OF DRAWINGS

[0017] The accompanying drawings incorporated in the specification and forming a part of the specification illustrate embodiments consistent with the present disclosure and, together with the description, serve to explain the principles of the present disclosure, and do not constitute an undue limitation on the present disclosure.

[0018] Figure 1 FIG. 1 is a schematic diagram of a system structure for determining optical attenuation characteristics of a display device according to an example embodiment; Figure 2 FIG. 2 is a flowchart of a method for determining optical attenuation characteristics of a display device according to an example embodiment; Figure 3 FIG. 3 is a flowchart of a model training stage and a model application stage in a method for determining optical attenuation characteristics of a display device according to an example embodiment; Figure 4 FIG. 4 is a schematic diagram of internal data processing of a multi-gray scale linkage module according to an example embodiment; Figure 5 FIG. 5 is a schematic block diagram of a device for determining optical attenuation characteristics of a display device according to an example embodiment. Figure 6 This is a schematic diagram of the internal structure of an electronic device according to an exemplary embodiment. Detailed Implementation

[0019] To enable those skilled in the art to better understand the technical solutions of this disclosure, the technical solutions in the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings.

[0020] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented in orders other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure. The terms "comprising," "including," or any other variations thereof are intended to cover a non-exclusive inclusion, such that a process, method, product, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, product, or apparatus. Without further limitation, the presence of other identical or equivalent elements in a process, method, product, or apparatus that includes said elements is not excluded. For example, the use of terms such as "first," "second," etc., is to denote names and does not indicate any specific order.

[0021] The method for determining the optical attenuation characteristics of a display device disclosed herein can be applied to, for example... Figure 1 In the implementation scenario shown. Figure 1 A schematic diagram of a system structure for implementing a method for determining the optical attenuation characteristics of a display device, illustrated in an exemplary embodiment, shows that the system can be divided into multiple modules or units according to physical or logical functions, or a functional module can include multiple units that implement specific functions. In some specific embodiments, the steps of the method can be implemented in... Figure 1The processing unit shown may include at least one processor, which can store / load / execute executable instructions of a pre-built attenuation prediction model to implement the method steps described in any embodiment of this disclosure. The processor may have functions such as data reception, processing, storage, and transmission, and may also interact with other devices or units. The processor may include a processor of a local terminal, such as an MCU (Microcontroller Unit), or a processor of a remote server. In some embodiments, the local processor and the processor of the remote server may work together to determine the optical attenuation characteristics of the display device. The processor may be integrated into various terminal devices, including but not limited to in-vehicle devices, personal computers, laptops, smartphones, tablets, wearable devices, medical devices, VR (Virtual Reality) devices, etc. The server may also be a single server, a server cluster, a distributed subsystem, a cloud processing platform, a server containing blockchain nodes, or a combination thereof. The processors described in this disclosure may include various control units capable of performing logic processing functions, including but not limited to CPU (Central Processing Unit), PLC (Programmable Logic Controller), ECU (Electronic Control Unit), MCU (Microcontroller Unit), and controllers composed of one or more logic function units, chips, etc.

[0022] To more clearly illustrate the technical solution of this disclosure, the following is combined with... Figure 1 The illustrated implementation scenarios provide information about some of the embodiments provided in this disclosure. For example... Figure 1As shown, the display may include a displacement platform module (containing a displacement platform for supporting fixed display devices, such as display panels, and capable of displacement), an optical acquisition module (camera + lens), a display control module (including a signal generator), a control module (industrial computer), a computing processing unit (data preprocessing, truth generation, model training, prediction, and output), and a data storage module, as well as their connection relationships. Arrows indicate the data flow direction, and dashed lines indicate that the camera's optical axis is perpendicular to the display panel and performs optical acquisition. The display device may include various display devices, components, units, etc., such as a display panel in this implementation scenario, and an industrial camera. As mentioned earlier, existing display devices such as MLA / Micro-OLED / Demura face the technical challenge of not being able to effectively separate MLA optical attenuation from the actual Mura during preprocessing. Current solutions only acquire images of the highest grayscale (e.g., RGB255) for attenuation estimation. This not only fails to cover all grayscale levels, but also results in pixel brightness approaching the sensor noise floor for lower grayscale levels (e.g., RGB32, RGB64), leading to a very low signal-to-noise ratio. This often causes confusion between Mura defects, optical attenuation, and random noise, making it impossible to obtain a reliable Mura true value. This further amplifies the undercorrection in low-light scenes. Although some existing technologies use optical attenuation models, these models generally use the fourth power of cosine law or a single radial attenuation function to fit the brightness attenuation introduced by the MLA. However, real MLAs have complex effects such as differences in microlens curvature, multiple scattering within the substrate, asymmetric aberrations, and ghosting. This leads to significant deviations between theoretical models and measured data, making it impossible to accurately describe the complex optical phenomena such as artifacts, stray light, and asymmetric aberrations present in real MLAs. As a result, inherent and unavoidable artifacts remain in the image even after model error attenuation stripping, severely affecting the accuracy of subsequent Mura recognition and compensation.

[0023] This disclosure provides a method for determining the optical attenuation characteristics of a display device. It abandons traditional theoretical models, acquiring data from physical measurements through multi-position scanning. Then, it autonomously generates high-precision full-image brightness truth data using an algorithm, achieving truth acquisition through physical measurements rather than model derivation. This fundamentally avoids model mismatch errors and significantly improves the accuracy of the matching data reference values ​​used in the model. Furthermore, the model used in this solution employs a multi-grayscale image joint processing structure, typically including high, medium, and low grayscale ranges. Data information at different grayscale levels usually exhibits different characteristics. For example, high grayscale (e.g., RGB255, RGB192) data provides the main signal, corresponding to the high-brightness scene of normal display panel display. At this time, the MLA transmittance is high and the attenuation curve is flat, capturing its attenuation characteristics under saturated transmittance. Medium grayscale (e.g., RGB128) corresponds to the medium brightness scene of daily display and is a transition zone of non-linear grayscale changes, reflecting the transition law of attenuation characteristics from flat to steep. Low grayscale data (e.g., RGB64, RFB32) can correspond to low-brightness scenarios on display panels (such as night mode or dark screen display). In this case, MLA scattering / diffraction effects are significant, and the attenuation curve is steep, which can solve the shortcomings of traditional methods such as low signal-to-noise ratio and inaccurate modeling at low grayscale. The attenuation prediction model of this scheme can generate ground truth values ​​for multiple grayscale independently and predict them separately, so that the model can more accurately capture the nonlinear relationship between grayscale and attenuation. In application, combining it with the full-image brightness ground truth data can output more accurate optical attenuation characteristic data.

[0024] It should be understood that the concepts of high, medium and low intervals, such as high grayscale and low grayscale, described in some embodiments of this disclosure can be set according to the feature value range of the object being processed, the scheme design, the constraints required to solve the technical problem, etc., to determine the number of intervals and the corresponding interval range. For example, in some examples, high, medium and low can correspond to 30%, 40% and 30% of the full feature value range, respectively.

[0025] Based on the above description, specifically, Figure 2 This is a flowchart illustrating a method for determining the optical attenuation characteristics of a display device according to an exemplary embodiment. Figure 2 As shown, in some embodiments provided in this disclosure, the method may include: S20: Obtain grayscale images of the display device under test at different grayscale levels, and determine the brightness data of pixel nodes at each grayscale level based on the grayscale images; S22: Input the brightness data into the attenuation prediction model to obtain the optical attenuation characteristic data of the display device under test. The sample data used by the attenuation prediction model includes: taking sample grayscale images of the display device at different grayscale levels at multiple sampling locations; generating attenuation prediction model input data based on the brightness data of pixel nodes at each grayscale level determined by the sample grayscale images; generating full-image brightness true value data at the corresponding grayscale level based on the brightness data at different grayscale levels at multiple sampling locations; and using the comparison result between the full-image brightness true value data and the measured brightness data of the display device as the attenuation label output by the attenuation prediction model.

