Intelligent correction method and system for outdoor LED multi-screen spliced content

Abstract

The present application relates to the technical fields of multi-screen splicing content, and discloses an intelligent correction method and system for outdoor LED multi-screen splicing content, which comprises the following steps: acquiring real-time illumination and an initial display image, extracting dynamic illumination and standard color features after gray-scale processing, splicing the features into a feature vector and classifying the feature vector by SVM to obtain an environment perception category, and matching an illumination mapping table to obtain illumination intensity.If the intensity exceeds a threshold value, then collect difference data of each display unit screen and denoise the data, use random forest to regress the difference data and the environment category to obtain color correction values, update the pixel values of the initial display image according to the color correction values, and obtain a target image through edge enhancement.Obtain viewing condition feedback, and if the feedback is below a threshold value, then re-collect illumination data.The present application realizes adaptive color correction of multi-screen splicing under complex illumination.The present method can realize adaptive color correction of outdoor multi-screen splicing display under complex illumination.

Description

Technical Field

[0001] This invention relates to the field of multi-screen splicing content technology, and in particular to an intelligent correction method and system for outdoor LED multi-screen splicing content. Background Technology

[0002] Currently, in the field of outdoor LED displays, multi-screen splicing systems are widely deployed in outdoor advertising, live sports broadcasts, and public information dissemination scenarios due to their ability to provide large-size, high-resolution visual presentations. However, outdoor ambient lighting conditions vary drastically, from strong midday sunlight to low nighttime illumination. In addition, the inherent brightness and color differences between display units due to variations in manufacturing batches and aging processes pose significant challenges to the color consistency and visual coherence of multi-screen spliced ​​images.

[0003] In existing technologies, content calibration for outdoor LED multi-screen splicing often employs offline calibration methods based on fixed parameters or manual adjustment of each screen based on subjective experience. This primarily focuses on the color reproduction of individual display units under standard conditions, using preset gamma curves and white balance parameters for calibration. When splicing display devices are deployed outdoors and face real-time changes in ambient light, offline calibration methods with fixed parameters cannot dynamically respond to sudden changes in light intensity, and the differences in optical attenuation among display units are difficult to accurately compensate for through manual subjective adjustments. Furthermore, in high-end industrial scenarios such as fiber optic manufacturing, outdoor splicing screens are often used to remotely display production line operating status and product quality data in real time, requiring extremely high levels of color accuracy, consistency, and recognizability. This further amplifies the shortcomings of existing calibration methods in terms of dynamic environmental adaptability and multi-unit consistency compensation.

[0004] In summary, existing technologies lack a collaborative mechanism for ambient light perception and multi-display unit difference compensation, making it difficult to achieve adaptive color correction for outdoor multi-screen splicing displays under complex lighting conditions, resulting in poor color consistency. Summary of the Invention

[0005] This invention provides an intelligent correction method and system for outdoor LED multi-screen splicing content, so as to realize adaptive color correction of outdoor multi-screen splicing display under complex lighting conditions.

[0006] In a first aspect, to solve the above-mentioned technical problems, the present invention provides an intelligent correction method for the content of outdoor LED multi-screen splicing, comprising: The system acquires real-time illumination data and initial display image of the splicing display device, converts the initial display image to grayscale to obtain a grayscale image, and extracts dynamic illumination features and standard color features based on the real-time illumination data and the grayscale image. The dynamic illumination features and the standard color features are concatenated to obtain a feature vector. The feature vector is then classified using a preset support vector machine model to obtain an environmental perception category. The light intensity value is obtained by matching the environmental perception category with a preset illumination mapping table. If the light intensity value is greater than the preset light intensity threshold, then the screen difference data of each display unit in the splicing display device is obtained, and the screen difference data is denoised to obtain the target difference data. The target difference data and the environmental perception category are processed using random forest to obtain color correction values. The pixel values ​​of the initial display image are updated and the edges are enhanced according to the color correction values ​​to obtain the target display image. Obtain the viewing condition feedback value corresponding to the target display image. If the viewing condition feedback value is lower than the preset feedback threshold, generate a state matching failure signal and reacquire real-time illumination data.

[0007] Secondly, the present invention provides an intelligent correction system for the content of outdoor LED multi-screen splicing, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the intelligent correction method for the content of outdoor LED multi-screen splicing as described above.

[0008] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the intelligent correction method for outdoor LED multi-screen splicing content as described above.

[0009] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention acquires real-time illumination data and an initial display image, converts the initial display image to grayscale to obtain a grayscale image, and extracts dynamic illumination features and standard color features based on the real-time illumination data and the grayscale image. This series of operations establishes a perceptual bridge between environmental optical information and display content. By using sliding window variance calculation to capture transient changes in illumination and combining it with two-dimensional discrete cosine transform to extract color presentation features, it achieves real-time quantitative perception of complex outdoor illumination environments and solves the problem that fixed parameter correction in the prior art cannot respond to dynamic changes in illumination.

[0010] (2) This invention concatenates dynamic illumination features and standard color features to obtain feature vectors, uses a preset support vector machine model for classification to obtain environmental perception categories, and matches them with a preset illumination mapping table to obtain illumination intensity values. Through the hyperplane spatial mapping of the support vector machine and the parsing of environmental state labels, continuous illumination perception data can be discretized into environmental perception categories with clear physical meanings, providing a categorical decision basis for subsequent difference compensation and color correction, and realizing adaptive classification of environmental perception and accurate quantification of illumination intensity.

[0011] (3) When the light intensity value exceeds the preset light intensity trigger threshold, the present invention acquires the screen difference data of each display unit and performs noise reduction processing. A preset random forest algorithm model is used to process the target difference data and environmental perception category to obtain the color correction value. Based on this, the pixel value is updated and the edge is enhanced in the initial display image. The multi-decision tree integration mechanism of random forest effectively suppresses the interference of single-point noise on the correction value. Combined with edge enhancement processing, it improves color consistency while maintaining the clarity of image details, thus significantly improving the visual coherence of multi-screen splicing. Attached Figure Description

[0012] Figure 1 This is a schematic flowchart of the intelligent correction method for outdoor LED multi-screen splicing content provided in the first embodiment of the present invention; Figure 2 This is a heat map showing the original brightness deviation distribution of each display unit in the splicing display device provided in the first embodiment of the present invention; Figure 3 This is a heatmap of brightness deviation distribution after bilateral filtering and noise reduction of screen difference data provided in the first embodiment of the present invention; Figure 4 This is a grayscale image of the abnormal region edge location after edge detection of the smooth transition matrix provided in the first embodiment of the present invention. Detailed Implementation

[0013] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0014] Reference Figure 1 The first embodiment of the present invention provides an intelligent correction method for the content of outdoor LED multi-screen splicing, including the following steps: S11, acquire real-time illumination data and initial display image of the splicing display device, convert the initial display image to grayscale to obtain a grayscale image, and extract dynamic illumination features and standard color features based on the real-time illumination data and the grayscale image; S12, the dynamic illumination features and the standard color features are concatenated to obtain a feature vector, the feature vector is classified using a preset support vector machine model to obtain an environmental perception category, and the light intensity value is obtained by matching the environmental perception category with a preset illumination mapping table. S13, if the light intensity value is greater than the preset light intensity threshold, then obtain the screen difference data of each display unit in the splicing display device, and denoise the screen difference data to obtain the target difference data. S14, Random forest is used to process the target difference data and the environmental perception category to obtain color correction values. The pixel values ​​of the initial display image are updated and edge enhancement is performed according to the color correction values ​​to obtain the target display image. S15, obtain the viewing condition feedback value corresponding to the target display image. If the viewing condition feedback value is lower than the preset feedback threshold, generate a state matching failure signal and re-acquire real-time illumination data.

[0015] In step S11, real-time illumination data and initial display image of the splicing display device are acquired, the initial display image is converted to grayscale to obtain a grayscale image, and dynamic illumination features and standard color features are extracted based on the real-time illumination data and the grayscale image.

