A LED screen brightness equalization method based on a dynamic color adjustment algorithm

Abstract

The application discloses a kind of LED screen brightness equalization methods based on dynamic toning algorithm, comprising the following steps: the display content of LED screen is divided into multiple partitions according to physical structure, obtains partition display content data;Real-time acquisition of the intensity of illumination, temperature and actual brightness of each partition environment, obtains partition environmental parameter data;Partition display content data and partition environmental parameter data are carried out feature fusion, and generate partition joint feature data;Partition joint feature data is input to partition CycleGAN model, and output partition enhancement content data;The overall enhancement display content data obtained by seamless splicing of partition enhancement content data is output to LED screen driving system;Actual display effect and environmental change of each partition are continuously collected, and obtain partition feedback data, and input to partition CycleGAN model.The application is realized LED screen brightness equalization by using partition CycleGAN and multi-objective loss optimization.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of LED display and intelligent image processing technology, and particularly relates to an LED screen brightness equalization method based on a dynamic color adjustment algorithm. BACKGROUND

[0002] At present, LED display screens are widely used in the fields of advertising media, traffic indication, stage performance and monitoring command. With the continuous expansion of the area of LED screens and the continuous improvement of the requirements for display quality, the traditional overall brightness and color adjustment method has been difficult to meet the actual needs. In the prior art, LED screen brightness equalization usually relies on global brightness mapping, static color correction or a simple partitioned brightness adjustment scheme. These methods mainly adjust the whole screen or local area with fixed parameters or preset strategies, and cannot achieve fine response to different regional environmental changes and dynamic content. The existing partitioned adjustment method mainly uses a brightness or color sensor feedback based on hardware sampling, combined with a partitioned correction based on rules or linear algorithms, but the interaction between multiple partitions, seamless connection at the boundary, color consistency and dynamic adaptive ability are not considered, which easily leads to problems such as brightness discontinuity, color discontinuity and slow environmental adaptive response between partitions.

[0003] In addition, with the popularization of high-resolution large screens and outdoor complex environments, the demand for dynamic brightness equalization and color optimization of LED screens in strong light interference, temperature change and environmental diversity scenarios is more prominent. Although the existing deep learning-based image enhancement or style transfer methods have made progress in the field of image processing, they are mostly aimed at natural images or single content enhancement, and lack systematic solutions to the problems of multi-source data fusion, partition-level adaptive color adjustment and seamless splicing in the partitioned scenario of LED screens.

[0004] Therefore, how to provide an LED screen brightness equalization method based on a dynamic color adjustment algorithm is a problem that those skilled in the art need to solve. SUMMARY

[0005] One object of the present application is to provide an LED screen brightness equalization method based on a dynamic color adjustment algorithm. The present application combines partition feature extraction, partitioned CycleGAN neural network, partitioned environmental parameter fusion and multi-objective loss optimization, and describes in detail the overall process of realizing dynamic adaptive equalization of brightness and color, seamless splicing and online feedback optimization in a multi-partitioned LED screen. The method has the advantages of strong brightness equalization, high color consistency, excellent environmental adaptability and natural and continuous display effect, and can effectively improve the intelligent display quality and user experience of various LED display screens in complex scenarios.

[0006] According to the LED screen brightness equalization method based on the dynamic color adjustment algorithm, the display content of the LED screen is divided into multiple partitions according to the physical structure, and partition display content data is obtained.

[0007] The display content of the LED screen is divided into multiple partitions according to the physical structure, and partition display content data is obtained.

[0008] For each partition, the ambient light intensity, temperature and actual brightness of the corresponding area are collected in real time to obtain partition environment parameter data.

[0009] The partition display content data and the partition environment parameter data are fused to generate partition joint feature data.

[0010] The partition joint feature data is input into the partition CycleGAN model, and partition enhanced content data is output.

[0011] The partition enhanced content data is seamlessly spliced to generate overall enhanced display content data, and the overall enhanced display content data is output to the LED screen driving system.

[0012] The actual display effect and environmental changes of each partition are continuously collected to obtain partition feedback data, and the partition feedback data is input as online feedback into the partition CycleGAN model to dynamically optimize the partition CycleGAN model parameters.

[0013] Optionally, the partition display content data includes pixel data of an original image in the partition, spatial position information of the partition in the LED screen, resolution parameters of the partition, average brightness value of the partition, average color temperature value, color histogram, contrast value, and time stamp of partition content collection.

[0014] Optionally, the partition environment parameter data includes ambient light intensity value, ambient temperature value, actual brightness value, spatial position information of the partition in the LED screen, and time stamp of partition environment parameter collection.

[0015] Optionally, the partition display content data and the partition environment parameter data are fused to generate partition joint feature data, which specifically includes:

[0016] The original image pixel data, spatial position information, resolution parameters, average brightness value, average color temperature value, color histogram, contrast value, and time stamp of partition content collection in the partition display content data are extracted.

[0017] The ambient light intensity value, ambient temperature value, actual brightness value, spatial position information, and time stamp of partition environment parameter collection in the partition environment parameter data are extracted.

[0018] The partition display content data and the partition environment parameter data are matched according to the partition space position information and the timestamp, so that the data are one-to-one corresponding;

[0019] The partition display content data and all parameters in the partition environment parameter data are feature spliced to generate partition joint feature data, and all the partition joint feature data are numbered and stored according to the partition index.

