A multi-screen spliced image color consistency correction method and system

Abstract

The application discloses a multi-screen spliced image color consistency correction method and system, and relates to the field of image processing. The multi-screen spliced image color consistency correction method comprises the following steps: collecting a plurality of image sequences of a spliced screen and performing geometric normalization to obtain a front view image sequence; extracting a seam area as a zero radiance reference; using the seam area as the zero radiance reference, combining a space-time consistent pixel block to estimate a camera inverse response function, and converting the front view image sequence into a linear radiance image sequence; defining a sampling band on both sides of the seam; extracting a cross-screen pixel pair based on a local flatness condition; and estimating a color mapping function from each to-be-corrected screen to a reference screen through multi-frame statistical regression. The application generates independent correction lookup tables for each screen through color mapping, and outputs a weighted mixed correction value of adjacent screens in a seam fusion band, so that the multi-screen color consistency correction is realized, and the multi-screen splicing can maintain stable and unified high-definition display effects for a long time.

Description

Technical Field

[0001] This invention relates to the field of image processing, specifically to a method and system for color consistency correction of multi-screen spliced ​​images. Background Technology

[0002] With the increasing demand for large-size displays, display walls composed of multiple LED displays are widely used in command and control centers, broadcast and television studios, large-scale exhibitions, and commercial advertising. By physically combining multiple independent screens, splicing screens achieve ultra-large screen and ultra-high resolution display effects at a relatively low cost.

[0003] During long-term use of splicing screens, each screen will exhibit varying degrees of light decay due to differences in cumulative lighting time, operating temperature, and the gradual changes in the material properties of the LED beads themselves. This decay is usually uneven across the three primary colors of red, green, and blue, leading to gradual differences in color performance between the screens. When one of the screens is replaced with a brand new backup screen of the same model due to a malfunction, the color difference between the new and old screens becomes particularly noticeable. When displaying pure colors or uniform images, the observer can clearly see the color boundary at the seam.

[0004] Among them, the limitations of the existing technology include at least the following problems: during the long-term operation of the splicing screen, when a screen is replaced with a brand new backup screen of the same model due to a fault, there is a significant color drift between the new screen and the old screen due to the different lighting time of the LED beads. The old screen has undergone light-induced decay after long-term use, and is generally warmer and darker, while the new screen retains its factory characteristics and is cooler and brighter. When displaying the same signal image, the actual light color emitted by the old and new screens under the same driving value has been separated. A clear color difference boundary can be seen at the seam in a pure color or large area of ​​uniform image. However, in situations where the continuity of the image is strictly required, such as command and control, broadcasting and television, it is difficult to interrupt the display at will to perform color correction. Furthermore, the color difference between the old and new screens is not limited to a certain brightness level. The degree of offset varies across different gray levels and hues, making it difficult to eliminate this multi-point color deviation at once while maintaining normal playback. This results in a persistent contradiction in actual operation and maintenance between ensuring uninterrupted business operations and the difficulty in eliminating multi-point color differences. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method and system for color consistency correction of multi-screen spliced ​​images, which solves the problem of simultaneously eliminating color differences at multiple points between new and old screens without interrupting business operations.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for color consistency correction of multi-screen spliced ​​images, comprising the following steps: acquiring a multi-frame image sequence of the spliced ​​screen and performing geometric normalization on the image sequence to obtain a front view image sequence, wherein the spliced ​​screen includes at least a reference screen and the remaining screens as screens to be corrected; extracting the seam region from the front view image sequence as a zero-radius reference, estimating the inverse response function of the camera using the zero-radius reference and detected spatiotemporally consistent pixel blocks, and applying the inverse response function to convert the front view image sequence into a linear radius image sequence; in the linear radius image sequence, defining sampling bands belonging to adjacent screens on both sides of the seam, extracting cross-screen pixel pairs from the sampling bands based on local flatness conditions, and estimating the color mapping function from each screen to be corrected to the reference screen through multi-frame statistical regression; generating an independent color correction lookup table for each screen to be corrected based on the color mapping function, and defining a fusion band on both sides of the seam; within the fusion band, weighted mixing of the color values ​​of adjacent screens after correction by their respective color correction lookup tables, as the final output, thereby achieving color consistency correction of multi-screen spliced ​​images.

[0007] Further, the specific steps for geometric normalization of the image sequence are as follows: Select one frame from the multi-frame image sequence and detect the straight lines formed by the seams; divide the detected straight lines into horizontal and vertical groups according to their angles, perform hierarchical clustering on each group of straight lines, and take the mean straight line of each cluster to obtain the horizontal and vertical seam lines respectively; determine the corner points of each screen by the intersection of the horizontal and vertical seam lines, and take the four outermost corner points as the source vertices of the perspective transformation; take the rectangle enclosed by the corner points of the reference screen as a reference, set the width and height of the target rectangle, and take the vertices of the target rectangle as the target vertices of the perspective transformation; use direct linear transformation to solve the homography matrix from the source vertex to the target vertex, and perform perspective transformation on each frame in the image sequence to obtain the front view image sequence.

[0008] Further, the specific steps for detecting spatiotemporally consistent pixel blocks are as follows: Within the display area of ​​each screen, multiple sampling areas are preset, and a preset range of neighborhoods is selected as candidate image blocks for each pixel; for candidate image blocks within the sampling areas, spatial smoothing conditions are applied for screening: candidate image blocks with a brightness standard deviation of a preset size pixel block below a first preset threshold are selected; for candidate image blocks that pass the spatial smoothing screening, short-term temporal stability conditions are applied for screening: candidate image blocks with an average brightness change of less than a second preset threshold within a consecutive preset number of frames are selected; for candidate image blocks that pass the short-term temporal stability screening, long-term temporal smoothing monotonic conditions are applied for screening: the average brightness time series of the image block in all frames is tracked, and candidate image blocks whose time series shows monotonic variation as a whole and whose cumulative amplitude of reverse fluctuation within the series does not exceed a preset proportion of the cumulative amplitude of positive change are selected; candidate image blocks that pass the above screening are included as spatiotemporally consistent pixel blocks.

[0009] Further, the specific steps for estimating the inverse response function of the camera are as follows: In the front view image sequence, extract the region within the preset pixel width range on both sides of the stitching seam as the zero-radius reference area; calculate the temporal mean of all pixels in the zero-radius reference area in multiple frames to obtain the dark current bias vector of the camera, and calculate the temporal variance of each pixel in multiple frames to obtain the stitching seam leakage noise variance; use the dark current bias vector provided by the zero-radius reference area as the zero-radius hard constraint; use the spatiotemporally consistent pixel block as the data constraint; use the dark area prior constraint constructed based on the stitching seam leakage noise variance; under the joint constraints of the zero-radius hard constraint, data constraint, and dark area prior constraint, obtain the inverse response function of the camera's color channel.

[0010] Further, the specific steps for extracting cross-screen pixel pairs based on local flatness conditions are as follows: Using the horizontal and vertical seam lines as seam center lines; taking a range from a first preset pixel width to a second preset pixel width on both sides of each seam center line as sampling bands belonging to adjacent screens; for each frame of linear radiance image, extracting a pair of cross-screen pixels pixel by pixel along the seam center line direction within the sampling bands of adjacent screens; taking the preset percentile of the standard deviation of the brightness of all sliding windows within a preset neighborhood range along the seam center line direction in the entire frame as an adaptive flatness threshold; based on the adaptive flatness threshold, performing a local flatness test on each pair of cross-screen pixels, and taking the pair of cross-screen pixels that passes the test as a cross-screen pixel pair.

[0011] Furthermore, the specific steps of the local flatness test are as follows: For each pair of cross-screen pixels, calculate the first luminance variance along the center line of the seam within the sampling band on the first side of the seam, the second luminance variance perpendicular to the center line of the seam within the sampling band on the first side of the seam, the third luminance variance along the center line of the seam within the sampling band on the second side of the seam, and the fourth luminance variance perpendicular to the center line of the seam within the sampling band on the second side of the seam; when the first luminance variance, the second luminance variance, the third luminance variance, and the fourth luminance variance are all lower than the adaptive flatness threshold, the pair of cross-screen pixels is determined to have passed the local flatness test.

[0012] Furthermore, the specific steps for estimating the color mapping function from each screen to be calibrated to the reference screen are as follows: Using the value of each pixel in the linear radiance image sequence on the color channel as the channel linear radiance value, and using the color channel as the coordinate axis to construct a color space; using the channel linear radiance value of the screen to be calibrated in the cross-screen pixel pair as coordinates, uniformly divide the color space within the respective value range of the color channels to obtain several sampling nodes; for each sampling node, calculate the weight of each cross-screen pixel pair with the node as the center and the distance within the color space, and introduce a confidence factor based on the frequency of occurrence of the cross-screen pixel pair into the weight; perform weighted regression on the cross-screen pixel pairs in the neighborhood of each sampling node to obtain the initial mapping output value of each sampling node, and apply prior regularization; for sampling nodes with a weighted sample number lower than a preset threshold in the neighborhood, fill by interpolation of neighboring nodes to obtain the color mapping function from each screen to be calibrated to the reference screen.

[0013] Furthermore, the specific steps for generating an independent color correction lookup table for each screen to be calibrated are as follows: Take the color mapping slice of the color mapping function at the center position of the corresponding screen to be calibrated, as the global color correction mapping for each screen to be calibrated; perform uniform sampling within the respective value range of the color channels in the color space to generate several input sampling points; calculate the corresponding correction output value for each input sampling point through the global three-dimensional color correction mapping; and store the correspondence between all input sampling points and correction output values ​​as the color correction lookup table for each screen to be calibrated.