[0026] The measured brightness data includes brightness data obtained by capturing images of the display device at different gray levels from a fixed position. In this embodiment, an attenuation prediction model can be pre-built, training samples can be constructed, and the attenuation prediction model can be trained. After training, grayscale images of the display device under test at different gray levels can be obtained during the application phase. Based on the grayscale images, the brightness data of pixel nodes at each gray level can be determined, and the brightness data can be input into the attenuation prediction model to obtain the optical attenuation characteristic data of the display device under test. Generally, the model selected in this disclosure can meet the requirements of covering multi-grayscale nonlinear responses, reflecting spatial angle dependence (vignetting effect), and performing end-to-end modeling using deep learning. Examples include the Lambert-Beer law improved model and the Monte Carlo enhanced attenuation model.

[0027] Figure 3This is a flowchart illustrating the model training and application stages of a method for determining the optical attenuation characteristics of a display device, according to an exemplary embodiment. During the model training stage, data acquisition can be performed. In some embodiments of this disclosure, data acquisition may include the acquisition of input sample data required for model training and the acquisition of label data output by the model. The acquisition of input sample data can be performed using sparse scanning at multiple angles to simulate the displacement of the display device under test in the X or Y direction by a displacement platform in actual production applications. Sample grayscale images of the display device at different grayscale levels can be captured at multiple sampling positions determined by sparse scanning. Based on the brightness data of pixel nodes at each grayscale level determined by the sample grayscale images, input data for the attenuation prediction model is generated. Specifically, for example, the displacement platform is controlled to drive the display device to move sparsely along a raster-like path in the XY plane, with the movement step size and the number of acquisition positions pre-set (typically covering the entire screen space of the display device). At each acquisition position, the display device is driven to sequentially display a positioning image and a multi-grayscale image, which are simultaneously captured by a camera. Multi-grayscale brightness data and spatial angle data of each sub-pixel are extracted based on the positioning image and combined into an input feature vector. The sparse multi-angle scanning is used to simulate scenarios where the fixed panel angle and XY direction displacement occur in actual production, enriching the spatial angle diversity of the input samples. In actual production, the display device maintains a fixed angle on the stage, but slight displacement occurs in the XY direction due to transmission errors, resulting in differences in the spatial angle of the pixels captured by the camera. Existing technologies mostly use single-position sample acquisition, making it difficult for the model to adapt to this scenario and resulting in weak generalization ability. This solution simulates the XY displacement in mass production through sparse multi-angle scanning, acquiring input samples at different spatial positions. Even with slight displacement of the display panel, the model can still accurately predict the attenuation coefficient.

[0028] The acquisition of label data output by the model corresponding to the input data typically involves multi-location acquisition of the display device, obtaining brightness data at different gray levels at multiple sampling locations. General label data requires high-density data to ensure accuracy, resulting in even higher acquisition density. Typically, this covers the full resolution of the display device, acquiring brightness data at different gray levels for all pixels, thus adapting to S-shaped close-step scanning. For example, in some implementations, a camera can be used to acquire data from each pixel of the display device from the frontal view, moving one pixel step at a time. Alternatively, pixels within a preset offset angle can be predefined as belonging to the frontal view. For instance, when shooting a pixel from the frontal view, pixels within a shooting angle range of 5 degrees can be considered as data acquired from the frontal view. This allows for the acquisition of a region's data at once. Pixels within the preset offset angle belong to the pixel in the very center of the shooting area and can be considered label data acquired from the frontal view. Subsequently, one or more pixel steps can be moved again to acquire data from other regions, ensuring that each pixel in the display device is scanned at least once from the frontal view, acquiring brightness data for each pixel from the frontal view.

[0029] The multi-angle sparse scanning of the input sample data and the S-shaped scanning of the model output label data can be processed separately (one scan can be performed first, followed by the other), each generating the input data and label data required by the model. In some implementations, the generation of full-image brightness ground truth data can also be based on the brightness data obtained from the acquired input data and processed by Demura. That is, the brightness data of pixel nodes at each gray level determined based on the sample grayscale image, which serves as the model input data, can also serve as the brightness data at different gray levels at multiple sampling locations required for the model output label data, without the need for separate scanning processing. Of course, the acquisition of input data can also use S-shaped scanning or sparse scanning for multi-angle input data acquisition, etc. During the acquisition or processing of input data, the pixel node can be a single pixel or a group of pixels, such as adjacent 2×2 or 8×8 pixel blocks. This disclosure describes the scheme using a single pixel as an example.

[0030] In some specific implementation examples, sample grayscale images can be acquired. A display device of the same model / batch as the display device under test (such as a Micro-OLED panel) can be selected as the sample device, fixed on a displacement platform, with the camera facing the panel and in a fixed position. The displacement platform can be controlled to move the sample device in the XY plane; the target position, sequence, or path of movement can be preset. In some embodiments of this disclosure, movement can be along a preset path (such as an S-shaped close-step scanning path), driving the sample device to display test images of different grayscale levels at multiple sampling positions (such as 5000 sampling positions, which can be set according to the display device resolution, data processing accuracy requirements, etc. to ensure data density). The camera simultaneously captures sample grayscale images at each sampling position. Generally, the acquired sample grayscale images cover the entire screen space of the display device. The selected grayscale levels typically cover high, medium, and low grayscale ranges (acquiring multiple grayscale ranges determined by a specific method), and at least one sample grayscale image is acquired for each grayscale range.

[0031] After acquiring the sample grayscale images, algorithms can be used to determine the brightness data of pixel nodes at each grayscale level, and the input data for model training can be generated based on this brightness data. In one embodiment, a display panel (e.g., 3468×3704 resolution) can be fixed on a displacement platform, with a camera facing the platform and fixed in place. The display screen driving the display panel sequentially displays a positioning map and a set of multi-grayscale images (e.g., RGB255, 192, 128, 64, 32), so that the pixel nodes in the display device are displayed at different grayscale levels in the red (R), green (G), and blue (B) primary color channels, and the camera captures sample grayscale images. The brightness value of each pixel at different grayscale levels can be calculated using the Demura processing algorithm (Demura processing typically involves determining the position of each pixel through the captured positioning map, and then extracting the different brightness levels of the target pixel at the pixel position; this implementation will not elaborate further). Of course, in other embodiments of this disclosure, the angle between each pixel and the camera's optical axis can also be calculated during this process (a shooting angle; in some embodiments of this solution, this angle may refer to the line connecting the midpoint of a pixel or pixel block to the light source point, the angle between this line and the horizontal and vertical directions of the plane where the display device is located, which can be calculated based on the camera height or relative position to the display device, the coordinates of the pixel node, and other relevant parameters). The mobile platform can be controlled to perform S-shaped scanning and taking pictures in the XY plane according to the scanning step size. For example, starting from the left side of the first scanning line (the camera has a scanning interval during scanning, usually referring to the line spacing, which can scan multiple lines at once; the first scanning line here can refer to the first line of the scanning interval), it gradually moves to the right. When scanning to the far right, the camera moves down the scanning interval and starts scanning from the right side to the left. Of course, the downward position can also be smaller than the scanning interval, so that the collected sample grayscale image can have data information of repeated pixel nodes). Thus, sample grayscale images at different shooting positions and different grayscale levels can be obtained based on the scanning and taking pictures. As mentioned above, the sample grayscale image usually includes the spatial position of the entire display screen to collect brightness data of all pixels at different grayscale levels. The Demura algorithm is used to calculate the brightness values ​​of different gray levels for each pixel, and these values ​​are combined to generate an input feature vector, which serves as the input data for the attenuation prediction model. The pixels of the display panel consist of three sub-pixels: red (R), green (G), and blue (B). By independently controlling the gray level value of each sub-pixel, different colors and brightness levels can be mixed to create images. In this embodiment, RGB255, 192, 128, 64, and 32 are used. This can be understood as each color channel being driven independently according to these five gray levels, obtaining the brightness data of pixel nodes at different gray levels under each color channel. For example, in one embodiment, the input feature vectors for different gray levels under the red channel (Input-Red) are: [R255_lv, R192_lv, R128_lv, R64_lv, R32_lv]; The input feature vectors for different gray levels under the green channel Input-Green are: [G255_lv, G192_lv, G128_lv, G34_lv, G32_lv]; The input feature vectors for different gray levels under the blue channel Input-Blue are: [B255_lv, B192_lv, B128_lv, B64_lv, B32_lv].