[0020] Optionally, the inputting the partition joint feature data into the partition CycleGAN model and outputting partition enhanced content data specifically includes:

[0021] The partition joint feature data of each partition is input into the partition CycleGAN model, and the partition CycleGAN model includes a partition generator and a partition discriminator;

[0022] The partition generator receives the partition joint feature data of each partition to generate partition enhanced content data;

[0023] The partition discriminator receives the partition enhanced content data and respectively outputs true-false discrimination scores, brightness consistency evaluation scores, color consistency evaluation scores, and style and structure consistency determination results;

[0024] Based on the true-false discrimination scores, the brightness consistency evaluation scores, the color consistency evaluation scores, and the style and structure consistency determination results, a multi-objective loss optimization function is constructed, and the partition CycleGAN model parameters are jointly optimized to obtain a trained partition CycleGAN model;

[0025] The trained partition CycleGAN model is used to infer the partition joint feature data to output partition enhanced content data, and the partition enhanced content data is numbered and stored.

[0026] Optionally, the partition generator receives the partition joint feature data of each partition to generate partition enhanced content data specifically includes:

[0027] The partition generator receives the partition joint feature data of each partition, and the partition generator includes an input feature fusion layer, an encoding layer, a feature mapping layer, and a decoding layer;

[0028] The input feature fusion layer receives the partition joint feature data of each partition, splices and normalizes the partition display content data and the corresponding partition environment parameter data at a feature level, and outputs a fusion feature vector;

[0029] The encoding layer receives the fusion feature vector, extracts deep features of the fusion feature vector, and outputs an encoding feature representation;

[0030] The feature mapping layer receives the encoded feature representation, adopts an adaptive adjustment mechanism for dynamic changes of the partition content and environmental features to perform partition-level optimization processing on the brightness and color features, and outputs a feature mapping result;

[0031] The decoding layer receives the feature mapping result, performs up-sampling and restoration on the feature mapping result, outputs partition enhanced content data, and numbers and stores all the partition enhanced content data.

[0032] Optionally, the partition discriminator receives the partition enhanced content data, and outputs true-false discrimination scores, brightness consistency evaluation scores, color consistency evaluation scores, and style and structure consistency determination results, specifically including:

[0033] The partition discriminator receives the partition enhanced content data, and the partition discriminator includes an input feature extraction layer, a local discrimination layer, and a consistency judgment layer;

[0034] The input feature extraction layer receives the partition enhanced content data, performs multi-scale feature extraction on the partition enhanced content data, and outputs partition multi-scale feature representation;

[0035] The local discrimination layer receives the partition multi-scale feature representation, and calculates true-false discrimination scores based on a discrimination network;

[0036] The consistency judgment layer receives the partition multi-scale feature representation, calculates brightness consistency evaluation scores, color consistency evaluation scores, and style and structure consistency determination results;

[0037] The true-false discrimination scores, the brightness consistency evaluation scores, the color consistency evaluation scores, and the style and structure consistency determination results are outputted.

[0038] Optionally, the multi-objective loss optimization function is constructed to jointly optimize the partition CycleGAN model parameters to obtain a trained partition CycleGAN model, specifically including:

[0039] The multi-objective loss optimization function of the partition CycleGAN model is set, and the multi-objective loss optimization function is a weighted combination of an adversarial loss function, a partition cycle consistency loss function, a partition brightness balance loss function, and a partition color consistency loss function;

[0040] The adversarial loss function is constructed based on the true-false discrimination scores output by the partition discriminator, specifically as follows: each partition discriminator outputs the probability of real content and generated content;

[0041] The partition cycle consistency loss function is constructed based on the consistency of the partition content, and is independently calculated at the partition level based on the CycleGAN cycle consistency loss function;

[0042] A partition luminance balance loss function is constructed based on a luminance consistency evaluation score of the generated content, and an error between the luminance of the generated content and the target luminance is measured.

[0043] A partition color consistency loss function is constructed based on a color consistency evaluation score of the generated content, and an error between the color distribution of the generated content and the target color distribution is measured.

[0044] A multi-objective loss optimization function is used to jointly optimize the partition CycleGAN model parameters through back propagation, and iterative training is performed until the multi-objective loss optimization function converges or a preset condition is met, and the trained partition CycleGAN model is output.

[0045] Optionally, the seamless splicing of the partition enhanced content data to generate the overall enhanced display content data and the output of the overall enhanced display content data to the LED screen driving system specifically include:

[0046] According to the physical structure partition index of the LED screen, all partition enhanced content data are sequentially sorted to determine the corresponding spatial position of each partition enhanced content data on the screen, and the sorted partition enhanced content data set is output.

[0047] The boundary pixels of each partition enhanced content data in the sorted partition enhanced content data set and the boundary pixels of adjacent partition enhanced content data are subjected to luminance and color smoothing processing to generate a set of partition enhanced content data subjected to seamless fusion processing.

[0048] The set of partition enhanced content data subjected to seamless fusion processing is spliced according to the corresponding spatial position to generate the overall enhanced display content data.

[0049] The overall enhanced display content data is output to the LED screen driving system.

[0050] Optionally, the continuous acquisition of the actual display effect and environmental changes of each partition to obtain partition feedback data and the dynamic optimization of the partition CycleGAN model parameters specifically include:

[0051] During the dynamic display of the LED screen, the actual display effect data of each partition is collected in real time.