[0016] In one implementation, the method includes: Ambient light intensity values ​​are collected by an optical sensor as real-time illumination data, and the current frame is extracted from the video stream buffer of the splicing display device as the initial display image based on the ambient light intensity values. The initial display image is parsed to obtain an initial pixel matrix, and a weighted average is calculated from the initial pixel matrix to obtain a grayscale image.

[0017] In one implementation, the method further includes: Dynamic illumination features are obtained by performing sliding window variance calculation on the real-time illumination data; A two-dimensional discrete cosine transform is performed on the grayscale image to obtain color rendering features. If the color rendering features are greater than a preset color threshold, the color rendering features are scaled to obtain initial normalized features. The initial normalized features are matched with a preset color lookup table to obtain standard color features.

[0018] The video wall display is deployed in an outdoor plaza. Its frame is equipped with a multi-channel photoelectric sensor array as an optical sensor, with a sampling frequency set to 10Hz. The ambient light intensity value reflects the overall illumination level of the outdoor scene, measured in lux. Inside the optical sensor, the ambient light signal is converted from analog to digital into a digital light intensity value and then reported in real-time to the control terminal of the display device via an RS485 bus. Upon receiving the ambient light intensity value, the system retrieves the current display frame from the video stream buffer of the video wall display device based on an internal clock counter, using it as the initial display image. The resolution of the initial display image is consistent with the physical pixel array of the video wall display device; in this embodiment, it is 1920×1080. The initial display image contains three channels: red, green, and blue, with each channel's pixel value ranging from 0 to 255.

[0019] A weighted average is used to convert the initial pixel matrix to grayscale. The input consists of the red, green, and blue channel values ​​of each pixel in the initial pixel matrix. A linear weighted sum is then applied, with a weight of 0.299 for the red channel, 0.587 for the green channel, and 0.114 for the blue channel, outputting a single-channel grayscale image matrix. The weights 0.299, 0.587, and 0.114 are derived from the relative contribution ratios of each channel to the human visual response in the CIE 1931 color matching function, which are industry-recognized standard parameters for image processing. Simultaneously, histogram statistical analysis is performed on the grayscale image to determine the proportion of bright pixels with grayscale values ​​between 200 and 255, thus obtaining the brightness mapping characteristics.

[0020] The preset threshold for the proportion of bright pixels was 0.50. In the laboratory, 30 observers with normal vision were invited to view the same set of test images under simulated outdoor light intensities. The observers used a 5-point Likert scale to rate their perception of the overall brightness of the image (1 point indicates significantly dark, and 5 points indicates significantly bright). A score of ≥4 points was defined as the overall brightness of the image. The proportion of bright pixels corresponding to the observer's judgment that the overall image was bright was recorded. The 50th percentile of the 30 sets of data was taken as the threshold, and the result was 0.50.

[0021] When the brightness mapping feature exceeds the threshold, gamma correction is applied to the grayscale image using a non-linear mapping. A gamma value of 1.2 is used. Different gamma values ​​(0.8, 1.0, 1.2, 1.5, and 2.0) are applied to correct 30 typical overexposed images. Thirty observers subjectively rate the detail of the corrected image, and the gamma value of 1.2, with the highest overall score, is used as the fixed parameter. This value effectively darkens overexposed areas while preserving details in shadows. If the brightness mapping feature is less than or equal to the preset highlight ratio threshold, the image is determined not to be overexposed overall, and the gamma correction step is skipped. The grayscale image is then directly used as the target grayscale image in the subsequent two-dimensional discrete cosine transform process.

[0022] The sliding window has a length of 50 sampling points, corresponding to a data length of 5 seconds. Within each sampling period, the variance of the 50 ambient light intensity values ​​within the window is calculated to obtain dynamic illumination characteristics. These dynamic illumination characteristics reflect the degree of fluctuation in outdoor illumination over a short period of time.

[0023] The output of the two-dimensional discrete cosine transform (DCT) is a coefficient matrix reflecting different frequency components of the image. The upper left corner of the coefficient matrix corresponds to the low-frequency components (reflecting the overall brightness and darkness distribution of the image), and the lower right corner corresponds to the high-frequency components (reflecting image details and texture). The low-frequency component coefficients of the upper left 8×8 region are extracted from the coefficient matrix. The absolute values ​​of these low-frequency component coefficients are summed, and the global maximum value of the summation of the absolute values ​​of the DCT low-frequency coefficients of 50 typical outdoor display images is 3.85×10⁻⁶. 4 As the normalization denominator, the value is normalized to the 0-100 range using a linear mapping formula. This value represents the color rendering characteristic, reflecting the overall contrast energy of the current image in the spatial frequency dimension. A preset color threshold of 60 is obtained through statistical analysis of the color rendering characteristics of 50 typical outdoor display images (including sunny, cloudy, and rainy scenes) under laboratory conditions. This threshold is the mean of the color rendering characteristics of the 50 images minus one standard deviation, corresponding to the lower limit boundary where image changes in outdoor scenes significantly affect the display effect. When the color rendering characteristic exceeds this preset color threshold, minimum and maximum scaling are used to linearly map the color rendering characteristic to the 0-100 range, outputting the initial normalized characteristic.

[0024] It should be noted that the minimum and maximum scaling values ​​are calculated by taking the historical minimum value of the color presentation characteristics (0 in this embodiment based on 50 typical outdoor display images) and the historical maximum value of the color presentation characteristics (100 in this embodiment based on the aforementioned 50 images). This value range boundary is stored in the control terminal memory and serves as the input domain parameter for each scaling operation. When the actual input exceeds this value range, it is clamped according to the boundary value.

[0025] If the color rendering feature is less than or equal to the preset color threshold of 60, it is determined that the overall contrast energy of the current image in the spatial frequency dimension is low. The color rendering feature is processed according to the above normalization formula to obtain the initial normalized feature. The minimum and maximum scaling steps are skipped, and the preset color lookup table matching process is directly entered.

[0026] The preset color lookup table is a two-dimensional lookup table calibrated in the laboratory using a spectrometer. The horizontal axis represents the initial normalized characteristic (value range 0 to 100), and the vertical axis represents the standard color characteristic (value range 0 to 100). The lookup table relationship reflects the nonlinear mapping relationship between the input normalized characteristic and the target color response of the display device under standard viewing conditions. For a single splicing display device of the same model, test patterns with different initial normalized characteristics were applied under a standard D65 light source. The target color response was measured using a spectrometer, constructing no fewer than 100 data pairs. A monotonically increasing nonlinear mapping curve was obtained by fitting using the least squares method and discretized into a lookup table stored in the control terminal memory.

[0027] During runtime, the current initial normalized features are input into the lookup table, and the corresponding standard color features are obtained through linear interpolation, completing the entire data processing flow of step S11. This feature extraction and normalization mechanism effectively preserves the key state information of ambient lighting and the displayed image, providing highly discriminative input for subsequent classification models. For example, some mapping data pairs in the preset color lookup table include: initial normalized feature 0 corresponding to standard color feature 0, initial normalized feature 50 corresponding to standard color feature 48, and initial normalized feature 100 corresponding to standard color feature 100. These mapping data pairs are only used to verify the consistency of the reproduction results.

[0028] In step S12, the dynamic illumination features and the standard color features are concatenated to obtain a feature vector. A preset support vector machine model is used to classify the feature vector to obtain an environmental perception category. The illumination intensity value is then obtained by matching the environmental perception category with a preset illumination mapping table. This includes: The feature vector is obtained by zero-padding alignment of the dynamic lighting features and the standard color features; A set of hyperplane distances is obtained by spatial mapping of the feature vectors using a preset support vector machine model, and the environmental state label is obtained by querying a preset category mapping matrix based on the set of hyperplane distances. If the environmental status label meets the preset perception conditions, the environmental status label is parsed to obtain the environmental perception category; The corresponding light intensity value is obtained by looking up and matching the light intensity value in a preset light mapping table according to the environmental perception category.