[0032] The brightness data of pixel nodes in the aforementioned sample grayscale image at each grayscale level are the brightness values ​​of each channel of the pixel at different grayscale levels as measured by actual physical measurements. Although processed by some computers or algorithms, this value is already data with extremely high fidelity. Therefore, data labels can be further constructed based on this data, which can be used as attenuation labels for the attenuation prediction model training output. In some specific implementations, the method of testing or measuring the display during the production and application stages can be simulated. Images of the display device can be taken individually, such as taking pictures of the center position of the display device and obtaining images at a specified grayscale level or multiple different grayscale levels, and obtaining the brightness data of the image (for ease of description and distinction, this can be referred to as measured brightness data). The brightness data of pixel nodes at different grayscale levels under each color channel obtained above can be directly used as the full-image brightness true value data of the corresponding grayscale level of the display device, or it can be called a full-resolution reference true value map, such as containing each pixel node in a resolution of 3468×3704, and containing brightness data of each color channel RGB at five grayscale levels (255, 192, 128, 64, 32). Of course, in other embodiments, the brightness data of pixel nodes at different gray levels under each color channel can also be processed in some preset ways to obtain the true brightness data of the entire image at the corresponding gray level, such as filtering, merging, weighting, and blending. The true brightness data of the entire image can be compared with the measured brightness data, such as calculating the ratio or difference, to obtain the comparison result.

[0033] The comparison result reflects the difference between the measured brightness data and the actual brightness data due to various hardware, system, and environmental factors. For example, the ratio can characterize the brightness attenuation feature, while the difference can characterize the brightness deviation. This comparison result can be used as an attenuation label for model training, serving as the output target, i.e., the output attenuation label. The attenuation label can be an attenuation coefficient, a difference, etc. Some embodiments of this disclosure use ratios as examples. For instance, the ratio obtained by dividing the true brightness data of the entire image by the measured brightness data can be recorded as the attenuation coefficient. In this embodiment, the attenuation coefficient labels for the five brightness channels corresponding to the three color channels can be as follows: The attenuation coefficient C-Red of the red channel is: [R255_c, R192_c, R128_c, R64_c, R32_c]. The attenuation coefficient C-Green for the green channel is: [G255_c, G192_c, G128_c, G34_c, G32_c]; The attenuation coefficient C-Blue of the blue channel is: [G255_c, G192_c, G128_c, G34_c, G32_c]; The aforementioned generated input feature vectors Input-Red, Input-Green, and Input-Blue can be paired with attenuation labels to form three training datasets: (Input-Red, C-Red), (Input-Green, C-Green), and (Input-Blue, C-Blue). The attenuation prediction model is then independently trained on the corresponding channels using the training data from these three datasets until convergence is achieved. The measurement data disclosed herein is brightness data obtained from grayscale images of display devices taken at fixed locations. This data simulates the actual measurement scenario of the device under test (i.e., a single fixed-position shot on the production line), typically containing a mixture of real Mura defects, MLA optical attenuation artifacts, and random noise. The comparison result with the full-image brightness ground truth data (attenuation labels) directly corresponds to the degree of influence of MLA optical attenuation on the measurement data. This ensures that the attenuation prediction model learns the attenuation rules under actual measurement scenarios, guaranteeing that the optical attenuation feature data output by the model can be applied to the attenuation characteristic detection of display devices. This achieves full-screen attenuation output in a single shot, balancing measurement accuracy and mass production efficiency, and is suitable for batch product testing.

[0034] This embodiment uses grayscale images at different gray levels, typically no less than three gray levels, such as from low grayscale RGB32 to high grayscale RGB255. One of its functions is to more comprehensively capture the nonlinear characteristics of optical attenuation of display devices as brightness changes, avoiding the limitations of traditional single grayscale measurements. The selection of different gray levels can satisfy the requirement of covering the nonlinear attenuation range. Generally, the selected grayscale values ​​need to include the key inflection point gray levels of the nonlinear change in MLA transmittance. The key inflection point gray levels include the low grayscale inflection point (64±8) and the high grayscale inflection point (192±8), and at least one grayscale value is located in the low grayscale range (32-64), at least two grayscale values ​​are located in the medium grayscale range (64-192), and at least one grayscale value is located in the high grayscale range (192-255), ensuring that the selected grayscale levels can both cover the full brightness range and accurately capture the nonlinear attenuation characteristics. Therefore, in some embodiments, the different grayscale levels include grayscale values ​​selected by dividing a specified grayscale range equally or proportionally. For example, taking 5 gray levels and the standard full gray level range (0-255) of the display device as an example, the gray level range can be specified as 32-255. Excluding low gray levels such as full black (0) and near black (e.g., 16) which have no practical display significance, 32 is selected as the starting low gray level and 255 as the highest gray level. The interval is calculated as (highest gray level - lowest gray level) / (number of gray levels - 1) and is approximately 55.75. The actual gray level values ​​selected can be RGB32, RGB64, RGB128, RGB192, and RGB255.

[0035] In this solution, a displacement platform moves the sample device, selecting a sufficient number of spatial sampling points across the entire screen to capture grayscale images. This allows for the acquisition of attenuation-free brightness data in the near-optical axis center region, avoiding the problem that traditional grayscale images from a single sampling location can only cover attenuation-free data in a local center region. This provides a data foundation for subsequently generating true brightness data for the entire image, addressing the industry pain point of lacking a global attenuation-free brightness benchmark. Moreover, this true brightness data covers the entire screen of the display device, meaning it is essentially a physical measurement rather than a theoretically derived brightness benchmark. Unlike traditional truth estimation based on simplified models such as the fourth power of cosine law, this embodiment, by collecting and generating true brightness data from multiple central regions, completely avoids model mismatch errors and accurately reflects the ideal brightness distribution of the display device without MLA attenuation interference. This is one of the core prerequisites for ensuring the accuracy of attenuation labels.

[0036] As can be seen from the above embodiments, some embodiments of this disclosure generate true brightness data for the entire image by fusing multiple sampling locations. This abandons the traditional simplified model and directly constructs attenuation-free benchmark data based on physical measurements, significantly reducing the error of the attenuation label and providing high-precision attenuation data support for subsequent Demura correction. Furthermore, by acquiring grayscale images at different grayscale levels, the attenuation prediction model can learn the attenuation patterns at each grayscale level (such as high attenuation coefficients due to enhanced scattering at low grayscale levels and low attenuation coefficients due to saturation at high grayscale levels). The output optical attenuation characteristic data can be adapted to all scenes with multiple grayscale levels, from low to high, solving the problem of grayscale drift-type brightness non-uniformity. This significantly improves the brightness uniformity after grayscale correction based on the optical attenuation characteristic data of this scheme.