[0052] The actual display effect data includes the luminance value, color distribution, smoothness evaluation result of the display content boundary transition area, and real-time physical environment parameters of the partition currently output by the partition.

[0053] The actual display effect data and the partition environment parameter data are compared synchronously to generate the partition feedback data.

[0054] The partition feedback data includes a deviation between a partition actual brightness and a target brightness, a difference between a partition actual color distribution and a target color distribution, and a partition environment parameter change amount.

[0055] The partition feedback data is input as online feedback into the trained partition CycleGAN model, and a partition brightness equalization loss function and a partition color consistency loss function are combined to correct the partition CycleGAN model parameters in real time.

[0056] The present application has the following advantages:

[0057] The LED screen brightness equalization method based on the dynamic color adjustment algorithm can realize fine-grained brightness and color dynamic adaptive adjustment of the LED screen multi-partition.

[0058] The present application introduces a pixel-level smoothing and seamless splicing mechanism in the partition boundary processing, which improves the unnatural brightness and color transition problems between partitions.

[0059] In addition, the online feedback mechanism is used to continuously collect and dynamically adjust the actual display effect and environment parameters of each partition, support real-time optimization of the partition CycleGAN model, and effectively improve the response speed and intelligent adaptive level of the system to external environmental changes and dynamic display content. BRIEF DESCRIPTION OF DRAWINGS

[0060] The accompanying drawings are included to provide a further understanding of the present application, and constitute a part of the specification, illustrate the present application together with the embodiments thereof, and explain the present application, but do not limit the present application. In the drawings:

[0061] Figure 1 The flowchart of the LED screen brightness equalization method based on the dynamic color adjustment algorithm is shown in the drawings.

[0062] Figure 2 A partition generator structure schematic diagram in an LED screen brightness equalization method based on a dynamic color adjustment algorithm is provided for the present application.

[0063] Figure 3 A partition discriminator structure schematic diagram in an LED screen brightness equalization method based on a dynamic color adjustment algorithm is provided for the present application.

[0064] Figure 4 A partition CycleGAN model training flowchart in an LED screen brightness equalization method based on a dynamic color adjustment algorithm is provided for the present application. DETAILED DESCRIPTION

[0065] The present application will now be further described in detail with reference to the drawings. These drawings are simplified schematic diagrams and only show the basic structure of the present application in a schematic manner, and thus only show the components related to the present application.

[0066] REFERENCE Figures 1-4 An LED screen brightness equalization method based on a dynamic color adjustment algorithm includes the following steps:

[0067] The display content of the LED screen is divided into multiple partitions according to the physical structure, and partition display content data is obtained;

[0068] For each partition, the ambient light intensity, temperature and actual brightness of the corresponding area are collected in real time to obtain partition environment parameter data;

[0069] The partition display content data and the partition environment parameter data are feature fused to generate partition joint feature data;

[0070] The partition joint feature data is input into a partition CycleGAN model, and partition enhanced content data is output;

[0071] The partition enhanced content data is seamlessly spliced to generate overall enhanced display content data, and the overall enhanced display content data is output to an LED screen driving system;

[0072] The actual display effect and environmental changes of each partition are continuously collected to obtain partition feedback data, and the partition feedback data is input as online feedback into the partition CycleGAN model to dynamically optimize the partition CycleGAN model parameters.

[0073] In this embodiment, the partition display content data includes the pixel data of the original image in the partition, the spatial position information of the partition in the LED screen (including row index and column index or pixel coordinate range), the resolution parameters of the partition (including width and height), the average brightness value of the partition, the average color temperature value, the color histogram, the contrast value, and the timestamp of the partition content collection;

[0074] An overall content image of a display area of the LED screen to be processed is acquired, denoted as original display content data;

[0075] According to a preset partition scheme, the original display content data is divided into a plurality of partitions according to the physical structure parameters of the LED screen, each partition having a unique row index and column index;

[0076] For each partition, a corresponding sub-area in the original display content data is extracted to generate partition display content data, and all the partition display content data are numbered and stored according to the partition index;

[0077] The partition display content data includes pixel data of the original image in the partition, spatial position information of the partition in the LED screen (including row index and column index or pixel coordinate range), resolution parameters of the partition (including width and height), average brightness value of the partition, average color temperature value, color histogram, contrast value, and time stamp of the partition content collection.

[0078] In the embodiment, the partition environment parameter data includes environmental illumination intensity value, environmental temperature value, actual brightness value of the partition corresponding area, spatial position information of the partition in the LED screen (including row index and column index or pixel coordinate range), and time stamp of the partition environment parameter collection;

[0079] An environment collection device is arranged in each partition area of the LED screen to collect the environmental illumination intensity, environmental temperature and actual brightness of the partition;

[0080] During the operation of the LED screen, the environment collection device of each partition is periodically controlled to collect data to obtain the environmental illumination intensity value, environmental temperature value and actual brightness value of each partition, and the corresponding partition row index and column index and time stamp information are assigned to each collection;

[0081] The environmental illumination intensity value, environmental temperature value, actual brightness value, spatial position information of the partition and collection time stamp of each partition are integrated to form corresponding partition environment parameter data, and all the partition environment parameter data are numbered and stored according to the partition index.