[0029] Since dynamic illumination characteristics are scalar values, the unit is lux. 2 Standard color features are scalar values, dimensionless. A zero-padding alignment algorithm is used to concatenate two one-dimensional features in a predetermined order, and two zeros are added at the end to expand the feature dimension to four dimensions, outputting a 4-dimensional feature vector. Zero-padding alignment is a mature and commonly used technique in machine learning data preprocessing, and will not be elaborated further. The pre-defined support vector machine (SVM) model uses a radial basis function (RBF) as the kernel function, with a penalty parameter of 2 and a kernel parameter of 0.2. For nonlinearly separable problems, the kernel function maps the original features to a high-dimensional Hilbert space. The pre-training data for the pre-defined SVM model comes from real lighting scene samples accumulated during the deployment of splicing display devices in at least 10 commercial outdoor plazas in North China, East China, and South China. Each sample consists of a 4D feature vector synchronously collected by optical sensors and environmental perception categories (5 categories in total: clear skies, cloudy, uniform overcast, twilight transition, and low light at night) labeled by on-site engineers based on meteorological records, forming a training set of at least 3000 samples. After Z-score standardization of the training set (recording the mean and standard deviation of each feature dimension and storing them in the control terminal memory), the 4D feature vector input during the inference stage adopts the same standard as that used during the training stage. Z-score standardization was performed using the same mean and standard deviation to ensure consistency between the training and inference feature spaces. A grid search and 5-fold cross-validation were employed to find the coarse combination with the highest average classification accuracy within a coarse grid range with penalty parameter values ​​[0.1, 1, 10, 100] and kernel parameter values ​​[0.01, 0.1, 1, 10] (the optimal point of the coarse grid is C=1, γ=0.1). Then, a fine-grained search was performed in the neighborhood of the optimal point of the coarse grid (C values ​​[0.5, 1, 2, 5], γ values ​​[0.05, 0.1, 0.2, 0.5]). The final model parameters were fixed at C=2 and γ=0.2, and the model achieved a classification accuracy greater than 92% on the test set. The 4-dimensional feature vectors were input into a pre-defined support vector machine model. The model outputs a hyperplane distance for each environmental perception category, resulting in five distance values ​​arranged in order of category number to form a hyperplane distance set. It should be noted that a positive hyperplane distance indicates that the current feature vector is located on the positive side of the hyperplane of the corresponding class, and the larger the value, the higher the confidence of belonging to that class; a negative value indicates that it is located on the negative side.

[0030] The hyperplane distance set is then input into a preset category mapping matrix for querying. The preset category mapping matrix is ​​a two-dimensional matrix with dimension N×6, where N is the number of historical samples. In this embodiment, N=5000. The historical samples are composed of the hyperplane distance output of each validation sample and the corresponding category label when the preset support vector machine model performs 5-fold cross-validation on the training set during the pre-training stage. The first to fifth columns of each record are the standard values ​​of the hyperplane distance of the five environmental perception categories corresponding to the sample, and the sixth column is the corresponding environmental state label (an integer from 1 to 5). By calculating the Euclidean distance between the current hyperplane distance set and each row vector in the matrix, the value in the sixth column corresponding to the row with the smallest Euclidean distance is taken as the environmental state label output.

[0031] It should be noted that the preset support vector machine model itself can directly determine the category based on the maximum hyperplane distance. In this embodiment, a preset category mapping matrix is ​​further introduced for secondary matching. The purpose is to use the statistical distribution of historical samples to correct the confidence of the initial classification result of SVM: when the confidence margin of the maximum hyperplane distance of SVM (the difference between the maximum distance and the second largest distance) is greater than 0.3, the argmax output of SVM is used as the environment state label; when the confidence margin is less than or equal to 0.3 (i.e., fuzzy samples near the SVM classification boundary), the nearest neighbor output of the mapping matrix is ​​used as the standard, and the statistical robustness of historical samples is used to suppress misclassification caused by instantaneous noise.

[0032] The preset perception condition is that the identifier code of the environmental state label is within the integer range of 1 to 5. This range corresponds one-to-one with the 5 environmental perception categories defined in the pre-training phase of the preset support vector machine model. It is a fixed index boundary during model training, and there are no label outputs with values ​​outside this range. When the environmental state label is any integer from 1 to 5, a dictionary key-value pair extraction algorithm is used to parse the environmental state label. That is, the label identifier code is used as the key to query the pre-stored environmental perception category dictionary and output the corresponding environmental perception category name. In this embodiment, label 1 corresponds to clear skies without clouds, label 2 corresponds to cloudy skies with obstruction, label 3 corresponds to overcast skies with uniform light, label 4 corresponds to dawn / dusk transition, and label 5 corresponds to low light at night.

[0033] The preset illumination mapping table is a two-dimensional lookup table obtained in advance through outdoor field testing and calibration. The horizontal index represents the environmental perception category (5 categories in total), and the vertical index represents the corresponding typical illumination intensity value (in lux). The calibration process of the lookup table relationship is as follows: for each environmental perception category, at no less than 12 outdoor display device sites in different geographical locations and seasons, at least 100 sets of ambient light intensity values ​​measured by optical sensors are continuously collected during the typical illumination period of that category (e.g., sunny and cloudless corresponding to 11:00 AM to 1:00 PM). The median of the 100 sets of data for each category is taken as the typical illumination intensity value of that category. Finally, 5 sets of mapping relationships are fixed and stored in the preset illumination mapping table. In this embodiment, sunny and cloudless corresponds to an illumination intensity value of 80,000 lux, cloudy and obscured corresponds to an illumination intensity value of 450 lux, cloudy and uniform corresponds to an illumination intensity value of 5,000 lux, dawn and dusk corresponds to an illumination intensity value of 1,500 lux, and low light at night corresponds to an illumination intensity value of 50 lux. The light intensity value serves as the input for subsequent trigger determination and regression calculation.

[0034] In step S13, if the light intensity value is greater than a preset light intensity threshold, then the screen difference data of each display unit in the splicing display device is acquired, and the screen difference data is denoised to obtain the target difference data. This includes: The difference between the light intensity value and the preset light intensity trigger threshold is calculated to obtain the intensity difference value. If the intensity difference value is greater than the preset fluctuation amplitude threshold, the environmental perception category is compared with the pre-established static configuration category to obtain the category difference feature. The category difference feature is subjected to difference mapping processing to obtain the matching degree score. If the matching degree score is greater than the preset matching degree threshold, the initial pixel brightness matrix and initial color deviation vector of each display unit in the splicing display device are obtained. The initial pixel brightness matrix and the initial color deviation vector are fused to obtain screen difference data; The screen difference data is subjected to bilateral filtering to obtain a smooth transition matrix, and the smooth transition matrix is ​​subjected to edge detection to obtain target difference data.

[0035] The difference between the light intensity value and the preset light intensity trigger threshold is calculated, and the absolute value of the difference is the intensity difference. The preset light intensity trigger threshold is determined based on the calibration benchmark of the splicing display device at the factory stage under a standard D65 light source (color temperature 6500K), with a typical value of 400 lux, corresponding to the lower limit of the light intensity at which outdoor display devices need to trigger color correction; this threshold is fixed in the control terminal memory at the factory stage as a static reference benchmark.

[0036] The algorithm further employs a sliding window variance calculation to examine the intensity difference sequence within a historical time window (e.g., the most recent 30 sampling points, corresponding to a 3-second duration), outputting the fluctuation amplitude value. This fluctuation amplitude value reflects the stability of illumination over a short period. The preset fluctuation amplitude threshold is 10 lux. 2 (With dimensions consistent with the intensity difference variance), it was obtained through statistical analysis of intensity difference sequences under various outdoor lighting scenarios in the laboratory. The operational definition of the sunny and stable scenario is an outdoor scenario where the ambient light intensity variation coefficient (standard deviation divided by mean) is less than 0.1 within 30 consecutive minutes and the cloud cover is less than 30% throughout the day. The boundary between sunny and partly cloudy (cloud cover 30% to 70%) is defined by the cloud cover index. No less than 1,000 intensity difference sequences were collected under the sunny and stable scenario, and the 95th percentile of its fluctuation amplitude was taken as the preset fluctuation amplitude threshold, which can filter out false triggers caused by small natural fluctuations in lighting.