[0037] In some of the above embodiments, the optical attenuation feature data output by the model can be an attenuation label, such as an attenuation coefficient, or optical attenuation feature data obtained by further processing the attenuation label. For example, the optical attenuation feature data of the model data can be the attenuation coefficient after 4×4 downsampling in some of the aforementioned embodiments. In this case, 4×4 upsampling with bilinear interpolation can be performed to restore the original resolution of 3468×3704, resulting in an attenuation characteristic distribution map of the entire screen and resolution. The optical attenuation feature data obtained in the embodiments of this disclosure provides accurate attenuation characteristic distribution data for the optical correction of the display panel, which can be used in the subsequent Demura correction process to correct the brightness data of the display device under test and obtain brightness data without attenuation effects. Based on this, in other embodiments of this disclosure, the method may further include: The measured brightness data of the display device under test is corrected based on the optical attenuation characteristic data to obtain attenuation-corrected brightness data.

[0038] As mentioned earlier, the measured brightness data is typically obtained by capturing images of the display device at different gray levels from a fixed position. Here, the measured brightness data of the display device under test can be obtained by capturing images of the display device at different gray levels from a fixed camera position in actual production applications. For example, by facing the camera directly at the center pixel of the display device, images at full resolution are captured at five gray levels to obtain measured brightness data for each gray level. Then, the optical attenuation characteristic data disclosed herein can be used to correct the measured brightness data of the display device exhibiting attenuation, resulting in attenuation-corrected brightness data with no attenuation (or with significantly reduced attenuation effects or attenuation effects within an acceptable range). This embodiment provides a precise basis for Demura correction. Through attenuation correction, the effects of attenuation artifacts and gray level drift are accurately removed, greatly restoring the Mura defect in the corrected brightness data. The accuracy of Mura defect identification is significantly improved, enhancing the reliability of product quality inspection and overall product quality.

[0039] Furthermore, the brightness data after attenuation correction has already undergone attenuation preprocessing. The Demura system does not need to design additional complex attenuation compensation algorithms and can focus solely on the identification and compensation of genuine Mura defects, thus simplifying the process. At the same time, the above method can be fully automated without manual intervention, significantly reducing the Demura correction time per display panel, effectively improving mass production efficiency and reducing industrial application costs.

[0040] As described in some of the preceding embodiments, some embodiments of this disclosure can also calculate the shooting angle during sample data acquisition and generation, or tag data acquisition and generation. Taking a single sub-pixel as an example, the angle between each sub-pixel and the camera optical axis can be calculated, which can describe the spatial position of the sub-pixel relative to the camera's imaging optical center. In specific implementation examples, it can refer to the angle between the line connecting the geometric center of the sub-pixel to the optical center of the camera lens (i.e., the incident light ray of the sub-pixel) and the camera optical axis (the central axis of the camera lens, usually a reference line perpendicular to the display panel). The shooting angle can include two dimensions. One is the X-axis angle, which can be denoted as θ. The angle between the incident light ray and the X-axis in the plane of the display panel reflects the degree to which a sub-pixel deviates from the optical axis in the horizontal direction (X-axis) of the panel; the other is the angle along the Y-axis, which can be denoted as θ. The angle between the incident light ray and the Y-axis in the plane of the display panel reflects the degree to which a sub-pixel deviates from the optical axis in the vertical direction (Y-axis) of the panel. Based on this, the camera's optical axis can be understood as perpendicular to the display panel, and can be considered the Z-axis. When the camera captures a pixel perpendicularly, both the X-axis and Y-axis angles of that pixel are 90 degrees. The specific method for calculating the shooting angle can use a checkerboard calibration board based on Zhang Zhengyou's calibration algorithm. Alternatively, it can be based on coordinates of the display panel, pixel nodes, and the camera (mobile platform) coordinates, through corresponding transformations and mappings.

[0041] Existing methods (such as single-grayscale static calibration) rely solely on luminance data modeling, ignoring the influence of spatial angle on attenuation, resulting in large attenuation prediction errors for edge sub-pixels. Some embodiments of this disclosure can address this by introducing θ... θ This expands the input of the attenuation prediction model from a single brightness dimension to a dual dimension of brightness and spatial angle. Specifically, in the data processing stage, for the brightness data generated at each sampling location, brightness data within the region where the shooting angle is less than a preset angle threshold (e.g., the angle with the optical axis is less than 5 degrees) can be selected as candidate valid ground truth values. Then, the candidate valid ground truth values ​​from different sampling locations can be mapped to the global coordinate system according to the world coordinates of the displacement platform, thereby obtaining the full-image brightness ground truth data. Based on this, this disclosure also provides some other embodiments of the method, in which the method further includes the following during the training stage of the attenuation prediction model: S24: Obtain the shooting angle of the pixel node; S220: Generate full-image brightness true value data at corresponding gray levels based on brightness data at different gray levels of multiple sampling locations, including: selecting brightness data in areas where the shooting angle is less than a preset angle threshold as candidate valid true values, and generating full-image brightness true value data based on the candidate valid true values.

[0042] The shooting angle reflects the distance of a pixel node from the optical center during imaging. Generally, the smaller the angle, the weaker the optical attenuation introduced by the MLA (microlens array), and the closer the brightness data is to the ideal state of no attenuation. In this embodiment, the brightness data within the range where the shooting angle is less than a preset angle threshold usually corresponds to the central area directly below the camera. From the brightness data at multiple sampling positions, the brightness data of pixel nodes whose shooting angle meets the preset threshold is selected. Since the attenuation of this type of data is negligible, it can be used as an effective value for generating the full-image brightness true data. For each pixel node, candidate effective true values ​​can be selected based on the shooting angle. Through spatial coordinate mapping and some data fusion algorithms, the candidate effective true values ​​at different sampling positions can be stitched together to form full-image brightness true value data covering the entire screen of the display device. Such full-image brightness true value data has the physical characteristic of serving as a zero-attenuation true value benchmark (attenuation is negligible), providing an accurate data comparison benchmark for attenuation label generation.

[0043] Specific methods for concatenating and fusing candidate valid ground truth values ​​can be implemented using selected or custom algorithms, which will not be elaborated here. For example, a Gaussian weighted fusion algorithm can be used to ensure that the central region has a high weight and the edge region has a low weight, thus eliminating concatenation artifacts. Of course, the input feature vector generated based on the shooting angle can also include the shooting angle of the pixel node under the corresponding color channel. For example, the input feature vector of a pixel node under different gray levels in the red channel Input-Red can be: The input feature vectors for other color channels, such as green and blue, are similar.

[0044] In some other embodiments of the method provided in this disclosure, the attenuation prediction model can employ a channel-specific independent training strategy, constructing an improved dual-path fusion feature network model with identical structure for each of the R, G, and B color channels. For example, the model input dimension can be 7×867×926, where the 7 channels correspond to 5 grayscale brightness data and 2 spatial angle data, and the output dimension is 5×867×926, with the 5 channels predicting the attenuation coefficients (attenuation labels) for the five grayscale levels. All three models employ a staged training approach, gradually increasing the training difficulty and constraints to accelerate convergence. Specifically, in some other embodiments of the method, the attenuation prediction model includes a dual-path fusion feature network, which includes a dual-path encoder, a cross-modal feature fusion module, a multi-grayscale linkage module, and a decoding output layer. The dual-path encoder includes a luminance encoding path and an angle encoding path for extracting grayscale luminance features and spatial angle features. The cross-modal feature fusion module fuses the dual-path features of the luminance encoding path and the angle encoding path through an attention mechanism. The multi-grayscale linkage module models the nonlinear relationship of grayscale through dynamic convolution and residual connections. The decoding output layer restores spatial resolution through transposed convolution and skip connections, and outputs the attenuation coefficients of each grayscale level in the color channel.