[0082] In the embodiment, the feature fusion of the partition display content data and the partition environment parameter data to generate the partition joint feature data specifically includes:

[0083] The original image pixel data, spatial position information, resolution parameters, average brightness value, average color temperature value, color histogram, contrast value and time stamp of the partition content collection in the partition display content data are extracted;

[0084] extracting an ambient light intensity value, an ambient temperature value, an actual brightness value, spatial position information, and a timestamp of partition environment parameter collection from the partition environment parameter data;

[0085] matching the partition display content data and the partition environment parameter data according to the spatial position information and the timestamp, to ensure one-to-one correspondence of the data;

[0086] performing feature splicing on all parameters in the partition display content data and the partition environment parameter data to generate partition joint feature data, and numbering and storing all partition joint feature data according to a partition index, wherein the partition joint feature data includes original image pixel data, spatial position information, resolution parameters, an average brightness value, an average color temperature value, a color histogram, a contrast value, a content collection timestamp, an ambient light intensity value, an ambient temperature value, an actual brightness value, and an environment parameter collection timestamp.

[0087] In the embodiment, inputting the partition joint feature data into the partition CycleGAN model and outputting partition enhanced content data specifically includes:

[0088] inputting the partition joint feature data of each partition into the partition CycleGAN model, wherein the partition CycleGAN model includes a partition generator and a partition discriminator, and the partition generator and the partition discriminator are independently set for each partition structure;

[0089] the partition generator receives the partition joint feature data of each partition to generate partition enhanced content data, wherein the partition enhanced content data has brightness equalization and color optimization characteristics, and the partition generator includes an input feature fusion layer, an encoding layer, a feature mapping layer, and a decoding layer, the input feature fusion layer is used to receive and fuse the partition joint feature data of each partition, the encoding layer is used to extract fused features, the feature mapping layer is used to perform adaptive transformation of brightness and color, and the decoding layer is used to output the partition enhanced content data;

[0090] the partition discriminator receives the partition enhanced content data and outputs true-false discrimination scores, brightness consistency evaluation scores, color consistency evaluation scores, and style and structure consistency determination results, respectively, wherein the partition discriminator includes an input feature extraction layer, a local discrimination layer, and a consistency judgment layer, the input feature extraction layer is used to extract multi-scale features of the partition enhanced content data, the local discrimination layer is used to perform fine-grained discrimination on the authenticity and style of the partition enhanced content data, and the consistency judgment layer is used to judge the consistency of the partition enhanced content data with target brightness distribution and color distribution;

[0091] Based on the true and false discrimination score, the brightness consistency evaluation score, the color consistency evaluation score and the style and structure consistency determination result, a multi-objective loss optimization function including the adversarial loss function, the partition cycle consistency loss function, the partition brightness balance loss function and the partition color consistency loss function is constructed, and the partition CycleGAN model parameters are jointly optimized to obtain the trained partition CycleGAN model.

[0092] The trained partition CycleGAN model is used to infer the partition joint feature data, output the partition enhanced content data, and number and store the partition enhanced content data as the input of the subsequent seamless splicing and overall enhanced display content generation.

[0093] In the embodiment, the partition generator receives the partition joint feature data of each partition, and generates the partition enhanced content data, which specifically includes:

[0094] The partition generator receives the partition joint feature data of each partition, and the partition generator includes an input feature fusion layer, an encoding layer, a feature mapping layer and a decoding layer.

[0095] The input feature fusion layer receives the partition joint feature data of each partition, performs feature-level splicing and normalization on the partition display content data and the corresponding partition environment parameter data, and outputs a fusion feature vector.

[0096] The encoding layer receives the fusion feature vector, performs deep feature extraction on the fusion feature vector, and outputs an encoded feature representation, which retains the correlation between the partition content and the environment features and realizes information aggregation of multi-modal features.

[0097] The feature mapping layer receives the encoded feature representation, and performs partition-level optimization processing on the brightness and color features using an adaptive adjustment mechanism for dynamic changes of the partition content and the environment features, outputs a feature mapping result, and realizes local adaptive color adjustment and brightness balance.

[0098] The decoding layer receives the feature mapping result, performs upsampling and restoration on the feature mapping result, and outputs the partition enhanced content data. All partition enhanced content data is numbered and stored, and the partition enhanced content data meets the target brightness balance and color optimization requirements and has the characteristics of seamless splicing with adjacent partitions.

[0099] In the embodiment, the partition discriminator receives the partition enhanced content data, and outputs the true and false discrimination score, the brightness consistency evaluation score, the color consistency evaluation score and the style and structure consistency determination result, which specifically includes:

[0100] The partition discriminator receives the partition enhanced content data, and the partition discriminator includes an input feature extraction layer, a local discrimination layer and a consistency judgment layer.