[0037] When the fluctuation amplitude value exceeds a preset fluctuation amplitude threshold, the environmental perception category is compared with the pre-established static configuration category to obtain category difference features. The determination that the intensity difference exceeds the preset fluctuation limit is implemented in this embodiment as a two-layer judgment: first, the intensity difference of the current frame is calculated; then, the fluctuation amplitude value is calculated using the intensity difference sequence of the most recent 30 sampling points. The condition is only satisfied when the intensity difference of a single frame is greater than 0 and the fluctuation amplitude value of the sliding window is greater than the preset fluctuation amplitude threshold.

[0038] Conversely, if the fluctuation amplitude is less than or equal to the preset fluctuation amplitude threshold, it is determined that the current illumination fluctuation is within the range of natural small fluctuations, and the current display state is maintained without entering the comparison process between the environmental perception category and the static configuration category.

[0039] It should be noted that the comparison operation here involves two layers of judgment. The first layer judges whether the intensity difference of a single frame is greater than 0 lux (i.e., the illumination is indeed higher than the trigger threshold). The second layer judges whether the fluctuation amplitude of the sliding window is greater than the preset fluctuation amplitude threshold of 10 (lux). 2 (That is, the light fluctuation has reached the level that needs to be triggered), and only when both layers of judgment are satisfied will the subsequent category difference feature comparison process begin.

[0040] The static configuration category is determined by the on-site engineer during the initial power-on and debugging of the splicing display equipment. This data is written into the control terminal memory and permanently stored based on annual climate statistics released by the local meteorological bureau. If the annual precipitation is greater than 800mm and the annual sunshine duration is less than 1800h, it is classified as cloudy / occupied; if the annual precipitation is less than or equal to 800mm and the annual sunshine duration is greater than or equal to 1800h, it is classified as clear / cloudless. This avoids subjective judgment discrepancies. Both the environmental perception category and the static configuration category are converted into 5-dimensional one-hot encoded vectors. The five components of the one-hot encoded vector correspond to the five environmental perception categories (component 1 corresponds to clear / cloudless, component 2 to cloudy / occupied, component 3 to uniformly cloudy, component 4 to dawn / dusk transition, and component 5 to low light at night). For each environmental perception category, its corresponding component is set to 1, and the remaining components are set to 0.

[0041] Subsequently, the cosine distance between the two one-hot encoded vectors in the 5-dimensional semantic space is calculated, outputting a two-dimensional feature vector (the first dimension is the semantic space cosine distance, and the second dimension is the one-hot encoded Hamming distance). This two-dimensional feature vector is the class difference feature. The preset difference mapping matrix is ​​a 2×1 column vector, with two components representing the weights corresponding to the semantic space cosine distance and Hamming distance, respectively. The weight values ​​are obtained through regression of historical outdoor multi-scene data. In this embodiment, the weights are set to 85 and 15, respectively, corresponding to the relative importance ratio of these two components in distinguishing significant class differences from minor class differences.

[0042] Under laboratory conditions, semantic cosine distance and Hamming distance were calculated for 50 samples with known class significance. Using the significance of expert ratings as labels, a binary linear regression using the least squares method was performed, resulting in a weight ratio of 85:15. The specific evaluation criteria for the significance of expert ratings were as follows: independent ratings were given by three engineers with over five years of experience in outdoor display technology, using a 1-10 scale. A score of 1 indicates complete equality between the two classes, and a score of 10 indicates a significant difference between the two classes. The arithmetic mean of the three engineers' ratings was used as the significance score for that sample. A score ≥7 was considered significant, a score of 4-6 was considered moderate, and a score ≤3 was considered slight. Multiplying the class difference feature by the difference mapping matrix yielded a scalar, which was the matching score.

[0043] The preset matching threshold is 80 and the state lasts for more than 3 detection cycles (corresponding to 3 × sampling cycle, which is 300ms in this embodiment). The matching score threshold of 80 is the lower boundary of significant category difference obtained from outdoor multi-scene data statistics (the 75th percentile value of 50 samples). The duration constraint is used to filter out instantaneous category jumps caused by transient light jitter.

[0044] If the matching score is less than or equal to the preset matching threshold, or if the matching score is greater than the preset matching threshold but the duration is less than 3 detection cycles, it is determined that the current environmental change is insufficient to trigger color correction, the current display state is maintained, and the process re-enters the next round of S11 data acquisition and processing.

[0045] When the matching score is greater than the preset matching threshold, grayscale histogram statistics are performed on the current frame image of each display unit, and an initial pixel brightness matrix with dimension W×H is output, where W and H are the horizontal and vertical numbers of display units in the splicing display device (in this embodiment, 16 display units in 4×4 are used as an example). The matrix elements are the average brightness statistics of the current frame image of the corresponding display unit. The average brightness statistics are obtained by averaging the grayscale values ​​of all pixels in the current frame image of the display unit and normalizing them to the range of 0 to 100. The current color temperature data of each display unit is collected synchronously and converted to a preset reference color temperature (typically 6500K, corresponding to a standard D65 light source). That is, the RGB color space is converted to the CIE 1931 XYZ space, and then to the CIE Lab space. The color difference distance of each display unit relative to the reference white point in the Lab space is calculated, and an initial color deviation vector with dimension W×H×2 is output according to the two color channels a* and b*. Each display unit position corresponds to a set of a* deviation and b* deviation values.

[0046] Since the original value range of the a* and b* channels in CIELab space is -128 to 127 when encoded with 8-bit integers (corresponding to the 8-bit color depth configuration of the splicing display device in this embodiment), and the brightness statistics have been normalized to the range of 0 to 100, in order to unify the dimensions, the a* deviation and b* deviation are both normalized to the range of 0 to 100 by using a linear mapping algorithm. The normalized value is equal to the sum of the original value and 128, divided by 255, and then multiplied by 100. The normalized a* deviation and b* deviation are consistent with the brightness statistics in terms of numerical distribution.

[0047] Subsequently, the initial pixel brightness matrix and the normalized initial color deviation vector are fused along the channel dimension and concatenated by channel to output a screen difference data tensor with dimensions W×H×3, where W and H are the number of display units horizontally and vertically, respectively, and 3 is the number of channels, corresponding to brightness statistics, a* deviation normalized value, and b* deviation normalized value.

[0048] Bilateral filtering is applied to the screen difference data. In this embodiment, the standard deviation of the spatial kernel function is set to 3 pixels, and the standard deviation of the domain kernel function is set to 15, outputting a smooth transition matrix. Then, the Laplacian operator is used to calculate the second-order gradient of the smooth transition matrix, outputting abnormal fluctuation values. If the abnormal fluctuation value exceeds a preset abnormal fluctuation threshold, Cannibal edge detection is used, with a high threshold of 60 and a low threshold of 24 (i.e., the low threshold is 0.4 times the high threshold). The Sobel operator kernel size is set to 3×3 to locate edges in the smooth transition matrix, outputting edge-preserving features. The output of the Laplacian operator after applying it to the smooth transition matrix is ​​a second-order gradient matrix with the same dimension as the input. The maximum absolute value of all elements in this matrix is ​​taken as the abnormal fluctuation value.

[0049] The preset abnormal fluctuation threshold is 30, dimensionless, corresponding to the second-order gradient magnitude of the Laplacian operator applied to an 8-bit grayscale image normalized to 0-100. Fifty sets of typical outdoor screen difference data were collected under laboratory conditions. Experts labeled the actual physical edge locations. Abnormal fluctuation values ​​were calculated for each set of data, and the 5th percentile of the abnormal fluctuation value corresponding to the actual edge location was used as the threshold, ensuring effective identification of the actual edges. Finally, a morphological opening and closing operation algorithm was used for secondary purification of the edge-preserving features, removing isolated discrete pixels and outputting the target difference data.