[0045] In some of the foregoing embodiments, multi-location input sample data acquisition is completed, generating Input_Red, Input_Green, Input_Blu and attenuation coefficient labels C-Red, C-Green, and C-Blue. The original high-resolution data (3468×3704) can be converted to low-resolution data through, for example, 4×4 average pooling downsampling processing. Therefore, after obtaining the brightness data and comparison results of pixel nodes at each grayscale level determined based on the sample grayscale image, the method further includes: Match the brightness data corresponding to each color channel with the corresponding comparison results to generate the initial sample training data for each color channel; The initial sample training data is downsampled to generate independent sample training data for each color channel.

[0046] Training datasets can be constructed based on the principle of independent color channels, ensuring that the training of each channel's model is not affected by the attenuation characteristics of other channels. Since optical attenuation characteristics are spatially continuous and do not undergo abrupt changes, the downsampling method in this embodiment significantly reduces the amount of data and computational complexity while maintaining the overall trend characteristics of the attenuation distribution, without causing any loss of accuracy (the loss is minimal and negligible). This significantly reduces computational complexity and memory usage, accelerating model training. Simultaneously, the averaging operation introduced during downsampling effectively suppresses random noise interference, improving the signal-to-noise ratio of the data. The model trained based on this data exhibits minimal error growth as the interference appears or increases in the device under test, even with slight noise interference (non-linear synchronization), thus giving the trained model better generalization ability and robustness.

[0047] In some embodiments of this disclosure, a dual-path feature fusion network is designed to address the coupling problem between multi-grayscale nonlinear response and spatial angle dependence in the optical attenuation characteristics of MLA. This network achieves accurate modeling of attenuation characteristics through multi-scale feature extraction, cross-modal adaptive fusion, and physical constraint optimization. Specifically, taking a 7-channel network with 5 multi-grayscale brightness data points and 2 spatial angle data points as the network input, the input data dimension is n_batch×7×H×W, where n_batch is the batch size (e.g., 16), and H×W is the downsampled feature map size (e.g., 867×926), corresponding to the original panel resolution of 3468×3704 after 4×4 average pooling. In the first 5 channels, the 5 grayscale brightness data of any color channel of R / G / B, such as R255_Lv, R192_Lv, R128_Lv, R64_Lv, and R32_Lv of the red channel, are extracted from the grayscale image and normalized to the range of [0,1] by the Demura processing algorithm; In the last two channels, the spatial angle data of the corresponding pixel nodes Of course, the unit can also be converted to radians and then normalized. (Corresponding to 0°-5°, it can better adapt to the characteristics of the low attenuation region of MLA).

[0048] In other embodiments, the training data can be randomly flipped horizontally, vertically, and rotated by ±1° to simulate the slight offset of the display panel in actual generation applications. This ensures that the network learns the decay law that does not change the spatial position, and further improves the model's generalization ability.

[0049] The dual-path independent encoding design of this disclosure can extract two types of modal features through the luminance encoding path and the angle encoding path respectively, avoiding mutual interference. The specific structure and operation example are as follows: Luminosity encoding path (extracting grayscale luminosity features): Input: 5-channel multi-grayscale brightness data (n_batch×5×H×W); Structure: It contains a 5-level downsampling module. Each level can be cascaded with a 3×3 convolutional layer (the number of convolutional kernels is 64→128→256→512→1024 in sequence), an instance normalization layer, a GELU activation function, and an SE channel attention mechanism to extract brightness features at different gray levels and calibrate channel importance through the attention mechanism.

[0050] During processing, downsampling can be performed first. This can be achieved by using convolutional layers with a stride of 2 to halve the feature map size (final output size is H / 32×W / 32), gradually increasing the feature abstraction. The SE channel attention mechanism performs global average pooling (GAP) on each level of output features. Channel weights are generated through fully connected layers (e.g., compression ratio 16:1), ReLU activation, fully connected layers (restoring the original number of channels), and Sigmoid activation, weighting and calibrating the luminance feature channels (e.g., higher grayscale channels have higher weights to highlight strong signal features). The output of the luminance encoding path can be a 512-channel grayscale luminance feature map (n_batch×512×H / 32×W / 32).

[0051] Angle encoding pathway (extracting spatial angle features): Input: 2-channel spatial angle data (n_batch×2×H×W); Structure: A 3-level lightweight convolutional module is used, with each level consisting of a 3×3 convolutional layer (the number of convolutional kernels is 32→64→128 in sequence), layer normalization, GELU activation function, and sequential concatenation. Here, the SE channel attention mechanism can be omitted to reduce parameter redundancy.

[0052] The number of convolutional kernels in the angle encoding path can be half that of the luminance encoding path, and only 3 levels of downsampling are needed (the final output size is the same as the luminance encoding path, H / 32×W / 32), thus balancing the accuracy of angle feature extraction and computational efficiency. Considering the more uniform distribution of angle data, this implementation example can use layer normalization instead of instance normalization, which can accelerate model convergence. The output can then be a 128-channel spatial angle feature map (n_batch×128×H / 32×W / 32).

[0053] The cross-modal feature fusion module enables interaction between dual-path features. The 128-channel feature map output from the angle encoding path is upsampled using a 1×1 convolutional layer (compressing the number of channels to 512) and bilinear interpolation to ensure its channel count and spatial size are consistent with the 512-channel feature map of the luminance encoding path (n_batch×512×H / 32×W / 32), achieving feature size alignment. The aligned luminance features are then concatenated with the angle features (dimension n_batch×1024×H / 32×W / 32), and input into a small convolutional network consisting of a 3×3 convolutional layer (64 channels), GELU activation, a 1×1 convolutional layer (1 channel), and Sigmoid activation. This generates a spatially adaptive attention weight map (n_batch×1×H / 32×W / 32), where bright areas in the weight map indicate that the luminance feature is more important at that location, and dark areas indicate that the angle feature is more important. In dual-path feature fusion processing, weighted modulation and element-wise addition strategies can be adopted. For example, for brightness feature weighting, the brightness feature map can be multiplied by the corresponding attention weight map to enhance the brightness information of key areas. When retaining angle features, the angle feature map can be multiplied by (1 - attention weight map) to retain the angle information of non-key areas. Then, the weighted brightness features and the weighted angle features are added together to obtain an n_batch×512×H / 32×W / 32 cross-modal fusion feature map.