[0014] Furthermore, the specific steps for weighted mixing of the color values ​​of adjacent screens after correction by their respective color correction lookup tables are as follows: Taking the center line of the seam as a reference, extend a preset pixel width to both sides as a fusion band; for each pixel of the adjacent screens covered by the fusion band, correct it by the color correction lookup table of its respective screen to obtain its corrected color value; for each pixel in the fusion band, calculate the mixing weight based on its directed distance to the center line of the seam; and weight and mix the corrected color values ​​according to the mixing weight to obtain the final output color value of the pixel.

[0015] A multi-screen splicing image color consistency correction system includes: a geometric self-calibration unit for acquiring a multi-frame image sequence of the splicing screen and performing geometric normalization on the image sequence to obtain a front view image sequence, wherein the splicing screen includes at least a reference screen and the remaining screens as screens to be calibrated; a response blind calibration unit for extracting the seam area from the front view image sequence as a zero-radius reference, estimating the inverse response function of the camera using the zero-radius reference and detected spatiotemporally consistent pixel blocks, and applying the inverse response function to convert the front view image sequence into a linear radius image sequence; and inter-screen color mapping. The estimation unit defines sampling bands belonging to adjacent screens along the seam on both sides of the linear radiance image sequence. It extracts cross-screen pixel pairs from the sampling bands based on local flatness conditions and estimates the color mapping function from each screen to be calibrated to the reference screen through multi-frame statistical regression. The color calibration fusion unit generates an independent color calibration lookup table for each screen to be calibrated based on the color mapping function and defines a fusion band on both sides of the seam. Within the fusion band, it performs a weighted fusion of the color values ​​of adjacent screens after calibration using their respective color calibration lookup tables to obtain the final output color value.

[0016] The present invention has the following beneficial effects: (1) The multi-screen splicing image color consistency correction method eliminates the interference caused by shooting distortion and device acquisition deviation through geometric normalization and camera inverse response function processing. It also extracts splicing seam lines in the image for clustering processing, calculates the perspective transformation matrix to convert all acquired frames into a front view image, and ensures that each pixel matches the real point position on the screen. At the same time, it uses the splicing screen's built-in light-free splicing seam to calculate the camera's dark current offset value, and combines it with spatiotemporally consistent pixel blocks obtained through three layers of screening. The inverse response function is solved by multiple constraints, converting the original image output by the camera into a radiance image that is proportional to the screen's luminous intensity. This allows the device to complete data acquisition while playing daily videos normally, without interrupting the original business display on the large screen, and also avoids the interference caused by the camera's own light sensitivity deviation on color contrast, thereby improving the correction accuracy.

[0017] (2) The multi-screen splicing image color consistency correction method filters effective cross-screen pixel pairs through layered screening and adaptive flatness verification, builds a color mapping relationship covering the complete color gamut, adapts to the light emission difference generated by long-term use of new and old screens, and delineates a dedicated sampling area far away from the splicing halo. Then, it filters pixel blocks from multiple dimensions, retains only the points with stable scene brightness for calculation, and generates an adaptive judgment standard when processing cross-screen pixels. It counts the brightness fluctuation value of the area from multiple directions. Only pixel pairs with smooth images on all four sides will be retained. Then, it divides nodes with color grids and assigns calculation weights based on the frequency of pixel occurrence. Grids with insufficient samples are filled by interpolation of surrounding nodes. Finally, a complete color mapping is generated, which can accurately restore the light emission difference between new and old screens under different gray levels and different hues.

[0018] (3) The multi-screen splicing image color consistency correction method uses an independent correction lookup table combined with weighted mixing of the splicing seam fusion band to unify the overall visual experience of the multi-screen splicing and completely eliminate the color boundary visible to the naked eye at the splicing seam. Each screen with a different usage time will generate a dedicated correction lookup table. When the screen plays any image, the image pixels will first be converted through the corresponding table to make the output images of all screens aligned with the light emission performance of the reference screen. That is, a fusion area is defined on both sides of the splicing seam, which will not occupy the main display pixels of the screen. Only a small number of pixels close to the splicing seam are mixed. Each pixel in the fusion area will be assigned a corresponding weight according to its distance from the center of the splicing seam. When it is close to the main position of its own screen, its own corrected image occupies a higher proportion. The closer it is to the center line of the splicing seam, the closer the proportions of the corrected images of the two screens are, and the smoother the image transition is completed. This makes the mixing operation based on the color values ​​that have been aligned with the reference screen, without superimposing additional color differences. Under uniform images such as pure white and gradient, there will be no more color blocks at the splicing seam.

[0019] (4) The multi-screen splicing image color consistency correction system automatically completes image acquisition and geometric normalization correction through the geometric self-calibration unit, accurately restores the true imaging field of the screen, the response blind calibration unit completes camera response calibration based on the splicing seam zero-radius brightness reference and spatiotemporally stable pixel blocks, removes the inherent error of the image acquisition link, the inter-screen color mapping estimation unit adaptively selects effective cross-screen pixel samples and regresses to obtain accurate inter-screen mapping relationship, and the color correction fusion unit generates independent screen correction lookup tables in batches and completes smooth mixing output of splicing seam pixels, thereby adapting to the color difference calibration requirements of large screens in online operation, thus ensuring that multi-screen splicing can maintain a stable and uniform high-definition display effect for a long time.

[0020] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0021] Figure 1 This is a flowchart of a multi-screen spliced ​​image color consistency correction method according to the present invention; Figure 2 This is a flowchart illustrating the specific steps involved in estimating the color mapping function from each screen to be corrected to the reference screen in a multi-screen splicing image color consistency correction method of the present invention. Figure 3 This is a block diagram of a multi-screen spliced ​​image color consistency correction system according to the present invention. Detailed Implementation

[0022] Please see Figure 1This invention provides a technical solution: a method for color consistency correction of multi-screen splicing images, comprising the following steps: during normal playback of any dynamic content on the splicing screen, a multi-frame image sequence of the splicing screen is acquired using a camera with locked parameters, and the image sequence is geometrically normalized to obtain a front view image sequence that can be used to accurately extract the seam area, wherein the splicing screen includes at least one reference screen and the remaining screens as screens to be corrected, and there are physical seams between adjacent screens of the splicing screen; The stitching seam region is extracted from the front view image sequence as a zero-radiance reference. Using the zero-radiance reference and spatiotemporally consistent pixel blocks detected in the normally playing content on the screen, the inverse response function of the camera is estimated online. The inverse response function is then applied to convert the front view image sequence into a linear radiance image sequence. In the linear radiance image sequence, sampling bands belonging to adjacent screens are defined on both sides of each physical seam. Cross-screen pixel pairs are extracted from the sampling bands based on local flatness conditions. The extracted cross-screen pixel pairs are then used to estimate the color mapping function from each screen to be calibrated to the reference screen through multi-frame statistical regression. Based on the color mapping function of each screen to be calibrated, an independent color calibration lookup table is generated for each screen. A fusion band with a preset number of pixels is defined on both sides of the physical seam. Within the fusion band, the color values ​​of adjacent screens after calibration by their respective color calibration lookup tables are weighted and mixed to serve as the final output, thereby achieving color consistency calibration of multi-screen spliced ​​images.

[0023] The specific steps for capturing multi-frame image sequences of the splicing screen using a parameter-locked camera are as follows: During the normal playback of any dynamic content on the splicing screen, one screen in the splicing screen is selected as the reference screen, and the remaining screens are used as screens to be calibrated. Use a camera to capture images of the splicing screen. Before capturing the images, lock all camera parameters, including exposure time, analog gain, digital gain, white balance, and focus, to ensure that the imaging parameters of all captured frames are consistent. After all the imaging parameters of the camera are locked, multiple frames of color images from the splicing screen are continuously acquired to form a multi-frame image sequence. After acquisition, the average pixel value of the physical stitching area is automatically detected to vary between frames. If the inter-frame fluctuation of each color channel is lower than the preset threshold, the parameter locking is successful and the acquired multi-frame image sequence can be used for subsequent processing; otherwise, the parameters are locked again and acquisition is resumed.

[0024] In one embodiment, the splicing screen is a display wall composed of four LED displays arranged in a 2x2 layout. Each screen has a physical resolution of 960 x 540 pixels, and the total resolution of the four screens spliced ​​together is 1920 x 1080 pixels. The top left screen is the reference screen (a spare screen of the same model that was newly replaced three months ago), and the other three screens are screens to be calibrated. The camera can be a regular USB color camera that supports the UVC protocol with a resolution of 1920 x 1080 pixels. Before capturing the image, the exposure time is locked to one-sixtieth of a second, the analog gain is locked to 1x, the digital gain is locked to 1x, the white balance is locked to 6500K color temperature, and the focus is locked to the stitching area. During the normal playback of news programs on the splicing screen, 600 frames of color images were continuously acquired at a frame rate of 30 frames per second. After the acquisition was completed, the fluctuation range of the pixel average of each color channel in the physical splicing area was automatically detected within 600 frames. The inter-frame fluctuation of the red channel was 0.3 digital level, the green channel was 0.2 digital level, and the blue channel was 0.4 digital level. None of them exceeded the preset threshold of 0.5 digital level, and the parameter locking was successfully determined.