[0050] The morphological opening and closing operation uses a 3×3 square structuring element. First, the opening operation is performed (erosion followed by dilation to remove small-sized noise), and then the closing operation is performed (dilation followed by erosion to fill small holes inside the edge). This can remove small-sized noise while filling small holes inside the edge, and output continuous target difference data.

[0051] If the abnormal fluctuation value is less than or equal to the preset abnormal fluctuation threshold, it is determined that the smooth transition matrix does not contain significant physical edges. The Canney edge detection step is skipped, and the smooth transition matrix is ​​directly sent as an edge-preserving feature into the morphological opening and closing operation.

[0052] like Figure 1 As shown, this is the original brightness deviation diagram in this embodiment. Figure 2 As shown in the figure, this diagram compares the distribution of brightness deviations in each display unit before and after screen difference data denoising. The comparison clearly shows that after the bilateral filtering and edge detection processing, the brightness deviations of the display units, which were originally randomly distributed over a large range (approximately -25 to +25) and exhibited severe local abnormal peaks, are significantly reduced. Figure 1 The random distribution of blue to red color patches was effectively smoothed and converged to the gradual transition interval between neighboring units. Figure 2 (The distribution of light tones tends to be more uniform), and the peak amplitude of abnormally bright areas is significantly reduced; furthermore, such as Figure 3 As shown, after performing Canny edge detection and morphological opening and closing operations on the smooth transition matrix, the real physical edges that were originally mixed in with noise are accurately extracted and purified (a clear bipartite distribution of black edge contours and gray-white smooth areas), and isolated discrete pixels are effectively removed. This result intuitively proves that this scheme can efficiently denoise and preserve the edges of the screen difference data of each display unit when the illumination intensity triggers the correction condition, and transform the original brightness deviation containing a lot of noise interference into target difference data with continuous edge features, providing a solid data foundation for subsequent random forest regression calculation of color correction values.

[0053] In step S14, the target difference data and the environmental perception category are processed using random forest to obtain color correction values. The pixel values ​​of the initial display image are updated and edge enhancement is performed based on the color correction values ​​to obtain the target display image.

[0054] In one implementation, the method includes: Based on the environmental perception category, ambient light intensity and color temperature distribution features are extracted, and tensor concatenation is performed on the ambient light intensity, the color temperature distribution features and the target difference data to obtain a multidimensional feature input matrix; The multidimensional feature input matrix is ​​processed using a pre-defined random forest algorithm model to obtain multiple decision tree models and the initial prediction weights corresponding to each decision tree model; If the initial prediction weight is lower than the preset weight threshold, the corresponding initial prediction weight is removed to obtain a set of valid prediction weights. The color correction value is obtained by weighting and summing the outputs of the multiple decision tree models based on the set of effective prediction weights.

[0055] In one implementation, the method further includes: The initial display image is linearly interpolated based on the color correction values ​​to obtain the second display image; The second displayed image is subjected to color space conversion processing to obtain a basic brightness matrix, and the basic brightness matrix is ​​subjected to edge detection to obtain gradient magnitude and edge direction; If the gradient magnitude is greater than a preset gradient threshold, then high-frequency detail features are extracted based on the edge direction; The target display image is obtained by performing Laplacian sharpening on the second display image based on the high-frequency detail features.

[0056] Ambient light intensity refers to the light intensity value obtained by looking up the preset light intensity mapping table in step S12, with the unit being lux. Both are the same physical quantity. The typical color temperature distribution characteristics (unit being K) are obtained by looking up the preset color temperature mapping table according to the environmental perception category. The calibration process of the preset color temperature mapping table is as follows: in the laboratory, for outdoor scenes under typical light conditions corresponding to each environmental perception category, a spectrophotometer (e.g., Konica Minolta CL-200A) is used to collect no less than 100 sets of measured color temperature data in no less than 12 different geographical locations and seasons. The median of the 100 sets of data for each category is taken as the typical color temperature distribution characteristics of that category. Finally, 5 sets of mapping relationships are solidified and stored in the preset color temperature mapping table.

[0057] It should be noted that the preset color temperature mapping table is a preset resource necessary for executing step S14. It is constructed and fixed in the control terminal memory once according to the above calibration process during the equipment manufacturing stage or the first debugging stage, and serves as the standard basis for the subsequent step S14 to extract color temperature distribution characteristics.

[0058] In this embodiment, a clear, cloudless day corresponds to a color temperature of 6500K, a cloudy day corresponds to a color temperature of 5500K, a cloudy day with uniform light corresponds to a color temperature of 6000K, a day between dawn and dusk corresponds to a color temperature of 4000K, and a nighttime day with low light corresponds to a color temperature of 3000K. Ambient light intensity, color temperature distribution characteristics, and target difference data are concatenated using tensors along the channel dimension, outputting a multidimensional feature input matrix with dimensions W×H×5. The five channels are: brightness statistics, a* deviation normalized value, b* deviation normalized value, ambient light intensity normalized value, and color temperature distribution characteristic normalized value.

[0059] Ambient light intensity is linearly normalized to the 0-1 range based on a typical outdoor value of 0 to 100,000 lux. The normalized value of ambient light intensity is equal to the ambient light intensity divided by 100,000. Color temperature distribution features are linearly normalized to the 0-1 range based on a typical outdoor value of 1500 to 10,000 K. The normalized value of color temperature distribution features is equal to the difference between the color temperature distribution feature value and 1500, divided by 8500. Both are used as scalars and copied to fill all pixel locations, resulting in a multi-dimensional feature input matrix.

[0060] The default random forest algorithm in this embodiment includes 100 decision trees, each with a maximum depth of 8. At each node split, features are randomly sampled at a ratio of 0.7 of the total number of features. During training, each tree samples with replacement at a ratio of 0.8 of the total number of samples. These hyperparameters are determined through grid search based on the principle of minimizing the mean squared error of 5-fold cross-validation. The specific search range includes the number of decision trees [50, 100, 200], maximum depth [4, 8, 12], feature sampling ratio [0.5, 0.7, 0.9], and sample sampling ratio [0.6, 0.8, 1.0], ultimately determining the hyperparameter combination.

[0061] During pre-training, a single splicing display device of the same model was subjected to known color correction values ​​(labels) under different environmental perception categories, light intensities, and color temperatures in a laboratory setting. Simultaneously, a multi-dimensional feature input matrix (input) was collected, forming a training set of no less than 5000 samples. The CART algorithm was used to train each decision tree independently, with the regression criterion being minimizing the mean squared error. Simultaneously, an independent benchmark validation subset (using out-of-bag data) was constructed for each decision tree. The mean squared error of the prediction output of each decision tree was calculated, and the reciprocal of the mean squared error was normalized to the range of 0 to 100 as the initial prediction weights for that decision tree model. Higher weights indicate better prediction reliability. Training was implemented using the RandomForestRegressor from the scikit-learn library. After training, the structure and weight parameters of all decision trees, as well as the initial prediction weights for each decision tree, were fixed.

[0062] The preset tree weight threshold is 60. 5000 training samples are divided into a training set and a validation set (4:1 ratio). This 4:1 split is independent of the OOB mechanism of the random forest and is specifically used for calibrating the preset tree weight threshold. The 4000 training samples are used for decision tree growth and the calculation of initial OOB prediction weights. The 1000 validation samples are used to independently calculate the prediction error of each decision tree. The weight value corresponding to the 75th percentile of the prediction error of all decision trees on the validation set is taken as the preset tree weight threshold (i.e., retaining approximately 75% of decision trees with lower prediction errors, about 80 trees). This filters out the 25% of low-quality decision trees with the largest prediction errors, preventing them from interfering with the final regression results. The OOB mechanism and the 4:1 split are built-in evaluation mechanisms of the random forest and do not consume additional training samples. The 4:1 split is used to independently calibrate the preset tree weight threshold and is an external validation mechanism. These are independent evaluation mechanisms for different purposes and do not interfere with each other. Traverse the initial prediction weights of 100 decision trees, remove decision trees with weights below 60 from the effective set, and retain decision trees with weights above 60 (approximately 80 in this embodiment), and output the effective prediction weight set.