[0054] The multi-grayscale linkage module establishes nonlinear relationships between different grayscale levels through dynamic convolution and residual connections. Traditional methods assume that the attenuation coefficients between grayscale levels change independently or linearly, which contradicts actual physical laws. This module achieves input-dependent, spatially adaptive multi-grayscale feature interaction through a dynamic weight generation mechanism. (Reference) Figure 4 , Figure 4 This is a schematic diagram illustrating the internal data processing flow of a multi-grayscale linkage module according to an exemplary embodiment. For example... Figure 4 As shown, the multi-grayscale linkage module extracts global statistical features of each grayscale channel through global average pooling (GAP), learns these global statistical features using a multilayer perceptron, and generates dynamic convolutional kernels related to the input features. These dynamic convolutional kernels are then used to perform multiple sets of convolutional processing on the input features, generating multiple sets of feature data reflecting the linkage relationship between different grayscale levels. Weights (which can be called spatial adaptive weights) are extracted from the input features through an attention mechanism, and these spatial adaptive weights are used to perform weighted fusion of the multiple sets of feature data. This allows for adaptive weight selection at different locations in different grayscale linkage modes. A specific implementation example is shown below. Figure 4 As shown, it may include: The dynamic weight generation subnetwork receives multi-grayscale features related to grayscale from the fused feature map, such as separating features corresponding to 5 grayscale levels from 512 channels, with dimensions of n_batch×5×H / 32×W / 32. Specifically, global average pooling (GAP) can be applied to each grayscale channel to extract global statistical features for each grayscale channel, resulting in n_batch×5×1×1 global statistical features. A multilayer perceptron (MLP) is used to learn complex interaction patterns between grayscale levels. Dynamic convolutional kernel parameters (n_batch×5×K) can be generated through a structure of "fully connected layer (256 neurons) + GELU activation + fully connected layer (K×5 neurons, K=8)". In the weight parameter reshaping process, the parameters can be reshaped to n_batch×5×1×1×K, corresponding to K groups of 5-channel 1×1 dynamic convolutional kernels.

[0055] Dynamic convolution processing involves performing 1×1 convolutions on the input 5-channel grayscale features using K sets of generated dynamic convolution kernels, resulting in K sets of output feature maps (n_batch×K×H / 32×W / 32). Each set of feature maps corresponds to a grayscale linkage mode (such as coupling between low and medium grayscale levels, or saturation characteristics of high grayscale levels), enabling information interaction and feature transformation between multiple grayscale channels. Through dynamic convolution processing, multiple sets of feature representations reflecting different grayscale linkage modes are generated, greatly enhancing the model's ability to characterize complex nonlinear relationships between grayscale levels.

[0056] In adaptive fusion, spatial adaptive weights are extracted from the original input through an attention mechanism. Multiple sets of features generated by dynamic convolution are weighted and fused to achieve adaptive selection of different locations across different grayscale linkage modes. For example, spatial weights are extracted from the original 5-channel grayscale input (generated as n_batch×K×1×1 weights through 1×1 convolution), and the K sets of feature maps are weighted and summed. Then, residual connections are performed, adding the fused feature map elements to the original 5-channel grayscale feature map, providing an identity mapping path, avoiding gradient vanishing, preserving original grayscale information, improving feature utilization, mitigating the gradient vanishing problem, and making deep networks easier to train. Then, the feature map can be calibrated using an n_batch×5×H / 32×W / 32 linkage (5 channels corresponding to 5 grayscale levels).

[0057] The decoding output layer can include both a decoder and an output layer. The decoder can contain 5 upsampling modules, each cascaded in the following order: transposed convolutional layer (stride 2, kernel count 512→256→128→64→5), instance normalization, and GELU activation, progressively restoring the feature map size from H / 32×W / 32 to H×W (867×926). This decoder employs a progressive upsampling structure, restoring spatial resolution through transposed convolution and making skip connections with features of the same scale in the encoder. The output of each upsampling module is element-wise added to the brightness feature map of the same scale in the encoder (matched via 1×1 convolution channels), preserving spatial details lost during downsampling (such as the attenuation gradient of sub-pixels at panel edges). In the output layer, channel compression can be performed, compressing the 5-channel feature map (corresponding to 5 gray levels) output from the last upsampling stage to 5 channels using 1×1 convolution (no need to change the number of channels, only adjust the weights). The Sigmoid activation function is used to strictly constrain the output values ​​to a physically reasonable range of [0,1] (the attenuation coefficient cannot be negative or greater than 1). Then, resolution restoration is performed by upsampling the output n_batch×5×867×926 feature map using 4×4 bilinear interpolation to restore it to the original panel resolution of 3468×3704, resulting in the attenuation coefficient distribution map (η_R255, η_R192, η_R128, η_R64, η_R32) for the five gray levels of the color channel (e.g., Red).

[0058] Repeat the above encoding-fusion-linkage-decoding process to train dual-path fusion feature networks for the R, G, and B color channels respectively, outputting the grayscale attenuation coefficients for each of the three channels. In the dual-path encoder + cross-modal attention fusion structure described above, the brightness path of the dual independently encoded paths focuses on extracting grayscale brightness differences (such as the brightness gradient between high and low grayscale levels), while the angle path focuses on capturing spatial location information (such as the angle increment of edge sub-pixels), avoiding modal interference. The attention fusion mechanism adaptively allocates modal weights (such as high brightness weight in the central region and high angle weight in the edge region), making the fused features more consistent with the MLA attenuation rule of "center relies on brightness, edge relies on angle". In feedback data from actual industrial applications, the edge sub-pixel attenuation prediction error can be reduced from 10% to less than 2.5%, and the spatial consistency of the full-screen attenuation distribution can be improved by about 40%. At the same time, the embodiments of this disclosure use a multi-grayscale linkage module, in which the dynamic convolutional kernel can autonomously learn the nonlinear relationship between grayscale levels (such as the scattering difference between R32 and R64, and the saturation characteristics of R255), without the need for manually preset grayscale association rules. Furthermore, the module design employs a shared network structure across all three channels. This allows for adjustments to the input color channel data without requiring redesign of the module parameters. The decoding output layer can directly output the full grayscale attenuation coefficients for each channel without additional data conversion. In practical training and mass production applications, the training cycle of the three-channel model can be significantly shortened. The device under test can acquire the attenuation data of all three channels simultaneously with a single multi-grayscale image capture, greatly reducing measurement time and improving measurement efficiency.

[0059] In some embodiments of the method provided in this disclosure, during the training process of the attenuation prediction model, a loss weight exponential warm-up function and a composite loss function are also designed, and a segmented training strategy is adopted. Introducing multiple optical constraints in this way can improve the reasonableness of the output results in conforming to the physical characteristics in actual generation applications and the accuracy of the results. The loss weight exponential warm-up function described in the embodiments of this disclosure generally refers to an exponential weight function used to dynamically adjust the weights of physical constraint terms. In one example, the loss weight exponential warm-up function beta can be... One of its functions is to achieve a gradual increase in constraint strength during the model training phase. The segmented training strategy, relying on the loss weight exponential warm-up function, divides model training into three stages: basic reconstruction, physical constraints, and fine-tuning. This progressively optimizes the model's learning process from data fitting to physical compliance, resolving the technical contradiction in traditional training where excessively strong constraints lead to convergence difficulties, while insufficient constraints result in physical distortion. The physical constraints can be understood as pre-set parameters (modifying / adding / reducing parameters, or changing them according to a preset method) set by the operator to ensure the output meets actual generation requirements. Therefore, in some embodiments of the method provided in this disclosure, the attenuation prediction model uses a segmented training approach during training. Each training stage gradually increases the training constraints. The training nodes include a physical constraint stage, where the loss weight exponential warm-up function dynamically adjusts the weights of the physical constraint loss term. The expression for the loss weight exponential warm-up function is: Where: beta_epoch is the weight coefficient of the physical constraint loss term in the epoch-th training round, beta is the maximum weight of the physical constraint loss term, k is the preset decay coefficient used to control the weight growth rate, and epoch is the training round.

[0060] In one implementation example, beta can range from 0.1 to 0.5, k can range from 0.05 to 0.01 to control the weight growth rate, and epoch can range from 1 to 200. It should be noted that the above parameter values ​​are merely illustrative descriptions of the implementation process of this solution and do not limit the scope of the implementation. For example, the number of epoch training rounds can be set to 10,000 or even more, and the maximum weight of beta can be greater than 0.5.