[0025] Specifically, the steps for geometric normalization of image sequences are as follows: Select one frame from a multi-frame image sequence and detect the straight line formed by the physical seam. Specifically: From the acquired multi-frame image sequence, select one frame with clear imaging and high contrast in the stitching area as the calibration frame. After converting the calibration frame into a grayscale image, use the Canny edge detection operator to extract edge information in the image. The low threshold of the Canny operator is set to 50 and the high threshold is set to 150. Based on the edge image, use the cumulative probability Hough transform to detect straight lines. The distance resolution of the Hough transform is set to 1 pixel, the angle resolution to 1 degree, the accumulator threshold to 80, the minimum line segment length to 100 pixels, and the maximum gap to 20 pixels. Straight lines with a length exceeding one-third of the image width and located in the middle region of the image (at least 20 pixels from the image edge) are selected from the detection results as the straight lines formed by physical stitching. In one implementation, the 300th frame is selected from the 600-frame image sequence as the calibration frame. The content displayed on the screen in this frame is the studio background of a news program, and the contrast between the seam area and the bright area of ​​the screen is relatively high. After converting the frame to a grayscale image, the Canny edge detection operator was used to extract the edges, with the low threshold set to 50 and the high threshold set to 150. Run the cumulative probability Hough transform on the extracted edge map, with the distance resolution set to 1 pixel, the angle resolution set to 1 degree, the accumulator threshold set to 80, the minimum line segment length set to 100 pixels, and the maximum gap set to 20 pixels. Twelve horizontal candidate lines and nine vertical candidate lines were detected. After filtering and removing line segments with a length less than one-third of the image width (i.e., 640 pixels) or located within 20 pixels of the image edge, eight horizontal lines and six vertical lines were retained as the lines formed by the physical stitching. The detected straight lines are divided into horizontal and vertical groups according to their angles. Hierarchical clustering is then performed on the lines in each group based on their intercepts. The mean line of each cluster is taken to obtain the horizontal and vertical seam lines, as follows: The detected straight lines are grouped according to the angle between them and the horizontal direction of the image. The angles within the range of 0 degrees ± 5 degrees are grouped into the horizontal group, and the angles within the range of 90 degrees ± 5 degrees are grouped into the vertical group. All lines within a horizontal group are hierarchically clustered from bottom to top based on their intercept values ​​with the bottom edge of the image. The clustering distance threshold is set to 15 pixels. The slope and intercept of all lines within each cluster are arithmetically averaged to obtain the mean line corresponding to that cluster. The two mean lines generated by the horizontal group clustering are used as two horizontal seam lines, and the two mean lines generated by the vertical group clustering are used as two vertical seam lines. In one embodiment, the angle range of the eight horizontal lines is between -2 degrees and 3 degrees, and the angle range of the six vertical lines is between 88 degrees and 93 degrees. The horizontal groups of straight lines are hierarchically clustered according to their intercept values ​​with the bottom edge of the image, using Euclidean distance metric and Wardlinkage criterion, with a distance threshold of 15 pixels. The clustering results in two clusters. The first cluster contains four straight lines with a mean intercept of 537 pixels, and the second cluster contains four straight lines with a mean intercept of 543 pixels. These two clusters serve as the two horizontal seam lines, namely the upper and lower edges of the horizontal seam. Similarly, the vertical seam is clustered to obtain two clusters with mean intercepts of 957 pixels and 963 pixels, which serve as the two vertical seam lines, namely the left and right edges of the vertical seam. The final horizontal seam lines are y=537 and y=543, and the vertical seam lines are x=957 and x=963. The average intercept of the upper and lower edge lines (537+543) / 2=540 is taken as the position of the horizontal seam center line, and the average intercept of the left and right edge lines (957+963) / 2=960 is taken as the position of the vertical seam center line. The corner points of each screen are determined by the intersection of the horizontal and vertical seam lines. The four outermost corner points are taken as the source vertices for perspective transformation, specifically as follows: Calculate the coordinates of all four intersection points of the two horizontal seam lines and the two vertical seam lines. The four intersection points correspond to the common corner points of each screen. Take the four corner points located on the outermost side of the image as source vertices. Store the four source vertices in the order of top left, top right, bottom right, and bottom left. In one implementation, two horizontal seam lines y=537 and y=543 intersect with two vertical seam lines x=957 and x=963, resulting in four intersection points: (957, 537), (963, 537), (957, 543), and (963, 543). These four intersection points correspond to the common corner points of the four screens of the 2x2 splicing wall. The top left corner (0, 0) of the image is taken as the top left source vertex, the top right corner (1920, 0) as the top right source vertex, the bottom right corner (1920, 1080) as the bottom right source vertex, and the bottom left corner (0, 1080) as the bottom left source vertex, and stored in order. Using the pixel dimensions of the rectangle enclosed by the corner points of the reference screen in the source image as a reference, and maintaining a consistent pixel sampling density, the width and height of the target rectangle are set. The four vertices of the target rectangle are then used as the target vertices for perspective transformation. Specifically: The reference screen is the most recently replaced or selected screen in the splicing screen as the color reference standard. The pixel width and pixel height occupied by the reference screen in the source image are calculated using the pixel coordinates of the upper left and lower right corners of the reference screen in the source image. The pixel width is equal to the x-coordinate of the bottom right corner minus the x-coordinate of the top left corner, and the pixel height is equal to the y-coordinate of the bottom right corner minus the y-coordinate of the top left corner. The pixel width and pixel height are directly used as the width and height of the target rectangle. The coordinates of the four vertices of the target rectangle are set to (0, 0), (W, 0), (W, H), and (0, H), respectively, where W is the width of the target rectangle and H is the height of the target rectangle, so as to ensure that the pixel sampling density of the reference screen area after perspective transformation is consistent with that of the source image. In one implementation, the reference screen is the upper left corner screen, with its upper left corner point being (0, 0) and its lower right corner point being the seam intersection point P1 (957, 537). Therefore, the baseline screen pixel width W = 957 - 0 = 957 pixels, the pixel height H = 537 - 0 = 537 pixels, the target rectangle width is set to 957 pixels, the height is set to 537 pixels, and the target vertices are set to (0, 0), (957, 0), (0, 537), (957, 537). The homography matrix from the source vertex to the target vertex is solved using a direct linear transformation. A perspective transformation is then performed on each frame of the image sequence to obtain a frontal view image sequence, specifically as follows: Substituting the coordinates of the four pairs of source vertices and the target vertex into the direct linear transformation equations, the eight degrees of freedom parameters of the 3x3 homography matrix are solved by singular value decomposition. After obtaining the homography matrix, perspective transformation is performed on each of the acquired frames. During the transformation, bilinear interpolation is used to calculate the pixel values ​​at the sub-pixel positions. After all frames have been transformed, a front view image sequence that can be used to accurately extract the seam area is obtained.

[0026] In one implementation, four pairs of corresponding points (0,0), (0,0), (1920,0), (957,0), (1920,1080), (957,537), (0,1080), (0,537) are substituted into the system of direct linear transformation equations, and the homography matrix is ​​solved by SVD decomposition. Perspective transformation was performed on all 600 frames of images by applying the homography matrix frame by frame, and the output pixel values ​​were calculated using bilinear interpolation. The resolution of the transformed front view image sequence was 957 x 537 pixels. The scaling ratio in the x-direction is approximately 0.498, and the scaling ratio in the y-direction is approximately 0.497. The original horizontal seam upper edge line y=537 is mapped to approximately 267 pixels (rounded up), the lower edge line y=543 is mapped to approximately 270 pixels (rounded up), and the horizontal seam center line is approximately 268 pixels (rounded up). The original left edge line of the vertical seam (x=957) is approximately 477 pixels (rounded down) after mapping, the right edge line (x=963) is approximately 480 pixels (rounded down) after mapping, and the center line of the vertical seam is approximately 478 pixels (rounded down). Therefore, the horizontal seam center line is located in row 268, the vertical seam center line is located in column 478, the top edge of the horizontal seam is located in row 267, and the bottom edge is located in row 270; the left edge of the vertical seam is located in column 477, and the right edge is located in column 480, with a seam width of approximately 3 to 4 pixels.

[0027] In the front view image sequence, the physical seams are presented as horizontal and vertical straight lines, and the correspondence between the coordinates of each pixel and the physical position on the screen is equidistant throughout the entire frame, so that the cross-screen pixel pairs extracted along both sides of the seam in subsequent steps correspond precisely in space.