[0101] The input feature extraction layer receives the partition enhanced content data, performs multi-scale feature extraction on the partition enhanced content data, and outputs a partition multi-scale feature representation;

[0102] The local discrimination layer receives the partition multi-scale feature representation, calculates a true-false discrimination score based on a discrimination network, and outputs the true-false discrimination score, which is used to determine whether the partition enhanced content data is real content data or generated content data;

[0103] The consistency judgment layer receives the partition multi-scale feature representation, calculates a brightness consistency evaluation score, a color consistency evaluation score, and a style and structure consistency determination result. The brightness consistency evaluation score is calculated based on a mean square error formula between the partition enhanced content data and a target brightness distribution. The color consistency evaluation score is calculated based on a correlation coefficient between a color distribution of the partition enhanced content data and a target color distribution. The style and structure consistency determination result is calculated based on a distance or a similarity score in a high-dimensional feature space between the partition enhanced content data and a reference style sample;

[0104] The true-false discrimination score, the brightness consistency evaluation score, the color consistency evaluation score, and the style and structure consistency determination result are outputted.

[0105] In the embodiment, the construction of the multi-objective loss optimization function and the joint optimization of the partition CycleGAN model parameters to obtain the trained partition CycleGAN model specifically include:

[0106] The multi-objective loss optimization function of the partition CycleGAN model is set, and the multi-objective loss optimization function is a weighted combination of an adversarial loss function, a partition cycle consistency loss function, a partition brightness balance loss function, and a partition color consistency loss function:

[0107] L total =λ1L adv +λ2L cyc +λ3L lum +λ4L col ;

[0108] Wherein, L total is the multi-objective loss optimization function, L adv is the adversarial loss function, L cyc is the partition cycle consistency loss function, L lum is the partition brightness balance loss function, L col is the partition color consistency loss function, and λ1, λ2, λ3, and λ4 are the weighting coefficients of each loss term, respectively;

[0109] The true and false discrimination scores based on the partition discriminator output are used to construct an adversarial loss function, specifically: each partition discriminator outputs a probability of the real content and the generated content:

[0110]

[0111] wherein N is the number of partitions of the LED screen, k is the partition index, x k is the real partition display content data of the kth partition, is the partition enhanced content data of the kth partition, p data (x k ) is the probability distribution of the real partition display content data of the kth partition, is the probability distribution of the partition enhanced content data of the kth partition, D k is the kth partition discriminator;

[0112] The adversarial loss function formula is derived on the basis of the generative adversarial network (GAN) loss function. The global discriminator and the generator in the original formula are expanded to a partition discriminator and a partition generator for each LED screen partition. The original GAN adversarial loss function is refined to a partition-level loss and averaged by partition. All components are probabilities or logarithmic probabilities, which are dimensionless. The formula has consistent physical dimensions and conforms to common sense.

[0113] A partition cycle consistency loss function is constructed based on the consistency of the partition content. On the basis of the CycleGAN cycle consistency loss function, it is expanded to a partition-level independent calculation:

[0114]

[0115] wherein C k is the kth partition display content data, G k is the kth partition generator, F k is the kth partition inverse generator, which is used to map the partition enhanced content data to the partition display content data, and ||·||1 is the L1 norm, representing the sum of absolute value distances of the corresponding partition data at each pixel point (i.e., the sum of absolute errors of all pixel points);

[0116] The partition cycle consistency loss function formula is derived based on the CycleGAN cycle consistency loss formula in combination with the LED screen partition scenario. The input and output are refined to partition display content data and partition enhanced content data. A forward and inverse generator is independently set for each partition. Finally, the global partition average loss is aggregated, wherein all components are pixel values or normalized pixels, which are dimensionless. The dimensions on both sides of the formula are consistent, which conforms to common sense.

[0117] The partition luminance balance loss function is constructed based on the luminance consistency evaluation score of the generated content, and the error between the luminance of the generated content and the target luminance is measured:

[0118]

[0119] wherein, is the luminance mean value of the kth partition enhanced content data, is the target luminance value of the kth partition;

[0120] The partition luminance balance loss function formula is derived from the mean square error (MSE) formula in mathematics, and the prediction and target value are respectively expanded into the luminance mean value of the partition enhanced content data and the target luminance value of the partition, and the average is taken on all partitions; wherein, and Both are luminance, both are luminance, and the square of the difference is the square of the luminance. After averaging the partitions, the dimension is unchanged, and the left and right dimensions are completely consistent.

[0121] The partition color consistency loss function is constructed based on the color consistency evaluation score of the generated content, and the error between the color distribution of the generated content and the target color distribution is measured:

[0122]

[0123] wherein, is the red channel mean value of the kth partition enhanced content data, is the target red channel mean value of the kth partition, is the green channel mean value of the kth partition enhanced content data, is the target green channel mean value of the kth partition, is the blue channel mean value of the kth partition enhanced content data, is the target blue channel mean value of the kth partition;

[0124] The partition color consistency loss function formula is derived from the mean square error (MSE) formula in mathematics, and the prediction and target value are respectively expanded into the RGB channel mean value of the partition enhanced content data and the target color channel mean value of the partition, and the average is taken on all partitions and channels; wherein, and Both are color component mean values, and the unit is normalized color value, which is dimensionless. After the square of the difference, the dimension is the square of the color component. Finally, the sum and average do not affect the dimension, and the left and right are consistent.

[0125] The partition CycleGAN model parameters are jointly optimized by back propagation using the multi-objective loss optimization function, so that the partition generator and the partition discriminator converge in terms of authenticity, consistency, brightness balance and color optimization, and the training is iterated until the multi-objective loss optimization function converges or meets the preset condition, and the trained partition CycleGAN model is output.