[0061] This embodiment introduces a loss weight exponential warm-up function, which can dynamically adjust the physical constraint weights to avoid training gaps caused by traditional fixed weights. This ensures that the initial training round weights are close to 0, prioritizing the model's learning of the basic mapping. As the training rounds increase, the weights gradually increase to beta, progressively strengthening the physical constraints while balancing convergence speed and physical rationality.

[0062] In the specific phased training process, this disclosure provides a complete composite loss function, as shown below: ; In the above formula, `loss_rec` is the multi-scale reconstruction loss function, which can be used in the initial training phase. `loss_mono` is the monotonicity constraint loss, `loss_smooth` is the spatial smoothness constraint loss, and `loss_continuity` is the grayscale continuity constraint loss. `alpha`, `beta`, `gamma`, and `delta` are the weights corresponding to the above four loss functions, respectively. This disclosure is not limited to the above four loss functions; other or more loss functions can be used. Specifically, one of `loss_rec`, `loss_mono`, `loss_smooth`, and `loss_continuity` can be selected as the initial loss function used in the basic reconstruction phase. Other loss functions besides the initial loss function can be used as physical constraint functions. One or more physical constraint functions can be added in the physical constraint phase for iterative training. In the physical constraint phase, all physical constraint functions can be added at once and used in combination as loss functions, or they can be added one by one or a portion at a time until all physical constraint functions are added. In a specific implementation, the phased training includes: In the basic reconstruction phase, the training epochs range from 1 to E_1, using the selected initial loss function as the primary loss function (e.g., the multi-scale reconstruction loss function) and training with the first learning rate. This decay prediction model uses the loss function loss_stage1 = loss_rec, and the learning rate can be set as follows: This stage focuses on data fitting, enabling the model to grasp the basic mapping between input features and attenuation coefficients, laying the foundation for subsequent physical constraints.

[0063] In the physical constraint introduction phase, the training epochs range from E_1+2 to E_toal. Physical constraint functions are added, and a loss function is constructed. The physical constraint functions use a loss weight exponential warm-up function to control the weight coefficients. For example, the loss function `loss_stage2` at this stage is: ; Training is performed using a second learning rate, such as In some embodiments, multiple physical constraint terms such as loss_mono, loss_smooth, and loss_continuity can be added one by one, multiple at a time, or all at once. This stage focuses on physical correction, gradually introducing constraints through dynamic weights to correct physical distortions in the underlying mapping.

[0064] In the fine-tuning phase, the training epochs range from E_2+1 to E_toal, where E_toal is the total number of training epochs, and the complete composite loss function is used. For example, the loss function loss_stage3 at this stage is: Use a third learning rate for training, such as 10. -5 This phase focuses on fine-tuning, using fixed weights to ensure the optimal balance between data accuracy and physical compliance.

[0065] During the training process described above, E_1 < E_2, and the first learning rate < the second learning rate < the third learning rate.

[0066] This disclosed embodiment introduces a loss weight exponential warm-up function to achieve a gradual increase in the weights of physical constraints. The model prioritizes learning the mapping of basic data, avoiding gradient explosion. As training progresses, the weights gradually increase, guiding the model to gradually conform to physical laws while fitting the data, fundamentally resolving the contradiction between convergence speed and physical rationality. This hierarchical optimization approach aligns with the cognitive laws of model learning, avoids goal confusion in single-stage training, and significantly reduces the parameter tuning time. In some production line applications, this time can be reduced from one week to one day. The success rate of solution reproduction is also greatly improved, which can significantly reduce production line adaptation costs, meet mass production requirements, and has significant practical value in product manufacturing and inspection processes.

[0067] It is understood that the various embodiments of the methods described in this specification are presented in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on its differences from other embodiments. Relevant details can be found in the descriptions of other method embodiments.

[0068] It should be understood that although the steps in the flowcharts shown in the accompanying drawings are displayed sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise expressly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some of the steps in the accompanying drawings may include multiple steps or stages, which are not necessarily completed at the same time, but may be executed at different times, and the execution order of these steps or stages is not necessarily sequential, but may be performed alternately or in turn with other steps or at least a portion of the steps or stages of other steps.

[0069] Based on the description of the above-described embodiments of the method for determining the optical attenuation characteristics of a display device, this disclosure also provides an apparatus for determining the optical attenuation characteristics of a display device. The apparatus may include a system (including a distributed system), software (application), module, component, controller, server, terminal, etc., using the method described in the embodiments of this specification, combined with necessary hardware implementation. Based on the same innovative concept, the apparatuses in one or more embodiments provided in this disclosure are as described in the following embodiments. Since the implementation schemes and methods for solving the problem by the apparatus are similar, the implementation of specific apparatuses in the embodiments of this specification can refer to the implementation of the foregoing method, and repeated details will not be repeated. As used below, the terms "unit" or "module" can refer to a combination of software and / or hardware that implements a predetermined function. Although the apparatuses described in the following embodiments are preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0070] Figure 5 This is a schematic block diagram of an optical attenuation characteristic determination device for a display device according to an exemplary embodiment. The device can be the aforementioned terminal device, or integrated into the terminal device module, component, device, control unit, etc. For details, please refer to... Figure 5 The device 100 may include: The acquisition module 50 can be used to acquire grayscale images of the display device under test at different grayscale levels, and determine the brightness data of pixel nodes at each grayscale level based on the grayscale images. The model processing module 52 can be used to input the brightness data into an attenuation prediction model to obtain optical attenuation characteristic data of the display device under test. The sample data used by the attenuation prediction model includes: sample grayscale images of the display device at different grayscale levels captured at multiple sampling locations; generating attenuation prediction model input data based on the brightness data of pixel nodes at each grayscale level determined by the sample grayscale images; generating full-image brightness true value data at the corresponding grayscale levels based on the brightness data at different grayscale levels from multiple sampling locations; and using the comparison result between the full-image brightness true value data and the measured brightness data of the display device as the attenuation label output by the attenuation prediction model. The measured brightness data includes brightness data obtained based on images of the display device captured at different grayscale levels from fixed locations.

[0071] Regarding the apparatus in the above embodiments, the specific methods by which each module performs its operations have been described in detail in the embodiments related to the method, and will not be elaborated upon here. Based on the description of the foregoing method embodiments, the apparatus provided by this disclosure may further include apparatus embodiments corresponding to the various method embodiments of this disclosure, to implement the execution steps or functions of the corresponding methods. For example, it may also include a first module, a second module, a third module, etc., which respectively execute the steps or functions of the corresponding methods. For instance, the model processing module 52 may include a training module, which includes the structure of a dual-path fusion feature network and the functions of its various hardware and functional units.

[0072] Based on the foregoing description of the relevant methods and apparatus embodiments, this disclosure also provides a computer-readable storage medium. When the instructions in the computer-readable storage medium are executed by a processor of an electronic device, the electronic device is able to implement the method for determining the optical attenuation characteristics of a display device as described in any embodiment of this disclosure. Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program can implement the method for determining the optical attenuation characteristics of a display device as described in any embodiment of this disclosure. The storage medium can be, for example, ROM / RAM, magnetic disk, optical disk, etc.

[0073] In an exemplary embodiment, a computer program product is also provided, including a computer program that, when executed by a processor, implements the method for determining the optical attenuation characteristics of a display device as described in any embodiment of this specification.

[0074] In an exemplary embodiment, an electronic device is also provided, comprising: At least one processor; Memory used to store the processor's executable instructions; The processor is configured to execute the instructions to implement the optical attenuation characteristic determination method of any display device according to any one of the present disclosures, or the electronic device includes the optical attenuation characteristic determination device of at least one display device according to the present disclosure.