[0063] If the initial prediction weight is greater than or equal to the preset tree weight threshold, the corresponding decision tree is directly retained in the effective set, and together with the decision trees with weights greater than 60, they constitute the effective prediction weight set.

[0064] The multidimensional feature input matrix is ​​averaged along the W×H dimension to obtain a 1×5 global feature vector, which is then input into the decision tree model. Each decision tree outputs a color correction value (ranging from 0 to 100). The output of each decision tree is multiplied by its corresponding initial prediction weight, and then all the results are summed and divided by the sum of the weights. The resulting scalar is the color correction value, which is 12 in this embodiment.

[0065] The color correction values ​​reflect the required color compensation intensity for the current display image and serve as the basis for subsequent pixel value updates. It should be noted that in S13, independent difference data for each display unit in three dimensions—brightness, a* deviation, and b* deviation—are collected. In this embodiment, the difference characteristics of all display units are aggregated by mean during the regression phase to reflect the overall color deviation trend of the splicing display device.

[0066] The RGB three-channel values ​​of each pixel in the initial pixel matrix of the initial displayed image are linearly fused with the color correction values. For each pixel, the R, G, and B channels are linearly interpolated using the corresponding color correction values. The R channel is calculated according to... G channel according to Channel B ,in , , The three-dimensional color correction vector components are obtained by averaging and aggregating the a* deviation, b* deviation, and brightness deviation of each display unit in S13, and the results are cropped to the range of 0 to 255. The resulting pixel matrix is ​​the second display image. This process completes the alignment of the global color reference, so that the overall color presentation of the display screen matches the current environmental perception category.

[0067] The channel correspondences Δa*→R channel, Δb*→G channel, and ΔL→B channel in the above pixel update formula are simplified engineering mappings based on experimental calibration. Experimental verification under a standard D65 light source shows that the chromaticity deviation of the a* axis (green-red versus the opposite axis) in CIELab space has the most significant impact on the human eye's perception of the red channel on LED display devices, the deviation of the b* axis (blue-yellow versus the opposite axis) has the most significant impact on the perception of the green channel, and the deviation of the L axis (luminance) has the most significant impact on the perception of the blue channel. In verification experiments covering five environmental perception categories, the above three sets of axis-channel correspondences can reduce the average color difference ΔE00 of multi-screen spliced ​​images from 6.8 before correction to 1.9 after correction. This engineering mapping relationship is stored in the control terminal memory and is used for lookup in the table during pixel value update operations.

[0068] A color space conversion algorithm is used to convert the second display image from RGB space to YCbCr space, and the Y channel (luminance channel) data is extracted as the basic luminance matrix. The dimensions of the basic luminance matrix are consistent with those of the second display image (1920×1080 in this embodiment). The Sobel edge detection operator is used to perform a two-dimensional convolution operation on the basic luminance matrix with a 3×3 convolution kernel. The first-order partial derivatives in the horizontal and vertical directions are calculated respectively, and the gradient magnitude and edge direction corresponding to each pixel are output. The gradient magnitude is the square root of the sum of the squares of the horizontal and vertical partial derivatives, and the edge direction is the arctangent function value.

[0069] With a preset gradient threshold of 45, 30 adult observers (aged 22 to 45) with normal or corrected vision (visual acuity ≥ 1.0) and no color vision abnormalities were invited to the laboratory to subjectively rate the edge sharpness of each set of 50 typical outdoor images containing four categories of content: text, people, scenery, and dynamic video frames. A 5-point Likert scale was used, with 1 point = completely unclear, 2 points = unclear, 3 points = moderate, 4 points = clear, and 5 points = very clear. The gradient amplitude of each pixel was calculated simultaneously at a standard viewing distance of 10m. The 50th percentile of the gradient amplitude corresponding to the clear and distinguishable score was taken as the threshold.

[0070] The gradient magnitude of each pixel is iterated, and pixels with a gradient magnitude greater than a preset threshold are marked as significant edge pixels. A non-maximum suppression (NMS) algorithm is then applied to these edge pixels to filter local features along the edge direction, eliminating redundant non-edge pixels and outputting a set of high-frequency detail features. NMS is a mature technique in image edge detection that ensures each edge retains only the single pixel with the largest response in the gradient direction, preventing edges from becoming thicker.

[0071] If the gradient magnitude is less than or equal to the preset edge gradient threshold, the corresponding pixel is determined to be a non-significant edge pixel. The high-frequency detail feature extraction and Laplacian sharpening steps are skipped, and the corresponding pixel in the updated display image directly retains its original value and enters the target display image.

[0072] The Laplacian sharpening algorithm is used to perform pixel-level fusion calculations on the second display image. High-frequency details are extracted from the image using a second-order Laplacian operator and then superimposed back onto the original image with a preset sharpening coefficient (0.5 in this embodiment). The sharpening coefficient of 0.5 is an empirical value; those skilled in the art can adjust it between 0.1 and 0.5 based on actual results. Laplacian sharpening is performed only on pixels within the high-frequency detail feature set; other pixels retain their original values. The final output is the target display image. This target display image serves as the final driving signal and is sent to the splicing display panel for display.

[0073] In step S15, the viewing condition feedback value corresponding to the target display image is obtained. If the viewing condition feedback value is lower than a preset feedback threshold, a state matching failure signal is generated and real-time illumination data is reacquired. This includes: Acquire the viewer’s facial video stream synchronously captured during the playback of the target display image, and perform facial region localization on the viewer’s facial video stream to obtain the region of interest for the eyes; The pupil dynamic change rate is obtained by contour fitting of the region of interest of the eye, and the viewing condition feedback value is obtained by linear mapping of the pupil dynamic change rate. If the viewing condition feedback value is lower than the preset feedback threshold, a state matching failure signal is generated. Based on the state matching failure signal, a light distribution acquisition instruction is constructed, and the photoelectric sensor array is driven to reacquire real-time illumination data according to the light distribution acquisition instruction.

[0074] A high-definition industrial camera is deployed directly in front of the video display device, sampling at 30fps, to synchronously acquire video stream data of the viewer's face during the playback of the target display image. The Haar feature classification algorithm is used to locate the facial region in each frame of the video stream. The input is a grayscale image of the video frame, and the output is the bounding box coordinates of the facial region. The cascaded classifier model used in the Haar feature classification algorithm employs the OpenCV open-source frontal face Haar feature cascade model. The model is pre-trained on the publicly available face dataset MIT-CBCL and includes no fewer than 20 classification stages, each using different combinations of Haar features and thresholds, ensuring a detection rate greater than 95% while keeping the false positive rate below 5%. After facial region localization, based on facial anatomy priors, the human eye is typically located in the upper 1 / 3 of the face, and the region of interest for the eye is extracted proportionally from the facial bounding box.

[0075] The Hough circle transform algorithm was used to fit the pupil contour to the region of interest (ROI) of the eye. The input was a grayscale image of the eye region, and the output was the coordinates of the pupil center and the radius. The radius search range of the Hough circle transform was set to 5 to 20 pixels, corresponding to the pupil imaging size at different distances. The accumulator threshold was set to 50. In the laboratory, no less than 6,000 frames of eye image data were collected from 30 observers with normal vision in an outdoor display scenario with a standard outdoor viewing distance of 10m and an ambient light intensity of 400 to 80,000 lux. 200 frames were collected from each observer, including normal pupil contraction, dilation, and slight deviation. Three engineers manually labeled the actual pupil position in each frame. The 5th percentile value of the Hough accumulator response distribution corresponding to all actual pupil positions was taken as the lowest confidence detection threshold, which was 50.