[0028] Specifically, the steps for detecting spatiotemporally consistent pixel blocks are as follows: Within the display area of ​​each screen, avoiding the location of physical seams, multiple sampling areas are preset, and a preset range of neighborhoods is selected as candidate image blocks for each pixel. Specifically: Within each screen, in the display area far from the physical seam and the physical edge of the screen, five rectangular sampling areas of 50 by 50 pixels are uniformly selected. The five sampling areas are located at the center of the screen and the four corner areas, respectively. The edge of each sampling area is at least 20 pixels away from the center line of the seam. Within each sampling area, the sampling area slides pixel by pixel with a step size of 1 pixel. A 3 by 3 pixel neighborhood centered on the current pixel is taken as a candidate image block. In one embodiment, the display area of ​​the reference screen (upper left screen) is row 0 to 268 and column 0 to 478 in the front view image sequence, with its center position at row 134 and column 239. The central sampling area is centered at (134, 239), with a size of 50 by 50 pixels, the upper left corner coordinates are (109, 214), and the lower right corner coordinates are (158, 263). The four corner sampling areas are located 30 pixels from each of the four corners of the screen. Each sampling area avoids the center line of the stitching seam by at least 20 pixels, which meets the requirements. Each sampling area extracts (50-2)×(50-2)=2304 candidate image blocks. The total number of candidate image blocks extracted from the twenty sampling areas of the four screens is approximately 46080. For candidate image blocks within the sampling area, a spatial smoothing condition is applied for filtering: candidate image blocks with a brightness standard deviation lower than a first preset threshold are selected, specifically as follows: For each 3x3 candidate image block, the standard deviation of the brightness values ​​of the 9 pixels within the block is calculated. The brightness value is the weighted average of the red, green, and blue channels, with weighting coefficients of 0.299, 0.587, and 0.114, respectively. The first preset threshold is set to 2.0. When the standard deviation of the brightness within a candidate image block is lower than 2.0, it indicates that the brightness change within the image block is gradual and the probability of it belonging to a flat region is high. It is then retained for the next round of screening. In one implementation, taking the 100th candidate image block in the central sampling area of ​​the reference screen as an example, its 9 pixel brightness values ​​are [128, 130, 129, 127, 128, 130, 131, 129, 128], with a standard deviation of 1.25, which is less than 2.0. Through screening, 1856 out of the 2304 candidate image blocks in this sampling area pass the spatial smoothing screening. For candidate image blocks selected through spatial smoothing, a short-term temporal stability condition is applied: candidate image blocks whose average brightness change is lower than a second preset threshold within a consecutive preset number of frames are selected, specifically as follows: For candidate image blocks that pass the spatial smoothing screening, the average brightness value of the block at the same spatial location in a continuous 3-frame orthographic image sequence is tracked in a sliding window manner, with a sliding step of 1 frame, covering all acquisition frames. The current window starts with three consecutive frames and then takes the next two frames. The average brightness change of the candidate image patch is calculated in these three frames. The average brightness change is defined as the difference between the maximum and minimum average brightness in the three frames. The second preset threshold is set to 1.5. When the change is less than 1.5, it indicates that the scene radiance corresponding to the candidate image block remains constant in a short period of time, and it is retained to enter the next round of screening. In one implementation, the average brightness of a candidate image block in frames 100, 101, and 102 is 126.3, 127.1, and 126.8, respectively, with a maximum difference of 0.8, which is less than 1.5. Through screening, 623 out of 1856 candidate image blocks in the central sampling area of ​​the reference screen pass the short-time temporal domain stability screening. For candidate image patches selected through short-time temporal stability screening, a long-time temporal smoothing monotonic condition screening is applied: the average brightness time series of the image patch in all frames is tracked, and candidate image patches whose time series shows monotonic variation as a whole and whose cumulative amplitude of reverse fluctuations within the series does not exceed a preset proportion of the cumulative amplitude of positive changes are selected, specifically as follows: Extract the average brightness time series of all acquired frames, calculate the absolute value of the difference between the average brightness of the first frame and the average brightness of the last frame of the time series, and require that the absolute value be greater than the preset monotonic threshold of 20. Calculate the sum of the absolute values ​​of the second difference of the time series, and require that the value be less than the preset smoothing threshold of 300. Statistically calculate the cumulative amplitude of the reverse fluctuation in the time series, and require that the cumulative amplitude of the reverse fluctuation do not exceed 10% of the cumulative amplitude of the positive change. When the time series meets all three conditions mentioned above, it indicates that the scene radiance of the candidate image patch changes in a monotonically smooth manner throughout the entire acquisition period, thus improving its reliability as a constant radiance reference. The preset smoothing threshold of 300 corresponds to an average second-order difference of no more than 0.5 brightness levels per frame in a 600-frame sequence. This value has been experimentally determined to be the optimal value in terms of balancing noise robustness and screening rate. In one implementation, the brightness of the first frame of a candidate image block time series is 45.2, and the brightness of the last frame is 152.7, with a difference of 107.5, which is greater than the monotonic threshold of 20. The sum of the absolute values ​​of the second-order differences of the sequence is 187, which is less than the smoothing threshold of 300; the cumulative upward amplitude in the sequence is 128.8, the cumulative downward amplitude is 8.2, and the reverse fluctuation accounts for 6.4%, which is less than 10%, so it passes the screening. Of the 623 candidate image blocks in the central sampling area of ​​the baseline screen, 47 passed the long-term screening, and a total of 312 spatiotemporally consistent pixel blocks were collected from the twenty sampling areas of the four screens. Candidate image blocks that pass the above screening are included as spatiotemporally consistent pixel blocks, specifically as follows: The center pixels of candidate image blocks that are sequentially filtered by spatial smoothness conditions, short-term temporal stability conditions, and long-term temporal smoothness monotonicity conditions are included as spatiotemporally consistent pixel blocks, and each included spatiotemporally consistent pixel block is associated with its pixel values ​​in each frame.

[0029] The specific steps for estimating the inverse response function of the camera are as follows: In the front view image sequence, the area within a preset pixel width on both sides of the physical seam center line is extracted as the zero-radius reference area, specifically as follows: The zero-radiance reference area originates from the physical seams of the splicing screen itself, which do not emit light. It is part of these physical seams. Based on the determined positions of the horizontal and vertical seam center lines, one pixel width is taken on each side of each seam center line, resulting in a total width of 3 pixels for the zero-radiance reference area. This area is entirely located inside the physical structure of the seams and does not contain any part of the screen's light-emitting pixels. In one implementation, based on the determined horizontal seam center line y=268 and vertical seam center line x=478, 1 pixel is taken above and below the horizontal seam center line, i.e., the y range is [267, 269], and 1 pixel is taken to the left and right of the vertical seam center line, i.e., the x range is [477, 479]. The horizontal seam reference area has approximately 957 x 3 = 2871 pixel positions, and the vertical seam reference area has approximately 537 x 3 = 1611 pixel positions. The two areas overlap by approximately 3 x 3 = 9 pixels at their center intersection. Therefore, the total number of pixel positions in the zero-radiance reference area is approximately 2871 + 1611 - 9 = 4473. The temporal mean of all pixels within the zero-radiance reference area is calculated across multiple frames to obtain the camera's dark current bias vector. The temporal variance of each pixel within this area is then calculated across multiple frames to obtain the stitching seam leakage noise variance, specifically: For each pixel location within the zero-radiance reference area, the sequence of red, green, and blue pixel values ​​in all acquisition frames is collected. The arithmetic mean of each color channel in the time dimension is calculated. The arithmetic mean of the three color channels constitutes the dark current bias vector, which reflects the background output level of the camera when there is no light input. Simultaneously, the variance of each pixel position in the time dimension is calculated, and the truncated average of the variances of all pixels in the zero-radiance reference area is taken as the seam leakage noise variance. The truncated average is the mean after removing the highest 5% and the lowest 5%. The seam leakage noise variance is used to characterize the degree of weak stray light fluctuation in the seam area. In one implementation, for approximately 4473 pixel locations within the zero-radiance reference area, the mean values ​​of R, G, and B for 600 frames are calculated respectively. The mean value of the R channel for all pixels is 12.3, the mean value of the G channel is 10.1, and the mean value of the B channel is 14.5, resulting in a dark current bias vector of (12, 10, 14) (rounded down). After calculating the temporal variance of each pixel location, the truncated average is taken to obtain the seam leakage noise variance: R channel 3.2, G channel 2.8, B channel 3.5. The dark current bias vector provided by the zero-radiance reference region is used as the zero-radiance hard constraint, specifically as follows: The pixel value corresponding to the dark current bias vector is used as the zero-point input of the inverse response function, forcing the output radiance value of the inverse response function at that pixel value to be zero. In one implementation, the dark current bias vector (12, 10, 14) is used as the zero constraint of the inverse response function, that is, the inverse response function g is forced to satisfy g_R(12)=0, g_G(10)=0, g_B(14)=0; Using spatiotemporally consistent pixel blocks to provide multiple sets of pixel value samples under the same scene radiance as data constraints, specifically: Each spatiotemporally consistent pixel block is considered to correspond to the same constant scene radiance during all acquisition frames. The pixel block presents different pixel values ​​in different frames due to camera noise and small fluctuations. These pixel values ​​constitute multiple samplings of the camera response function under the constant radiance. All spatiotemporally consistent pixel blocks together provide a large amount of data constraints covering multiple radiance levels from dark to bright. The unknown logarithmic radiance corresponding to each pixel block is used as a variable to be optimized in the solution. In one implementation, there are 312 spatiotemporally consistent pixel blocks, each associated with 600 frames of pixel values, providing a total of 312 × 600 = 187,200 sets of data constraints. Taking the 15th pixel block as an example, its average R channel pixel value over 600 frames is 87.3, with a fluctuation range of 82-93. It is considered as multiple samples under the same radiance, and its unknown logarithmic radiance is used as a variable to be optimized. By constructing a priori constraints for dark areas based on the variance of seam leakage noise, the inverse response function remains approximately linear within a preset dark area interval near the corresponding pixel value in the zero-radius reference region, thus preventing overfitting and bending in low-radius regions. Specifically: The ten pixel value levels near the corresponding pixel value of the zero-radiance reference area are defined as the preset dark area interval. Within this interval, a dark area prior constraint term is constructed, requiring that the slope of the inverse response function at each pixel value within this interval be consistent with the dark area linear slope obtained by converting the variance of the stitching leakage noise. The linear slope of the dark area is determined by the ratio of the standard deviation of the seam leakage noise to the equivalent radiance fluctuation caused by the seam leakage noise. This constraint term is added to the total energy function in the form of a penalty term to prevent unreasonable bending or oscillation due to noise interference in areas with extremely low pixel values ​​and sparse data constraints in the dark area, thus ensuring the physical rationality of the dark area mapping.