[0075] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. In particular, for hardware + program embodiments, since they are basically similar to the method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0076] In one exemplary embodiment, an electronic device is provided, which may be a controller, a device, or a device within a server. Its internal structure diagram may be as follows: Figure 6 As shown. The electronic device may include a processor, memory, input / output interface, communication interface, and input device, and may also include a display unit. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are connected to the system bus via the input / output interface. The processor of the electronic device provides computing and control capabilities. The memory of the electronic device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The input / output interface of the electronic device is used for exchanging information between the processor and external devices. The communication interface of the electronic device is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, Near Field Communication (NFC), or other technologies. When the computer program is executed by the processor, it implements the method for determining the optical attenuation characteristics of the display device according to any embodiment of this disclosure. The display unit of the electronic device is used to form a visually visible image and may be a display screen, a projection device, or a virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the electronic device can be a touch layer covering the display screen, or buttons, trackballs, or touchpads set on the casing of the electronic device, or external keyboards, touchpads, or mice, etc.

[0077] Those skilled in the art will understand that Figure 6 The structure shown is only a block diagram of a part of the structure related to the present application and does not constitute a limitation on the electronic device to which the present application is applied. The specific electronic device may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements.

[0078] It should be noted that the aforementioned apparatus, devices, electronic devices, etc., as described in the method embodiments, may also include other implementation methods. Specific implementation methods can be found in the descriptions of the relevant method embodiments. Furthermore, new embodiments formed by combinations of features from various methods, apparatuses, devices, and server embodiments still fall within the scope of this disclosure and will not be elaborated upon here.

[0079] For ease of description, the above devices are described in terms of function, divided into various modules. Of course, when implementing one or more of these specifications, the functions of each module can be implemented in the same or different software and / or hardware, or a module that performs the same function can be implemented by a combination of multiple sub-modules or sub-units. The device embodiments described above are merely illustrative. For example, the division of modules or units is only a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling and communication connections between the devices or units shown or described can be implemented through direct and / or indirect coupling / connection, through standard or custom interfaces or protocols, and can be implemented electrically, mechanically, or in other forms.

[0080] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This disclosure is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated directly or indirectly by the content of the specification.

[0081] It should be understood that this disclosure is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope.

Claims

1. A method for determining the optical attenuation characteristics of a display device, characterized in that, The method includes: Acquire grayscale images of the display device under test at different grayscale levels, and determine the brightness data of pixel nodes at each grayscale level based on the grayscale images; The brightness data is input into the attenuation prediction model to obtain the optical attenuation characteristic data of the display device under test. The sample data used by the attenuation prediction model includes: taking sample grayscale images of the display device at different grayscale levels at multiple sampling locations; generating attenuation prediction model input data based on the brightness data of pixel nodes at each grayscale level determined by the sample grayscale images; generating full-image brightness true value data at the corresponding grayscale level based on the brightness data at different grayscale levels at multiple sampling locations; and using the comparison result between the full-image brightness true value data and the measured brightness data of the display device as the attenuation label output by the attenuation prediction model.

2. The method for determining the optical attenuation characteristics of a display device according to claim 1, characterized in that, The attenuation prediction model includes a dual-path fusion feature network, which comprises a dual-path encoder, a cross-modal feature fusion module, a multi-grayscale linkage module, and a decoding output layer. The dual-path encoder includes a luminance encoding path and an angle encoding path, used to extract grayscale luminance features and spatial angle features. The cross-modal feature fusion module fuses the dual-path features of the luminance encoding path and the angle encoding path through an attention mechanism. The multi-grayscale linkage module models the nonlinear relationship of grayscale levels through dynamic convolution and residual connections. The decoding output layer restores spatial resolution through transposed convolution and skip connections, and outputs the attenuation coefficients of each grayscale level in the color channels.

3. The method for determining the optical attenuation characteristics of a display device according to claim 2, characterized in that, The multi-grayscale linkage module extracts the global statistical features of each grayscale channel, uses a multilayer perceptron to learn the global statistical features, and generates a dynamic convolution kernel related to the input features; the dynamic convolution kernel is used to perform multiple sets of convolution processing on the input features to generate multiple sets of feature data reflecting the linkage relationship between different grayscale levels. Spatial adaptive weights are extracted from the input features using an attention mechanism, and the spatial adaptive weights are then used to perform weighted fusion of the multiple sets of feature data.

4. The method for determining the optical attenuation characteristics of a display device according to claim 1, characterized in that, During the training phase of the decay prediction model, the method further includes: Obtain the shooting angle of the pixel node; The step of generating full-image brightness true value data at corresponding gray levels based on brightness data at different gray levels of multiple sampling locations includes: selecting brightness data in areas where the shooting angle is less than a preset angle threshold as candidate valid true values, and generating full-image brightness true value data based on the candidate valid true values.

5. The method for determining the optical attenuation characteristics of a display device according to claim 1, characterized in that, After obtaining the brightness data and comparison results of pixel nodes at each gray level determined based on the sample grayscale image, the method further includes: Match the brightness data corresponding to each color channel with the corresponding comparison results to generate the initial sample training data for each color channel; The initial sample training data is downsampled to generate independent sample training data for each color channel.

6. The method for determining the optical attenuation characteristics of a display device according to claim 1, characterized in that, The different gray levels include gray level values ​​selected by dividing the specified gray level range equally or proportionally.

7. The method for determining the optical attenuation characteristics of a display device according to claim 1, characterized in that, The training process of the decay prediction model employs a phased training approach, with each training phase progressively increasing the training constraints. The training phase includes a physical constraint phase, in which the weights of the physical constraint loss term are dynamically adjusted using a loss weight exponential warm-up function. The expression for the loss weight exponential warm-up function is as follows: Where: beta_epoch is the weight coefficient of the physical constraint loss term in the epoch-th training round, beta is the maximum weight of the physical constraint loss term, k is the preset decay coefficient used to control the weight growth rate, and epoch is the training round.

8. The method for determining the optical attenuation characteristics of a display device according to claim 7, characterized in that, The phased training includes: In the basic reconstruction phase, the training epochs range from 1 to E_1, the selected initial loss function is used as the loss function, and the first learning rate is used for training. In the physical constraint introduction phase, the training epoch values ​​range from E_1+2 to E_toal. A physical constraint function is added, and a loss function is constructed. The physical constraint function uses a loss weight exponential warm-up function to control the weight coefficients, and a second learning rate is used for training. In the fine-tuning phase, the training epochs range from E_2+1 to E_toal, where E_toal is the total number of training epochs. The complete composite loss function is used, and a third learning rate is employed for training. Where E_1 < E_2, and the first learning rate < the second learning rate < the third learning rate.

9. The method for determining the optical attenuation characteristics of a display device according to claim 1, characterized in that, The method further includes: correcting the measured brightness data of the display device under test based on the optical attenuation characteristic data to obtain attenuation-corrected brightness data.

10. A device for determining the optical attenuation characteristics of a display device, characterized in that, include: The acquisition module is used to acquire grayscale images of the display device under test at different grayscale levels, and determine the brightness data of the pixel nodes at each grayscale level based on the grayscale images. The model processing module is used to input the brightness data into the attenuation prediction model to obtain the optical attenuation characteristic data of the display device under test. The sample data used by the attenuation prediction model includes: taking sample grayscale images of the display device at different grayscale levels at multiple sampling locations; generating attenuation prediction model input data based on the brightness data of pixel nodes at each grayscale level determined by the sample grayscale images; generating full-image brightness true value data at the corresponding grayscale level based on the brightness data at different grayscale levels at multiple sampling locations; and using the comparison result between the full-image brightness true value data and the measured brightness data of the display device as the attenuation label output by the attenuation prediction model.

Images