[0076] Within N consecutive frames (N=30 in this embodiment, corresponding to a duration of 1 second), the standard deviation of the pupil radius sequence detected in each frame is calculated. The standard deviation divided by the mean pupil radius yields the pupil dynamic change rate, reflecting the amplitude of pupil accommodation when viewing the target displayed image. A larger pupil dynamic change rate indicates a more significant pupil accommodation response caused by the light stimulus of the current displayed image, corresponding to a poorer viewing comfort. A linear mapping algorithm is used to map the pupil dynamic change rate to the range of 0 to 100 to obtain the viewing condition feedback value. The maximum and minimum values ​​of the mapping are derived from the statistical extreme values ​​of pupil dynamic change rate of 30 observers with normal vision under a standard outdoor display screen in the laboratory.

[0077] It should be noted that when the Hough circle transform fails to detect a valid pupil circle satisfying the accumulator threshold of 50 in a certain frame (e.g., the viewer is blinking or the pupil is obscured by the eyelid due to rapid eye saccades), that frame is not included in the pupil radius sequence statistics. Simultaneously, the sampling window is shifted forward one frame to maintain the statistical frame count of 30, ensuring consistency in the statistical basis for standard deviation and mean calculations. If pupil detection fails for more than 5 consecutive frames (corresponding to approximately 167ms), it is determined that the viewer's eyes are not visible (e.g., a side profile or wearing sunglasses), the viewing condition feedback value is set to the preset median value of 50, and the process enters a degradation logic.

[0078] In this embodiment, the linear mapping relationship is achieved by multiplying the pupil dynamic change rate by 1000 and then cropping it to the range of 0 to 100. The general formula for calculating the mapping coefficient is that the mapping coefficient equals 100 divided by the 95th percentile of the statistical extreme value of the pupil dynamic change rate of the observer group. The 95th percentile of the statistical extreme value of the 30 observers in this laboratory is 0.1, which is dimensionless. Substituting this into the formula, the mapping coefficient is 1000. This statistical extreme value is mapped to the feedback value of 100, so that the range of the feedback value is consistent with the target range of 0 to 100. If there are differences in the extreme values ​​obtained from different experimental groups, the mapping coefficient can be adjusted accordingly (e.g., when the statistical extreme value is 0.08, the coefficient is taken as 1250). However, the fixed 1000 in this embodiment corresponds to the statistical results of the 30 observers in this laboratory. The obtained viewing condition feedback value objectively reflects the human eye's response to the light stimulus of the current display screen. It should be noted that the statistical robustness of the mapping coefficient 1000 is ensured as follows: In a group of 30 observers, a bootstrap method is used to resample 1000 times to calculate the 95% confidence interval of the 95th percentile, resulting in the interval [0.085, 0.115]. The mapping coefficient range corresponding to the center value of 0.1 is [870, 1176]. The integer value 1000 falls within this interval with a deviation of less than 5%, indicating that this fixed value has sufficient statistical robustness. If the characteristics of the actual viewer group in the deployment scenario differ significantly from those of the laboratory observer group (e.g., significantly different age distribution or vision status of the viewers), the calibration process can be re-executed and the mapping coefficient updated during the device initialization phase.

[0079] The preset viewing feedback threshold of 60 was obtained through ROC curve analysis based on the subjective ratings of 30 observers on viewing comfort and discomfort of the display screen under various outdoor lighting conditions, and the synchronously collected viewing condition feedback values. Specifically, the calibration process involved converting the binary label for viewing comfort and discomfort into a 5-level rating scale (VCS): 1 point represents extremely uncomfortable, 2 points represent uncomfortable, 3 points represent moderate, 4 points represent comfortable, and 5 points represent very comfortable. 30 observers rated the display screen under each lighting condition using the VCS. Scores ≥4 were converted to a viewing comfort label (positive example), and scores ≤3 were converted to a viewing discomfort label (negative example). The true positive rate and false positive rate were calculated for each data set, and an ROC curve was plotted. The viewing condition feedback value corresponding to the point on the curve that maximizes the Youden index (true positive rate minus false positive rate) was selected as the preset viewing feedback threshold. At this point, the true positive rate was 0.87, the false positive rate was 0.17, and the corresponding maximum Youden index was 0.70. Therefore, the preset viewing feedback threshold was 60.

[0080] The VCS scale uses a 4-point scale as the dividing line between viewing comfort and viewing discomfort. In the 5-level design of the VCS scale, a 3-point score (moderate) indicates a borderline neutral state where the observer cannot clearly determine whether the viewing experience of the current image is comfortable or uncomfortable. Classifying the neutral state as a negative example (discomfort) allows for a conservative triggering strategy. Comparative experiments show that when using a 3-point cutoff (>=3 for comfortable, <=2 for uncomfortable), the optimal ROC split point is 48 (Youden index 0.65), while using a 4-point cutoff point results in 60 (Youden index 0.70). The latter maintains a higher true positive rate of 0.87 while achieving a better Youden index, validating the rationality of the 4-point cutoff.

[0081] When the viewing condition feedback value is lower than the preset viewing feedback threshold, it indicates that the optical output of the current display screen has caused abnormal pupil adjustment, resulting in a mismatch with the actual viewing environment, and a state matching failure signal is immediately generated. The encoding format of the state matching failure signal is predefined in the initialization phase and includes three fields: mismatch level (2 bits), occurrence timestamp (32 bits), and current viewing condition feedback value (16 bits), which facilitates subsequent traceability and analysis.

[0082] Conversely, if the viewing condition feedback value is greater than or equal to the preset viewing feedback threshold, the current display state is maintained, the target display image is used as the output benchmark for the next round of dynamic adjustment of display parameters, no state matching failure signal is generated, and the system enters the waiting state for the next detection cycle.

[0083] If the high-definition industrial camera does not detect any viewer's face within a preset time window (60s in this embodiment), such as when no one passes by the equipment deployment area late at night or when there are few visitors in an outdoor plaza during inclement weather, the viewing condition feedback value cannot be calculated based on the pupil dynamic change rate. The viewing condition feedback value is set to 50, and the process of generating a state matching failure signal and re-acquiring the photoelectric sensor array in this step is skipped. The target display image is used as the output reference for the next round of dynamic adjustment of display parameters, and the system enters the waiting state for the next detection cycle until a new viewer's face is detected, at which point the complete closed-loop feedback process is restored.

[0084] It should be noted that the preset median value of 50 is based on the median of the viewing condition feedback values ​​corresponding to the viewing comfort label (VCS score >= 4) and the viewing discomfort label (VCS score <= 3) in historical closed-loop feedback data (72 and 38 respectively). The arithmetic mean of these two values ​​is calculated to be 55, which is then rounded to the nearest integer. Considering the system's conservative tendency in color correction in unattended scenarios—that is, to avoid the system failing to trigger a new round of environmental perception due to excessively high feedback values ​​and thus maintaining a potentially deviated display state—the preset median value was lowered from 55 to 50 to retain an appropriate safety margin, allowing the system to maintain a relatively stable correction output during degradation processing. This safety margin of 5 units corresponds to approximately 0.5 times the statistical standard deviation of the pupil dynamic change rate of a group of 30 observers, and has been verified to prevent display drift during degradation processing without excessively interfering with the normal closed loop.

[0085] The state matching failure signal is encapsulated using an instruction encoding algorithm to generate a light distribution acquisition instruction containing three fields: sensor addressing information (4 bytes, identifying the specific optical sensor number participating in the re-acquisition), sampling frequency (2 bytes, increased from the default 10Hz to 100Hz in this embodiment to improve response sensitivity), and sampling duration (2 bytes, set to 10s in this embodiment to cover a complete light fluctuation cycle). This instruction is sent to the multi-channel photoelectric sensor array deployed on the bezel of the splicing display device. The spatial illuminance scan is re-executed according to the sampling frequency and sampling duration specified in the instruction, and the updated real-time illumination data is output as the physical environment benchmark for the next round of dynamic adjustment of display parameters, forming a closed-loop optimization of the display effect.