[0030] In one implementation, taking the R channel as an example, the dark current bias is 12, the preset dark area range is 12 to 22 (10 pixel value level), the standard deviation of the seam light leakage noise is about 1.79 pixel values, and the resulting equivalent radiance fluctuation is about 1.79 divided by 255, which is approximately equal to 0.007. Therefore, the linear slope of the dark area is approximately 0.007 divided by 1.79, which is about 0.0039. The constraint requires that the sum of the squares of the differences between the slope of the inverse response function at adjacent pixel values ​​in the interval of 12 to 22 and 0.0039 be as small as possible. A weight of 0.5 is added to the total energy function. Under the combined constraints of zero-radiance hard constraints, data constraints, and dark area prior constraints, three inverse response functions for the red, green, and blue color channels of the camera are obtained, specifically as follows: The total energy function is composed of a weighted combination of three constraints. The first term is the data fidelity term, which means: for each spatiotemporally consistent pixel block, the sum of the squares of the difference between the radiance value after mapping the pixel value in each frame through the inverse response function and the estimated constant radiance value of the pixel block. The second term is a smoothing constraint, which means that the value of the inverse response function should change smoothly at adjacent pixel values, and the sum of the squares of the differences between adjacent values ​​should be weighted by a coefficient of 0.1. The third term is the prior constraint term for dark areas, which means that in the dark area interval near the corresponding pixel value of the zero radiance reference area, the slope of the inverse response function at adjacent pixel values ​​should be consistent with the dark linear slope obtained by converting the variance of the seam leakage noise. The sum of the squares of the difference between the actual slope and the reference slope is taken as the weighting coefficient of 0.5. The solution uses an alternating iterative approach: First, the inverse response function is fixed. At this time, the optimal estimate of the constant radiance of each spatiotemporally consistent pixel block is the arithmetic mean of its mapped radiance values ​​in each frame, which is directly calculated. Then, each constant radiance estimate is fixed, and the value of the inverse response function at each pixel value is updated. At this time, the total energy function degenerates into a quadratic optimization problem with respect to the value of the inverse response function. The updated inverse response function is obtained by solving the linear equation system, and a boundary constraint with zeros is applied. The above alternating process is repeated until convergence or the preset maximum number of iterations of 100 is reached. This solution process is performed independently for the red, green and blue color channels to obtain three inverse response functions.

[0031] It should be noted that the inverse response function obtained from the above solution process represents the relative radiance of the scene, and its absolute scale is not fixed. In order to unify the radiance magnitude of different color channels in subsequent calculations, the inverse response function of each channel is post-processed and scaled after the solution is completed, so that the mapping value of each channel at the pixel value of 255 is equal and is 1. This post-processing scaling only unifies the dimensions of each channel and does not affect the estimation accuracy of the subsequent color mapping function, because the color mapping function learns the relative mapping relationship of color values ​​across the screen, rather than the absolute radiance value.

[0032] Specifically, the steps for extracting cross-screen pixel pairs based on local flatness conditions are as follows: The horizontal and vertical seam lines are used as the center lines of the seam, specifically as follows: The equations of the horizontal and vertical seam lines in the front view image sequence are used as the position reference for the seam center line, which is the geometric center line of the non-illuminated black area of ​​the physical seam, and its position remains constant throughout the entire front view image sequence.

[0033] In one implementation, the determined horizontal seam centerline y=268 and vertical seam centerline x=478 are used as spatial references for the seam centerlines, and these positions remain fixed throughout the entire front view image sequence. On both sides of the center line of each seam, a range from a first preset pixel width to a second preset pixel width from the center line of the seam is selected as the sampling band belonging to the adjacent screen. The minimum distance of the sampling band from the center line of the seam is greater than the width of the halo scattering area of ​​the seam. Specifically: The boundary of the sampling band on the side closest to the center line of the seam is 4 pixels away from the center line of the seam, and the boundary of the sampling band on the side furthest from the center line of the seam is 9 pixels away from the center line of the seam, with each sampling band being 5 pixels wide. A safety distance of 4 pixels is maintained between the sampling band and the center line of the splicing seam. This distance ensures that the sampling band avoids the halo mixing area caused by scattering and reflection on both sides of the splicing seam, and the collected pixel values ​​purely reflect the light emission characteristics of their respective screens. In one implementation, taking the vertical seam center line x=478 as an example, the range of the left sampling band (belonging to the reference screen) x is [478-9, 478-4], i.e. [469, 474], with a total width of 5 pixels; The right sampling band (belonging to the right side of the screen to be calibrated) has an x ​​range of [478+4, 478+9], which is [482, 487], with a total width of 5 pixels. The horizontal seam is similar. The upper sampling band has a y range of [268-9, 268-4], which is [259, 264], and the lower sampling band has a y range of [268+4, 268+9], which is [272, 277]. For each frame of linear radiance image, within the sampling band of each adjacent screen, a pair of cross-screen pixels is extracted pixel by pixel along the center line of the seam. The cross-screen pixel is the median of all pixels within the width of its respective sampling band. Specifically: For each frame in the linear radiance image sequence, the image is processed position by position with a step size of 1 pixel along the center line of the stitching seam. At each position, the median of each of the 5 pixels in the sampling band in the red, green and blue channels is taken to form two three-dimensional linear radiance vectors of a pair of cross-screen pixels. The use of the median effectively suppresses the interference of individual abnormal pixels in the sampling band. In one implementation, taking the 300th frame with a vertical seam x=478 as an example, the position is extracted pixel by pixel from 20 to 248 along the y direction. This range is centered on the intersection of the vertical seam center line and the horizontal seam center line, excluding the cross intersection area of ​​about 20 pixels above and below. The intersection area involves pixels from four screens, where horizontal and vertical seams meet. The physical structure and scattering halo are more complex than in areas with a single seam direction. Excluding it ensures that cross-screen pixels within the sampling band come purely from the two opposite screens of a single seam, avoiding multi-screen light mixing interference in the intersection area. At y=100, the R channel values ​​of the 5 pixels in the left sampling band are [0.452, 0.448, 0.455, 0.450, 0.453], with a median of 0.452; the G channel median is 0.387; and the B channel median is 0.298. The R channel median of the 5 pixels in the right sampling band is 0.468, the G channel median is 0.382, and the B channel median is 0.305, forming a pair of cross-screen pixels ((0.452, 0.387, 0.298), (0.468, 0.382, 0.305)). The preset percentile of the standard deviation of the brightness of all sliding windows within a preset neighborhood along the center line of the stitching seam in the entire frame is taken as the adaptive flatness threshold, specifically: Within the entire frame, a sliding window of 7 pixels in length along the center line of the stitching is used to traverse all positions of the entire frame with a step size of 1 pixel, and the standard deviation of the pixel brightness in each window is calculated. Sort all sliding windows by standard deviation from smallest to largest, and take the 80th percentile as the adaptive flatness threshold. This threshold is automatically adjusted according to the content of each frame of the image. The threshold is lower when the image is generally flat and higher when the image has rich texture, so as to ensure that the filtering conditions match the image content. In one implementation, a 7-pixel sliding window is taken across the entire frame to calculate the standard deviation of brightness and sort them. The 80th percentile of a certain frame is 3.8, and this value is used as the adaptive flatness threshold for the current frame. Based on the adaptive flatness threshold, a local flatness test is performed on each pair of cross-screen pixels, and the pair of cross-screen pixels that passes the test is taken as the cross-screen pixel pair.

[0034] The specific steps for local flatness inspection are as follows: For each pair of cross-screen pixels, calculate the first luminance variance along the center line of the seam within the sampling band on the first side of the seam, the second luminance variance perpendicular to the center line of the seam within the sampling band on the first side of the seam, the third luminance variance along the center line of the seam within the sampling band on the second side of the seam, and the fourth luminance variance perpendicular to the center line of the seam within the sampling band on the second side of the seam. Specifically: Centered on the current cross-screen pixel position, take 7 adjacent pixels along the center line of the first side of the seam to form the first neighborhood, and calculate the variance of the brightness of these 7 pixels to obtain the first brightness variance. Five pixels are taken from the sampling band on the first side of the seam along the direction perpendicular to the center line of the seam, i.e., the width of the sampling band, to form the second neighborhood. The variance of the brightness of the five pixels is calculated to obtain the second brightness variance. The third luminance variance of the 7-pixel neighborhood along the center line of the seam and the fourth luminance variance of the 5-pixel neighborhood perpendicular to the center line of the seam are calculated in the same way within the sampling band on the second side of the seam. The four variances reflect the degree of texture change of the two screens in different directions within their respective sampling areas. In one implementation, taking the cross-screen pixel pair at y=100 as an example, the brightness values ​​of 7 pixels (y=97 to 103) along the seam direction (y direction) within the left sampling band are taken, with a mean brightness of 138.5 and a variance of 2.1. The variance of brightness for 5 pixels in the vertical seam direction (x direction 469 to 474) within the left sampling band is 1.8; the variance for 7 pixels along the seam direction within the right sampling band is 2.4; and the variance for 5 pixels in the vertical seam direction is 2.0. When the first luminance variance, the second luminance variance, the third luminance variance, and the fourth luminance variance are all lower than the adaptive flatness threshold, the pair of cross-screen pixels is determined to have passed the local flatness test, specifically as follows: The four calculated brightness variances are compared with the adaptive flatness threshold. If all four variances are lower than the threshold, it indicates that there are no obvious brightness fluctuations in the sampling areas on both sides of the seam along the seam direction and perpendicular to the seam direction, that is, both sides are flat or gradually changing areas. At this point, the color difference between the cross-screen pixels can be attributed to the difference in the color transfer function of the two screens themselves, rather than the natural texture change of the video content at the seam. Therefore, it is marked as a valid cross-screen pixel pair that passes the local flatness test. In one implementation, the current frame adaptive flatness threshold is 3.8, and the four variances 2.1, 1.8, 2.4, and 2.0 are all less than 3.8. Therefore, the cross-screen pixel pair is determined to pass the test and is regarded as a cross-screen pixel pair.