[0126] The preset condition includes, but is not limited to, one or more of the following situations: the variation amplitude of the multi-objective loss optimization function in consecutive iterations is lower than a preset threshold, the partition CycleGAN model training reaches a preset maximum training round, or the performance indicator of the partition CycleGAN model on the validation set does not improve in consecutive iterations.

[0127] In this embodiment, the seamless splicing of the partition enhanced content data to generate the overall enhanced display content data, and outputting the overall enhanced display content data to the LED screen driving system specifically includes:

[0128] According to the physical structure partition index of the LED screen, all partition enhanced content data are sequentially sorted to determine the corresponding spatial position of each partition enhanced content data on the screen, and the sorted partition enhanced content data set is output;

[0129] The boundary pixels of each partition enhanced content data in the sorted partition enhanced content data set and the boundary pixels of adjacent partition enhanced content data are subjected to brightness and color smoothing processing, and pixel-level weighted average, boundary fusion or other seamless splicing algorithms are adopted to generate the partition enhanced content data set subjected to seamless fusion processing, and the fusion processing ensures that there is no obvious mutation at the boundary between partitions, and realizes natural transition of brightness and color between partitions;

[0130] The partition enhanced content data set subjected to seamless fusion processing is spliced according to its corresponding spatial position to generate the overall enhanced display content data, and the overall enhanced display content data has brightness balance and color consistency in the full screen range;

[0131] The overall enhanced display content data is output to the LED screen driving system to realize dynamic brightness balance display of the LED screen.

[0132] In this embodiment, the continuous acquisition of the actual display effect and environmental changes of each partition to obtain partition feedback data, and the dynamic optimization of the partition CycleGAN model parameters specifically includes:

[0133] During the dynamic display of the LED screen, the actual display effect data of each partition is collected in real time, and the actual display effect data includes the brightness value, color distribution, smoothness evaluation result of the display content boundary transition area and real-time physical environment parameters of the partition currently output by the partition;

[0134] The actual display effect data and the partition environment parameter data are synchronously compared to generate partition feedback data. The partition feedback data includes: the deviation between the actual brightness of the partition and the target brightness, the difference between the actual color distribution of the partition and the target color distribution, and the change in partition environment parameters.

[0135] The partition feedback data is used as online feedback input to the trained partition CycleGAN model. The parameters of the partition CycleGAN model are corrected in real time by combining the partition brightness equalization loss function and the partition color consistency loss function.

[0136] By iteratively updating the parameters of the partition CycleGAN model, the partition enhancement content data output by the partition generator continuously matches the actual display requirements, improving the brightness balance and color consistency of the LED screen under different environmental changes.

[0137] Example 1:

[0138] To verify the feasibility of this invention in practice, it was applied to the daily operation of a large outdoor high-resolution LED advertising screen. This scenario involves commercial advertising in an outdoor plaza. The LED screen covers an area of ​​over 100 square meters and is divided into 96 physical zones (12×8). Each zone is equipped with ambient light, temperature, and brightness sensors to collect real-time environmental parameters. Daily advertising content includes high-definition video, dynamic graphics, and real-time weather data. The screen has extremely high requirements for brightness uniformity and color consistency, especially under changing light conditions (morning and evening) and in rainy weather, necessitating adaptive dynamic adjustment.

[0139] In this application scenario, the system first performs partitioned modeling of the LED screen's physical structure and then fuses the real-time environmental parameters of each partition with the displayed content into the partition CycleGAN model. The partition generator dynamically enhances the brightness and color features of different areas and time periods, while the discriminator outputs real-time true / false discrimination scores, brightness consistency evaluation scores, and color consistency evaluation scores. The model is optimized through a multi-objective loss function. All partition enhanced content data undergoes pixel-level seamless stitching, and the overall display content is uniformly output to the drive system. The display effect is monitored in real time through a feedback channel, and the partition CycleGAN parameters are fine-tuned online based on the monitoring data to ensure the system's adaptability to changes in the external environment.

[0140] Comparative testing over a period of actual operation clearly demonstrates the advantages of this invention's solution in terms of brightness balance, color consistency, and seamless splicing. Using the standard deviation of screen partition brightness as a metric, the standard deviation of brightness across different partitions is reduced to 2.1 cd / m² after adopting this invention's method. 2, while the standard deviation of the traditional global brightness equalization scheme is 5.8cd / m 2 In terms of color consistency, the mean difference of the partition color space is controlled within ΔE = 1.9, which is significantly better than the ΔE = 4.5 of the traditional method. For special weather and strong light environment, the dynamic brightness response time of the system is reduced from 7.2 seconds to 2.8 seconds. The user visual score consistency of the actual playing content seam area is improved from 73.5% to 95.2%. At the same time, the proposed scheme shows good stability under dynamic scene and high-frequency switching content, and there is no obvious lag or delay in long-time running. The following table is the comparison data of some key performance.