[0086] It should be noted that when the viewing condition feedback value is still lower than the preset viewing feedback threshold after a complete cycle of S11 to S15, the next iteration begins. The iteration terminates when the viewing condition feedback value is greater than or equal to the preset viewing feedback threshold for two consecutive cycles, or when the number of iterations reaches the upper limit of 3. If the upper limit is not met, the current iteration result is fixed as the display parameter for this cycle and the loop is restarted in the next detection cycle (typical interval 60s). When S11 is re-executed, the initial display image is extracted from the video stream buffer of the splicing display device in real time, and all steps are re-executed according to the above process.

[0087] In summary, this invention discloses an intelligent correction method for outdoor LED multi-screen splicing content. By using support vector machines for adaptive classification of ambient light perception data and random forests for regression correction of differences in multi-display unit data, combined with feedback closed-loop adjustment based on the dynamic change rate of the viewer's pupils, this invention solves the technical problems of existing fixed-parameter offline calibration methods being unable to dynamically respond to complex outdoor lighting changes and lacking a collaborative compensation mechanism for optical attenuation differences and color deviations among multi-display units. This enables adaptive color correction of outdoor multi-screen splicing displays under complex lighting conditions.

[0088] The second embodiment of the present invention provides an intelligent correction system for the content of outdoor LED multi-screen splicing, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the intelligent correction method for the content of outdoor LED multi-screen splicing as described above.

[0089] It should be noted that the intelligent correction system for outdoor LED multi-screen splicing content provided in this embodiment of the invention is used to execute all the process steps of the intelligent correction method for outdoor LED multi-screen splicing content in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0090] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0091] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for intelligent correction of content in outdoor LED multi-screen splicing, characterized in that, include: The system acquires real-time illumination data and initial display image of the splicing display device, converts the initial display image to grayscale to obtain a grayscale image, and extracts dynamic illumination features and standard color features based on the real-time illumination data and the grayscale image. The dynamic illumination features and the standard color features are concatenated to obtain a feature vector. The feature vector is then classified using a preset support vector machine model to obtain an environmental perception category. The light intensity value is obtained by matching the environmental perception category with a preset illumination mapping table. If the light intensity value is greater than the preset light intensity threshold, then the screen difference data of each display unit in the splicing display device is obtained, and the screen difference data is denoised to obtain the target difference data. The target difference data and the environmental perception category are processed using random forest to obtain color correction values. The pixel values ​​of the initial display image are updated and the edges are enhanced according to the color correction values ​​to obtain the target display image. Obtain the viewing condition feedback value corresponding to the target display image. If the viewing condition feedback value is lower than the preset feedback threshold, generate a state matching failure signal and reacquire real-time illumination data.

2. The intelligent correction method for outdoor LED multi-screen splicing content according to claim 1, characterized in that, The step of acquiring real-time illumination data and an initial display image of the splicing display device, and converting the initial display image to grayscale to obtain a grayscale image, includes: Ambient light intensity values ​​are collected by an optical sensor as real-time illumination data, and the current frame is extracted from the video stream buffer of the splicing display device as the initial display image based on the ambient light intensity values. The initial display image is parsed to obtain an initial pixel matrix, and a weighted average is calculated from the initial pixel matrix to obtain a grayscale image.

3. The intelligent correction method for outdoor LED multi-screen splicing content according to claim 1, characterized in that, The step of extracting dynamic lighting features and standard color features based on the real-time lighting data and the grayscale image includes: Dynamic illumination features are obtained by performing sliding window variance calculation on the real-time illumination data; A two-dimensional discrete cosine transform is performed on the grayscale image to obtain color rendering features. If the color rendering features are greater than a preset color threshold, the color rendering features are scaled to obtain initial normalized features. The initial normalized features are matched with a preset color lookup table to obtain standard color features.

4. The intelligent correction method for outdoor LED multi-screen splicing content according to claim 1, characterized in that, The process of concatenating the dynamic illumination features and the standard color features to obtain a feature vector, classifying the feature vector using a preset support vector machine model to obtain an environmental perception category, and matching the environmental perception category with a preset illumination mapping table to obtain an illumination intensity value includes: The feature vector is obtained by zero-padding alignment of the dynamic lighting features and the standard color features; A set of hyperplane distances is obtained by spatial mapping of the feature vectors using a preset support vector machine model, and the environmental state label is obtained by querying a preset category mapping matrix based on the set of hyperplane distances. If the environmental status label meets the preset perception conditions, the environmental status label is parsed to obtain the environmental perception category; The corresponding light intensity value is obtained by looking up and matching the light intensity value in a preset light mapping table according to the environmental perception category.

5. The intelligent correction method for outdoor LED multi-screen splicing content according to claim 1, characterized in that, If the light intensity value is greater than a preset light intensity threshold, then screen difference data of each display unit in the splicing display device is acquired, and the screen difference data is denoised to obtain target difference data, including: The intensity difference is obtained by calculating the difference between the light intensity value and the preset threshold. If the intensity difference is greater than the preset fluctuation limit, the environmental perception category is compared with the pre-established static configuration category to obtain the category difference feature. The matching score is obtained by performing difference mapping processing on the category difference features. If the matching score is greater than the preset matching threshold, the initial pixel brightness matrix and initial color deviation vector of each display unit in the splicing display device are obtained. The initial pixel brightness matrix and the initial color deviation vector are fused to obtain screen difference data; The screen difference data is subjected to bilateral filtering to obtain a smooth transition matrix, and the smooth transition matrix is ​​subjected to edge detection to obtain target difference data.

6. The intelligent correction method for outdoor LED multi-screen splicing content according to claim 1, characterized in that, The process of using random forest to process the target difference data and the environmental perception category to obtain color correction values ​​includes: Based on the environmental perception category, ambient light intensity and color temperature distribution features are extracted, and tensor concatenation is performed on the ambient light intensity, the color temperature distribution features and the target difference data to obtain a multidimensional feature input matrix; The multidimensional feature input matrix is ​​processed using a pre-defined random forest algorithm model to obtain multiple decision tree models and the initial prediction weights corresponding to each decision tree model; If the initial prediction weight is lower than the preset weight threshold, the corresponding initial prediction weight is removed to obtain a set of valid prediction weights. The color correction value is obtained by weighting and summing the outputs of the multiple decision tree models based on the set of effective prediction weights.

7. The intelligent correction method for outdoor LED multi-screen splicing content according to claim 1, characterized in that, The initial display image is updated with pixel values ​​and edge enhancement is performed based on the color correction values ​​to obtain the target display image: The initial display image is linearly interpolated based on the color correction values ​​to obtain the second display image; The second displayed image is subjected to color space conversion processing to obtain a basic brightness matrix, and the basic brightness matrix is ​​subjected to edge detection to obtain gradient magnitude and edge direction; If the gradient magnitude is greater than a preset gradient threshold, then high-frequency detail features are extracted based on the edge direction; The target display image is obtained by performing Laplacian sharpening on the second display image based on the high-frequency detail features.

8. The intelligent correction method for outdoor LED multi-screen splicing content according to claim 1, characterized in that, The step of acquiring the viewing condition feedback value corresponding to the target display image, and generating a state matching failure signal and reacquiring real-time illumination data if the viewing condition feedback value is lower than a preset feedback threshold, includes: Acquire the viewer’s facial video stream synchronously captured during the playback of the target display image, and perform facial region localization on the viewer’s facial video stream to obtain the region of interest for the eyes; The pupil dynamic change rate is obtained by contour fitting of the region of interest of the eye, and the viewing condition feedback value is obtained by linear mapping of the pupil dynamic change rate. If the viewing condition feedback value is lower than the preset feedback threshold, a state matching failure signal is generated. Based on the state matching failure signal, a light distribution acquisition instruction is constructed, and the photoelectric sensor array is driven to reacquire real-time illumination data according to the light distribution acquisition instruction.

9. An intelligent correction system for outdoor LED multi-screen splicing content, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the intelligent correction method for the content of outdoor LED multi-screen splicing as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the intelligent correction method for the content of outdoor LED multi-screen splicing as described in any one of claims 1 to 8.

Images