[0035] Specifically, such as Figure 2 As shown, the specific steps for estimating the color mapping function from each screen to be calibrated to the reference screen are as follows: The value of each pixel in the linear radiance image sequence on the color channel is used as the channel linear radiance value, and the color channel is used as the coordinate axis to construct the color space, which is as follows: The red channel linear radiance value, green channel linear radiance value, and blue channel linear radiance value of each pixel in the linear radiance image sequence are used as the three coordinate components of the pixel in the color space. The three coordinate axes of the color space correspond to the red channel, green channel, and blue channel, respectively. The value range of each coordinate axis is from zero to the maximum linear radiance value of each channel. Each point in the color space represents a color. In one implementation, the R, G, and B values ​​(range 0 to 1) of each pixel after mapping by the inverse response function are taken as three-dimensional coordinates. For example, the linear radiance value of a pixel is (0.452, 0.387, 0.298), which corresponds to a point in the color space. Using the linear radiance value of the channel of the screen to be calibrated as the coordinate, the color space is uniformly divided within the respective value range of each color channel to obtain several sampling nodes, specifically: The three-dimensional linear radiance values ​​belonging to the screen to be calibrated in the cross-screen pixel pair dataset are used as input sample coordinates in the color space. Uniform sampling is performed in the red, green and blue color channels between 0 and 1. Each channel is equally divided into 32 intervals, and each sampling node has a definite coordinate position in the color space. For each sampling node, the weight of each cross-screen pixel pair is calculated with that node as the center and the distance within the color space. A confidence factor based on the frequency of occurrence of the cross-screen pixel pair is introduced into the weight. This confidence factor is obtained by mapping the frequency to the interval of 0 to 1 through a preset mapping function, specifically as follows: For each sampling node, calculate the Euclidean distance in color space from the color value of the screen to be corrected to the sampling node for each cross-screen pixel pair. Use the cube kernel function value with this distance as the independent variable as the basic weight. The cube kernel function is in the form that when the distance is less than the bandwidth, the weight is equal to 1 minus the cube of the distance divided by the cube of the bandwidth, otherwise the weight is zero. The bandwidth is 0.05, and the value range of the basic weight is from 0 to 1. At the same time, the frequency of the cross-screen pixel pair appearing within a preset neighborhood radius centered on its color coordinates among all cross-screen pixel pairs is counted. The frequency is then mapped to a confidence coefficient between 0 and 1 using a preset mapping function. The mapping function is in the form of frequency divided by frequency plus a parameter (e.g., 10). The confidence coefficient is used to adjust the base weights, resulting in a comprehensive weight equal to the base weight multiplied by the confidence coefficient. The comprehensive weight reflects the dual contribution of color similarity and sample frequency: the higher the frequency of occurrence, the closer the confidence coefficient is to 1, and the base weight is fully preserved; when the frequency of occurrence is extremely low, the confidence coefficient is close to 0, and the influence of the sample is effectively suppressed. In the subsequent weighted regression, the comprehensive weights of all samples participating in the regression will be further normalized so that the total weight sums to 1.

[0036] In one implementation, taking the sampling node (0.5, 0.3, 0.2), bandwidth 0.05, and frequency mapping parameter k=10 as an example, the color value of a cross-screen pixel in the dataset to be calibrated is (0.51, 0.29, 0.21), and its Euclidean distance to the node is √((0.01)²+(0.01)²+(0.01)²)=0.0173; The basic weight of the three-cubic-kernel is (1-(0.0173 / 0.05)³)³=0.782. This pixel pair appears 35 times in the color space neighborhood, and the confidence coefficient is approximately 0.778. Therefore, the comprehensive weight is approximately 0.782×0.778≈0.608. In the subsequent weighted median regression, the comprehensive weight of this sample will be normalized together with the weights of other neighborhood samples to make the total weight sum to 1. Weighted regression is performed on cross-screen pixel pairs within the neighborhood of each sampling node to obtain the initial mapping output value of each sampling node. Physical prior regularization is then applied to constrain the color mapping function to tend towards diagonal dominance in the local Jacobian matrix. Specifically: For each sampling node, collect all cross-screen pixel pairs in the neighborhood with a comprehensive weight greater than zero, normalize the comprehensive weight of all samples in the neighborhood so that the sum is 1, take the channel linear radiance value of the reference screen in these cross-screen pixel pairs as the target output, and use weighted median regression to calculate the initial mapping output value of the sampling node. The physical prior regularization is performed as a post-processing step after regression. The specific operation is as follows: weighted median regression is performed on all sampling nodes to obtain the initial mapping output value of each node. For each sampling node, the Jacobian matrix at that node is calculated by difference using its mapping relationship with neighboring nodes. This matrix describes how a small change in the input color causes a change in the output color. Calculate the energy percentage of the off-diagonal elements in the Jacobian matrix, which is the ratio of the sum of squares of the off-diagonal elements to the sum of squares of all elements. When this ratio exceeds 15%, it indicates that the crosstalk component between channels in the mapping relationship is too large, which is inconsistent with the physical law of independent decay of the three primary colors of LED. At this point, the output value of the current node is shrunk and corrected: the corrected output value is equal to the initial output value shifted towards the pure diagonal mapping output value, and the shift is controlled by the shrinkage coefficient 0.2. That is, the corrected value is equal to 80% of the initial value plus 20% of the pure diagonal mapping output value, where the pure diagonal mapping output value is the output value that retains only the linear mapping relationship determined by the diagonal elements of the Jacobian matrix. The corrected output value replaces the original initial mapping output value and participates in subsequent interpolation filling.

[0037] In one implementation, for node (0.5, 0.3, 0.2), there are 183 cross-screen pixel pairs in the neighborhood. The sum of the weights of each sample after normalization is 1. Weighted median regression is performed using the normalized weights to obtain the initial mapping output value (0.522, 0.285, 0.193). Calculate the Jacobian matrix of this node. The energy percentage of the off-diagonal elements is 0.08, which is less than 0.15, so no shrinkage is needed. The off-diagonal percentage of the other node is 0.22, so its output value is corrected to 80% of the initial value plus 20% of the output value of the pure diagonal mapping. For sampling nodes with a weighted sample count below a preset threshold in their neighborhood, interpolation with neighboring nodes is used to fill the gap, obtaining a color mapping function from the screen to be calibrated to the reference screen that covers the entire color gamut. Specifically: Traverse all sampling nodes and count the number of weighted samples in the neighborhood of each sampling node. The preset threshold can be set to 10. For sampling nodes with a weighted sample count of less than 10; Trilinear interpolation is used to fill the color space using the estimated output values ​​of the neighboring nodes. The trilinear interpolation calculates the output value of the color space by weighting the distance between the node and the eight nearest completed nodes in the color space. After all nodes are filled, a color mapping function that covers the entire color gamut is formed in the color space. This color mapping function describes the mapping relationship between any input color value of the screen to be calibrated and the color value that the color should appear on the reference screen (i.e., the forward mapping from the screen to be calibrated to the reference screen).

[0038] In one implementation, 35,937 nodes are traversed, and 1,247 nodes have fewer than 10 weighted samples in their neighborhood. Taking node (0.1, 0.05, 0.95) as an example, the mapping output value of the node is calculated by trilinear interpolation of the eight estimated nodes around it. The final color mapping function covers the full color gamut range of R, G, and B channels from 0 to 1.

[0039] The specific steps for generating an independent color correction lookup table for each screen to be calibrated are as follows: The color mapping slice at the center of the corresponding screen to be calibrated is taken as the global color calibration map for each screen to be calibrated. Specifically: The color mapping function obtained by blind estimation describes the forward mapping relationship from the color of the screen to be calibrated to the color of the reference screen. When generating the color calibration lookup table, based on this forward mapping, a query and interpolation mechanism is constructed from the target color of the reference screen to the output color of the screen to be calibrated, in order to generate the inverse mapping. Specifically: For each target color in the standard color gamut, search for sample points that have been covered by the forward mapping in its color space neighborhood. Estimate the output value required for the screen to be calibrated to match the target color by performing weighted interpolation on the input values ​​of the screen to be calibrated corresponding to these sample points. For the many-to-one situation that may exist in the forward mapping (i.e., multiple different colors of the screen to be calibrated are attenuated and mapped to the same color of the reference screen), the interpolation process selects the sample point closest to the target color for estimation; for the dark or highlight areas in the target color gamut that are not covered by the forward mapping, they are filled in by neighborhood extrapolation. This yields a global correction mapping from the target color of the reference screen to the output color of the screen to be corrected. When the color mapping function is a global three-dimensional mapping and the color characteristics within the screen are uniform, there is no need to extract spatial slices; the above query and interpolation mechanism can be directly applied to the full mapping. In one implementation, taking the screen to be calibrated (the lower right corner of the screen) as an example, the color mapping function obtained by blind estimation is a forward three-dimensional color mapping from the screen to be calibrated to the reference screen, which includes the mapping relationship of multiple nodes. Based on this forward mapping, a query and interpolation mechanism is constructed to generate the inverse mapping, as follows: A reverse lookup structure is established from the reference screen color to the screen color to be calibrated. To do this, all 35,937 sampling nodes of the forward mapping are traversed. The color value of each node after forward mapping, which is located in the reference screen color gamut, is used as the index key, and its original input color value, which is located in the color gamut of the screen to be calibrated, is used as the storage value, forming a sparse three-dimensional mapping relationship set. In the standard target color gamut, the red, green and blue channels are uniformly sampled at 33 levels, generating a total of 35,937 target colors on regular grids. For each target color, the 8 index keys with the closest Euclidean distance are searched in the sparse mapping relationship set, and the storage values ​​corresponding to these index keys are obtained. Using the inverse distance from the target color to each index key as the weight, the eight stored values ​​are weighted and averaged to obtain the inverse mapping output value corresponding to the target color; the correspondence between all target colors and the inverse mapping output value constitutes the inverse mapping, which is used as the global color correction mapping for the screen to be corrected; Uniform sampling is performed within the respective value range of each color channel in the color space to generate several input sampling points, specifically as follows: In the color space of the standard target color gamut, the red, green and blue color channels are sampled uniformly at 33 levels in the range of 0 to 1, with a sampling interval of 1 / 32. A total of 33 x 33 x 33, or 35937 three-dimensional input sampling points are generated for the three channels. These sampling points uniformly cover the entire target color gamut space.