[0141] Table 1 Comparison of LED screen brightness equalization and color consistency test data

[0142]

[0143] As can be seen from the above table, the proposed scheme has certain improvement in the key performance indicators of LED screen brightness equalization and color consistency compared with the traditional global equalization scheme. First, in terms of the standard deviation of the partition brightness, the proposed scheme achieves an excellent level of 2.1cd / m 2 , which is much lower than the 5.8cd / m 2 of the traditional scheme, fully demonstrating that the partition dynamic color algorithm can effectively suppress the brightness unevenness between partitions and improve the overall display balance of the screen. In the test of the mean difference ΔE of the color space, the proposed method is only 1.9, while the traditional method is 4.5, which reflects the high-efficiency constraint ability of the proposed scheme on color consistency under multi-partition and dynamic scene, and reduces the color difference problem that can be perceived by the naked eye of the audience. In terms of the response ability to dynamic environment and strong light interference in actual use, the dynamic brightness response time of the system using the proposed scheme is shortened to 2.8 seconds, which is better than the 7.2 seconds of the traditional scheme, and enhances the adaptability of the LED screen to external environment changes.

[0144] In terms of seamless display in the joint area and user experience, the application also performs outstandingly. Through pixel-level seamless splicing and boundary smoothing algorithm, the visual score consistency of the joint area reaches 95.2%, which is significantly better than 73.5% of the traditional global balancing scheme, greatly improving the visual comfort and overall screen feeling of users in actual viewing. Further, in terms of long-time running average stability and extreme environment adaptability, the application achieves high scores of 97.4 and 93.6% respectively, which are higher than 85.9 and 68.2% of the traditional scheme. The average color saturation in rainy weather is also improved to 91.8%, and the robustness of the display system to complex environmental changes is stronger. In addition, the convergence period of the adaptive adjustment of the partition is shortened to 5.1 seconds, which is much lower than 13.6 seconds of the traditional scheme, indicating that the model of the application can complete adaptive adjustment and parameter optimization faster, effectively supporting real-time display requirements in high-frequency switching content and complex application scenarios.

[0145] From the above data analysis, it can be seen that the LED screen brightness equalization method based on the dynamic color algorithm proposed by the application has made technical progress in improving the partition balancing, color consistency, environmental adaptability and user experience, and can meet the high-standard and high-demand LED display screen engineering application, providing strong technical support for the intelligent and fine upgrading of the industry.

[0146] The above describes only the preferred specific embodiments of the application, but the protection scope of the application is not limited thereto, and any person skilled in the art can make equivalent replacements or changes to the technical solutions and inventive concepts of the application within the technical scope disclosed by the application, which should be covered within the protection scope of the application.

Claims

1. A method for equalizing the brightness of an LED screen based on a dynamic color adjustment algorithm, characterized in that, Includes the following steps: The LED screen's display content is divided into multiple zones based on its physical structure to obtain the zone display content data; For each zone, the ambient light intensity, temperature and actual brightness of the corresponding area are collected in real time to obtain the zone's environmental parameter data; Feature fusion is performed on the partition display content data and partition environment parameter data to generate partition joint feature data; The joint feature data of the partitions is input into the partition CycleGAN model, and the partition-enhanced content data is output. The enhanced content data of each zone is seamlessly spliced ​​to generate overall enhanced display content data, which is then output to the LED screen driver system. Continuously collect the actual display effect and environmental changes of each partition to obtain partition feedback data, and use the partition feedback data as online feedback input to the partition CycleGAN model to dynamically optimize the parameters of the partition CycleGAN model; The step of inputting the joint feature data of the partitions into the partition CycleGAN model and outputting the partitioned enhanced content data specifically includes: The joint feature data of each partition is input into the partition CycleGAN model, which includes a partition generator and a partition discriminator. The partition generator receives the partition joint feature data for each partition and generates partition-enhanced content data; The partition discriminator receives partition enhancement content data and outputs true / false discrimination score, brightness consistency evaluation score, color consistency evaluation score, and style and structure consistency judgment result respectively. Based on the true / false discrimination score, brightness consistency evaluation score, color consistency evaluation score, and style and structure consistency judgment results, a multi-objective loss optimization function is constructed, and the parameters of the partitioned CycleGAN model are jointly optimized to obtain the trained partitioned CycleGAN model. Using the trained partition CycleGAN model, inference processing is performed on the joint feature data of the partitions to output partition-enhanced content data, and the partition-enhanced content data is numbered and stored. The multi-objective loss optimization function is a weighted combination of the adversarial loss function, the partition cyclic consistency loss function, the partition brightness equalization loss function, and the partition color consistency loss function.

2. The LED screen brightness equalization method based on dynamic color adjustment algorithm according to claim 1, characterized in that, The partition display content data includes pixel data of the original image within the partition, spatial location information of the partition on the LED screen, resolution parameters of the partition, average brightness value, average color temperature value, color histogram, contrast value, and timestamp of partition content acquisition.

3. The LED screen brightness equalization method based on dynamic color adjustment algorithm according to claim 1, characterized in that, The partitioned environmental parameter data includes the ambient light intensity value, ambient temperature value, actual brightness value, spatial location information of the partition in the LED screen, and the timestamp of the partitioned environmental parameter collection.

4. The LED screen brightness equalization method based on dynamic color adjustment algorithm according to claim 1, characterized in that, The step of fusing the partition display content data and partition environment parameter data to generate joint partition feature data specifically includes: Extract the original image pixel data, spatial location information, resolution parameters, average brightness value, average color temperature value, color histogram, contrast value, and the timestamp of partition content collection from the partition display content data; Extract the ambient light intensity value, ambient temperature value, actual brightness value, spatial location information, and the timestamp of the partition environmental parameter data from the partition environmental parameter data; The partition display content data and partition environment parameter data are matched according to the partition spatial location information and timestamp to ensure that the data corresponds one-to-one; The partition display content data and all parameters in the partition environment parameter data are concatenated to generate partition joint feature data, and all partition joint feature data are numbered and stored according to the partition index.