[0040] In one implementation, in the sRGB linear space, 33 points are uniformly sampled within the 0-1 range of each of the R, G, and B channels, with a sampling interval of 1 / 32, generating a total of 35937 input sampling points, such as (0, 0, 0), (0, 1 / 32, 0), ..., (1, 1, 1). The corresponding corrected output value for each input sampling point is calculated using a global 3D color correction mapping, specifically as follows: For each 3D input sampling point, if the input sampling point falls exactly on the sampling node of the global color correction mapping, the mapping output value of that node is directly taken. If the input sampling point is located between sampling nodes, the output value is calculated using trilinear interpolation. The trilinear interpolation calculates the weighted average output based on the distance weight between the input point and the eight nearest nodes in the color space. In one implementation, the input sampling point (0.5, 0.5, 0.5) is interpolated based on the mapping values ​​of the surrounding 8 nodes through trilinear interpolation of the global color correction mapping to obtain the corrected output value (0.531, 0.484, 0.461). This output value represents the actual color value that the screen to be corrected should output when displaying the target color (0.5, 0.5, 0.5). The correspondence between all input sampling points and correction output values ​​is stored as a color correction lookup table for each screen to be calibrated, specifically as follows: All input sampling point coordinates and corresponding correction output values ​​are organized and stored in a preset data structure. The data structure header contains metadata such as the number of input channels (3), the number of output channels (3), and the number of sampling nodes for each channel (33). The volume data arranges the correction output values ​​of each node in order of priority of red channel, green channel, and blue channel. It can support real-time interpolation calculation of non-node input values ​​in hardware using tetrahedral interpolation.

[0041] Specifically, the weighted mixing steps for the color values ​​of adjacent screens after correction using their respective color correction lookup tables are as follows: Using the center line of the seam as a reference, a preset pixel width is extended to both sides as a blending band, specifically as follows: Taking a vertical seam as an example, the rightmost preset pixel column of the left side of the screen display area and the leftmost preset pixel column of the right side of the screen display area are taken as the fusion zone, and the width of the fusion zone on both sides is the preset number of pixels; In one implementation, taking a vertical seam as an example, according to the geometric normalization result, the left edge of the vertical seam is located in column 477, the right edge is located in column 480, the seam itself is approximately 2 pixels wide, occupying the physical gap between columns 478 and 479, and the half-width D of the blending band is set to 6.5 pixels, which only applies to the display pixels. The left screen blending band represents the last 6 pixel columns of its display area, specifically columns 472 to 477; the right screen blending band represents the first 6 pixel columns of its display area, specifically columns 480 to 485. The gaps in the physical area of ​​the seam, from column 478 to 479, are non-illuminated and are not weighted; they remain black. The same applies to the horizontal seams. The upper screen blending zone is the row range from row 262 to 267, and the lower screen blending zone is the row range from row 270 to 275. The physical area of ​​the seam, from row 268 to 269, remains black. For each pixel of an adjacent screen covered by the fusion band, the color value is corrected using the color correction lookup table of its respective screen, resulting in the corrected color value: Within the area covered by the fusion band, the pixels on the first side of the seam physically belong to the display area of ​​the first screen, and the pixels on the second side of the seam physically belong to the display area of ​​the second screen. For each pixel in the fusion band, the original input video signal color value corresponding to that pixel is obtained. The color value is fed into the color correction lookup table of the first screen and the color correction lookup table of the second screen for independent correction, resulting in a first corrected color value and a second corrected color value. Both corrected color values ​​have been mapped to a color space consistent with the visual perception of the reference screen by the lookup table of their respective screens.

[0042] In one embodiment, the pixel at column 474 within the fusion band is located on the reference screen side. Its original input RGB is normalized to (0.5, 0.5, 0.5). After correction by the reference screen LUT, the output first corrected color value is (0.502, 0.498, 0.495). After correction by the right screen LUT, the output second corrected color value is (0.531, 0.484, 0.461). For each pixel within the blending band, a blending weight is calculated based on its directed distance to the center line of the stitching seam. The blending weight approaches 1 at one edge of the blending band, smoothly transitions to 0.5 towards the center line of the stitching seam, and then smoothly transitions to 1 at the other edge of the blending band. Specifically: Using the seam center line (the vertical seam center line is located at column 478.5, i.e., the middle of column 478 and column 479; the horizontal seam center line is located at row 268.5, i.e., the middle of row 268 and row 269) as a unified benchmark, the directed distance d is defined as the vertical distance from the pixel to the seam center line. The directed distance of the pixels on the first side of the seam (left or top) is negative, and the directed distance of the pixels on the second side of the seam (right or bottom) is positive. The half-width D of the blending band is equal to 6.5 pixels. Taking the mixing weight of the first side correction color value as an example: the weight is 1 at the first side edge of the blending zone (d equals negative D), smoothly transitions towards the center line of the seam, reaches 0.5 at the center line of the seam (d equals 0), and continues to smoothly transition to 0 at the second side edge of the blending zone (d equals D). Specifically, when d is less than or equal to negative D, the mixing weight is equal to 1. When d is greater than negative D and less than or equal to 0, the mixed weight is equal to 0.5 minus 0.5 multiplied by d divided by D (because d is negative, the actual result smoothly decreases from 1 to 0.5). When d is greater than 0 and less than D, the mixed weight is equal to 0.5 minus 0.5 multiplied by d divided by D (because d is a positive value, the actual result smoothly decreases from 0.5 to 0); when d is greater than or equal to D, the mixed weight is equal to 0. The blending weight of the second-side corrected color value is 1 minus the blending weight, which smoothly transitions from 0 at the edge of the first side to 0.5 at the center line of the seam, and then smoothly transitions to 1 at the edge of the second side. This weight design is continuous and smooth throughout the blending zone, with equal weights on both sides at the center line of the seam.

[0043] In one implementation, the directed distance d of each pixel column is calculated based on the 478.5th column of the vertical seam centerline, and the blending weight is calculated. The half-width of the blending band D = 6.5 pixels, and the first side is the left side (the side where the reference screen is located) and the second side is the right side (the side where the screen to be calibrated is located). An example of the calculation results is shown in the table below: Example table of weight allocation for fusion band

[0044] The corrected color values ​​are then weighted and blended according to their respective blending weights to obtain the final output color value for that pixel. Specifically: The final output color value is equal to the first corrected color value multiplied by the weight of the first side plus the second corrected color value multiplied by 1 minus the weight of the first side. At the center line of the seam, the two corrected color values ​​each account for half of the weight. At the edge of the blending zone, the weight of the corrected color value of the screen to which the pixel belongs is close to 1 while the weight of the corrected color value of the screen on the opposite side is close to 0. The weighted and mixed color value is directly output to the screen driver board for display.

[0045] In one implementation, the final output at column 474 is approximately equal to 0.846 multiplied by (0.502, 0.498, 0.495) plus 0.154 multiplied by (0.531, 0.484, 0.461), which is approximately equal to (0.506, 0.496, 0.490). The final output at column 476 is approximately equal to 0.692 multiplied by (0.502, 0.498, 0.495) plus 0.308 multiplied by (0.531, 0.484, 0.461), which is approximately equal to (0.511, 0.494, 0.485). The weight of the first side at the center line of the seam is equal to 0.5. The output is the average of the correction values ​​on both sides. The splicing processor sends the weighted and mixed color value to the screen driver board for display.

[0046] Please see Figure 3 This invention provides a technical solution: a multi-screen splicing image color consistency correction system, comprising: a geometric self-calibration unit, used to acquire a multi-frame image sequence of the splicing screen and perform geometric normalization on the image sequence to obtain a front view image sequence, wherein the splicing screen includes at least a reference screen and the remaining screens as screens to be calibrated; and a response blind calibration unit, used to extract the seam area from the front view image sequence as a zero-radius reference, estimate the inverse response function of the camera using the zero-radius reference and the detected spatiotemporally consistent pixel blocks, and apply the inverse response function to convert the front view image sequence into a linear radius image sequence. The system comprises: a screen-to-screen color mapping estimation unit, which defines sampling bands belonging to adjacent screens along the seam on both sides of the linear radiance image sequence; extracts cross-screen pixel pairs from the sampling bands based on local flatness conditions; and estimates the color mapping function from each screen to be calibrated to the reference screen through multi-frame statistical regression; and a color calibration fusion unit, which generates an independent color calibration lookup table for each screen to be calibrated based on the color mapping function, defines a fusion band on both sides of the seam, and performs a weighted fusion of the color values ​​of adjacent screens after calibration using their respective color calibration lookup tables within the fusion band to obtain the final output color value.