5. The LED screen brightness equalization method based on dynamic color adjustment algorithm according to claim 1, characterized in that, The partition generator receives the joint partition feature data for each partition and generates partition-enhanced content data, specifically including: The partition generator receives joint feature data for each partition, and the partition generator includes an input feature fusion layer, an encoding layer, a feature mapping layer, and a decoding layer; The input feature fusion layer receives the joint feature data of each partition, performs feature-level concatenation and normalization of the partition display content data and the corresponding partition environment parameter data, and outputs the fused feature vector. The coding layer receives the fused feature vector, performs deep feature extraction on the fused feature vector, and outputs the coded feature representation; The feature mapping layer receives the encoded feature representation and, in response to the dynamic changes in the partition content and environmental features, adopts an adaptive adjustment mechanism to perform partition-level optimization of brightness and color features, and outputs the feature mapping results. The decoding layer receives the feature mapping results, upsamples and restores the feature mapping results, outputs the partitioned enhanced content data, and numbers and stores all the partitioned enhanced content data.

6. The LED screen brightness equalization method based on dynamic color adjustment algorithm according to claim 1, characterized in that, The partition discriminator receives partition-enhanced content data and outputs true / false discrimination scores, brightness consistency evaluation scores, color consistency scores, and style and structure consistency judgment results, specifically including: The partition discriminator receives partition-enhanced content data, and the partition discriminator includes an input feature extraction layer, a local discriminant layer, and a consistency judgment layer; The input feature extraction layer receives partitioned enhanced content data, performs multi-scale feature extraction on the partitioned enhanced content data, and outputs a multi-scale feature representation of the partition. The local discriminant layer receives multi-scale feature representations of the partitions and calculates true / false scores based on the discriminant network. The consistency judgment layer receives the multi-scale feature representation of the partition, calculates the brightness consistency evaluation score, the color consistency evaluation score, and the style and structure consistency judgment result; Output the true / false discrimination score, brightness consistency evaluation score, color consistency evaluation score, and style and structure consistency judgment results.

7. The LED screen brightness equalization method based on dynamic color adjustment algorithm according to claim 1, characterized in that, The specific steps of constructing a multi-objective loss optimization function and jointly optimizing the parameters of the partitioned CycleGAN model to obtain the trained partitioned CycleGAN model include: Define the multi-objective loss optimization function for the partitioned CycleGAN model; An adversarial loss function is constructed based on the true / false discrimination scores output by the partition discriminator, specifically: the probability output of each partition discriminator for real content and generated content; Based on the consistency of partition content, a partition-based cyclic consistency loss function is constructed, which is extended to be calculated independently at the partition level based on the CycleGAN cyclic consistency loss function. A partition brightness equalization loss function is constructed based on the brightness consistency evaluation score of the partition-generated content to measure the error between the brightness of the partition-generated content and the target brightness. A partition color consistency loss function is constructed based on the color consistency evaluation score of the partitioned generated content to measure the error between the color distribution of the partitioned generated content and the target color distribution. Using a multi-objective loss optimization function, the parameters of the partitioned CycleGAN model are jointly optimized through backpropagation. The model is iteratively trained until the multi-objective loss optimization function converges or meets the preset conditions, and the trained partitioned CycleGAN model is output.

8. The LED screen brightness equalization method based on dynamic color adjustment algorithm according to claim 1, characterized in that, The process of seamlessly stitching together the partitioned enhanced content data to generate overall enhanced display content data, and then outputting the overall enhanced display content data to the LED screen driving system, specifically includes: Based on the physical structure partition index of the LED screen, sort all partition enhancement content data in sequence, determine the corresponding spatial position of each partition enhancement content data on the screen, and output the sorted partition enhancement content dataset. Brightness and color smoothing are performed on the boundary pixels of each partition enhancement content data in the sorted partition enhancement content dataset and the boundary pixels of adjacent partition enhancement content data to generate a seamlessly fused partition enhancement content dataset. The seamlessly integrated partitioned enhanced content datasets are combined according to their corresponding spatial locations to generate overall enhanced display content data; The overall enhanced display content data is output to the LED screen driver system.

9. The LED screen brightness equalization method based on dynamic color adjustment algorithm according to claim 1, characterized in that, The process of continuously collecting the actual display effects and environmental changes of each partition, obtaining partition feedback data, and dynamically optimizing the partition CycleGAN model parameters specifically includes: During the dynamic display process of the LED screen, the actual display effect data of each zone is collected in real time; The actual display effect data includes the brightness value, color distribution, smoothness evaluation results of the transition area of ​​the display content boundary, and real-time physical environment parameters of the partition; The actual display effect data is compared synchronously with the partition environment parameter data to generate partition feedback data. The partition feedback data includes the deviation between the actual brightness and the target brightness of the partition, the difference between the actual color distribution and the target color distribution of the partition, and the change in the partition environmental parameters; The partitioned feedback data is used as online feedback input to the trained partitioned CycleGAN model. The parameters of the partitioned CycleGAN model are adjusted in real time by combining the partitioned brightness equalization loss function and the partitioned color consistency loss function.

Images