[0047] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the invention.

[0048] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for color consistency correction of multi-screen spliced ​​images, characterized in that, Includes the following steps: A multi-frame image sequence of the splicing screen is acquired, and the image sequence is geometrically normalized to obtain a front view image sequence. The splicing screen includes at least a reference screen and the remaining screens as screens to be calibrated. The stitching region is extracted from the front view image sequence as a zero-radiance reference. Using the zero-radiance reference and the detected spatiotemporally consistent pixel blocks, the inverse response function of the camera is estimated, and the inverse response function is applied to convert the front view image sequence into a linear radiance image sequence. In the linear radiance image sequence, sampling bands belonging to adjacent screens are defined on both sides of the seam. Cross-screen pixel pairs are extracted from the sampling bands based on local flatness conditions. The color mapping function from each screen to be calibrated to the reference screen is estimated through multi-frame statistical regression. Based on the color mapping function, an independent color correction lookup table is generated for each screen to be corrected, and a fusion band is defined on both sides of the splicing seam. Within the fusion band, the color values ​​of adjacent screens after being corrected by their respective color correction lookup tables are weighted and mixed to serve as the final output, thereby achieving color consistency correction for multi-screen spliced ​​images.

2. The method for color consistency correction of multi-screen spliced ​​images according to claim 1, characterized in that, The specific steps for geometric normalization of image sequences are as follows: Select one frame from a multi-frame image sequence and detect the straight line formed by the stitching seam; The detected straight lines are divided into horizontal and vertical groups according to their angles. Hierarchical clustering is performed on the straight lines in each group, and the mean straight line of each cluster is taken to obtain the horizontal seam line and the vertical seam line respectively. The corner points of each screen are determined by the intersection of the horizontal and vertical seam lines, and the four outermost corner points are taken as the source vertices of the perspective transformation. Using the rectangle enclosed by the corner points of the reference screen as a reference, set the width and height of the target rectangle, and use the vertices of the target rectangle as the target vertices of the perspective transformation; The homography matrix from the source vertex to the target vertex is solved by direct linear transformation. Perspective transformation is then performed on each frame of the image sequence to obtain the frontal view image sequence.

3. The method for color consistency correction of multi-screen spliced ​​images according to claim 1, characterized in that, The specific steps for detecting spatiotemporally consistent pixel blocks are as follows: Within the display area of ​​each screen, multiple sampling areas are preset, and a preset range of neighborhoods is taken as candidate image blocks for each pixel; For candidate image blocks within the sampling area, a spatial smoothing condition is applied for filtering: candidate image blocks with a brightness standard deviation lower than a first preset threshold are selected. For candidate image blocks that have passed the spatial smoothing screening, a short-term temporal stability condition screening is applied: candidate image blocks whose average brightness change is lower than the second preset threshold within a consecutive preset number of frames are selected. For candidate image blocks that pass the short-term temporal domain stability screening, apply the long-term temporal domain smoothing monotonic condition screening: track the average brightness time series of the image block in all frames, and select candidate image blocks whose time series is monotonic and whose cumulative amplitude of reverse fluctuation within the series does not exceed a preset proportion of the cumulative amplitude of positive change. Candidate image blocks that pass the above screening are included as spatiotemporally consistent pixel blocks.

4. The method for color consistency correction of multi-screen spliced ​​images according to claim 3, characterized in that, The specific steps for estimating the inverse response function of the camera are as follows: In the front view image sequence, the area within a preset pixel width on both sides of the seam is extracted as the zero-radiance reference area; Calculate the temporal mean of all pixels in the zero-radiance reference area across multiple frames to obtain the dark current bias vector of the camera, and calculate the temporal variance of each pixel across multiple frames to obtain the stitching seam leakage noise variance. The dark current bias vector provided by the zero-radiance reference region is used as the zero-radiance hard constraint. Spatiotemporally consistent pixel blocks are used as data constraints; Dark area prior constraints are constructed based on the variance of light leakage noise at the seams; Under the combined constraints of zero radiance hard constraint, data constraint, and dark area prior constraint, the inverse response function of the camera's color channel is obtained.

5. The method for color consistency correction of multi-screen spliced ​​images according to claim 2, characterized in that, The specific steps for extracting cross-screen pixel pairs based on local flatness conditions are as follows: Use the horizontal and vertical seam lines as the center lines of the seam; On both sides of the center line of each seam, a range from the first preset pixel width to the second preset pixel width of the center line of the seam is taken, and these ranges are respectively used as sampling bands belonging to adjacent screens; For each frame of linear radiance image, extract a pair of cross-screen pixels pixel by pixel along the center line of the seam within the sampling band of each adjacent screen; Take the preset percentile of the brightness standard deviation of all sliding windows in the preset neighborhood range along the center line of the stitching in the whole frame as the adaptive flatness threshold; Based on the adaptive flatness threshold, a local flatness test is performed on each pair of cross-screen pixels, and the pair of cross-screen pixels that passes the test is taken as the cross-screen pixel pair.

6. The method for color consistency correction of multi-screen spliced ​​images according to claim 5, characterized in that, The specific steps for local flatness inspection are as follows: For each pair of cross-screen pixels, calculate the first luminance variance along the center line of the seam within the sampling band on the first side of the seam, the second luminance variance perpendicular to the center line of the seam within the sampling band on the first side of the seam, the third luminance variance along the center line of the seam within the sampling band on the second side of the seam, and the fourth luminance variance perpendicular to the center line of the seam within the sampling band on the second side of the seam. When the first luminance variance, the second luminance variance, the third luminance variance, and the fourth luminance variance are all lower than the adaptive flatness threshold, the pair of cross-screen pixels is determined to have passed the local flatness test.

7. The method for color consistency correction of multi-screen spliced ​​images according to claim 1, characterized in that, The specific steps for estimating the color mapping function from each screen to be calibrated to the reference screen are as follows: The value of each pixel in the color channel of the linear radiance image sequence is used as the channel linear radiance value, and the color channel is used as the coordinate axis to construct the color space. Using the linear radiance value of the channel of the screen to be calibrated as the coordinate, the color space is evenly divided within the value range of each color channel to obtain several sampling nodes. For each sampling node, the weight of each cross-screen pixel pair is calculated with the node as the center and the distance in the color space, and a confidence factor based on the frequency of occurrence of the cross-screen pixel pair is introduced into the weight. Weighted regression is performed on cross-screen pixel pairs in the neighborhood of each sampling node to obtain the initial mapping output value of each sampling node, and prior regularization is applied. For sampling nodes with a weighted sample count lower than a preset threshold in the neighborhood, interpolation of neighboring nodes is used to fill the gaps and obtain the color mapping function from each screen to be calibrated to the reference screen.

8. The method for color consistency correction of multi-screen spliced ​​images according to claim 7, characterized in that, The specific steps for generating an independent color correction lookup table for each screen to be calibrated are as follows: Take the color mapping slice at the center of the corresponding screen to be calibrated, and use it as the global color calibration map for each screen to be calibrated; Uniform sampling is performed within the value range of each color channel in the color space to generate several input sampling points; The corresponding correction output value is calculated for each input sampling point through a global 3D color correction mapping. The correspondence between all input sampling points and the correction output values ​​is stored as a color correction lookup table for each screen to be calibrated.

9. The method for color consistency correction of multi-screen spliced ​​images according to claim 5, characterized in that, The specific steps for weighted mixing of the color values ​​of adjacent screens after correction using their respective color correction lookup tables are as follows: Using the center line of the seam as a reference, extend a preset pixel width to both sides as a blending zone; For each pixel of the adjacent screens covered by the fusion band, the color correction lookup table of its respective screen is used to correct the color value of each pixel. For each pixel within the fusion band, calculate the blending weight based on its directed distance to the center line of the stitching seam. The corrected color values ​​are weighted and mixed according to the mixing weights to obtain the final output color value for that pixel.

10. A multi-screen spliced ​​image color consistency correction system, employing the multi-screen spliced ​​image color consistency correction method according to any one of claims 1-9, characterized in that, include: The geometric self-calibration unit is used to acquire a multi-frame image sequence of the splicing screen and perform geometric normalization on the image sequence to obtain a front view image sequence. The splicing screen includes at least a reference screen and the remaining screens as screens to be calibrated. The response blind calibration unit is used to extract the stitching region from the front view image sequence as a zero-radiance reference, use the zero-radiance reference and the detected spatiotemporally consistent pixel block to estimate the inverse response function of the camera, and apply the inverse response function to convert the front view image sequence into a linear radiance image sequence. The inter-screen color mapping estimation unit is used to define sampling bands belonging to adjacent screens along the seam on both sides of the linear radiance image sequence, extract cross-screen pixel pairs from the sampling bands based on local flatness conditions, and estimate the color mapping function from each screen to be calibrated to the reference screen through multi-frame statistical regression. The color correction fusion unit is used to generate an independent color correction lookup table for each screen to be corrected based on the color mapping function, and to define a fusion band on both sides of the seam. Within the fusion band, the color values ​​of adjacent screens after being corrected by their respective color correction lookup tables are weighted and mixed to obtain the final output color value.

Images