Ultrahigh-pixel image splicing method based on deep neural network

Abstract

The invention discloses an ultrahigh-pixel image splicing method based on a deep neural network, relates to the technical field of image compensation splicing, and solves the technical problem that spliced images are mutually unnatural in color and brightness due to deviation of RGB and brightness of multiple images. The method comprises the following steps: extracting similar areas of a target image, obtaining RGB deviation based on RGB values of pixel points in the two similar areas, and obtaining a brightness deviation rate based on brightness values of the pixel points in the two similar areas; an operation instruction is sent based on the RGB deviation and the brightness deviation rate, and RGB value and brightness value compensation is carried out on the to-be-spliced image with the large number in the target image based on the operation instruction; inputting the image set subjected to RGB value and brightness value compensation into an image splicing model to obtain a composite image; according to the invention, the quality and efficiency of image splicing can be improved.

Description

Technical Field

[0001] The present invention belongs to the field of image processing and relates to image compensation stitching technology, specifically to an ultra-high pixel image stitching method based on a deep neural network. Background Art

[0002] Image stitching is the process of stitching two or more images with overlapping areas and similar viewpoints, which is equivalent to taking the union operation of multiple images, obtaining as much image information as possible, saving it in one image, and merging multiple images into a larger image while maintaining the same clarity, so as to show a wider field of view or more detailed scene information. This technology is often used to create panoramic images, such as in tourism photography, drone aerial photography, satellite remote sensing, and video surveillance.

[0003] At present, most ultra-high pixel image stitching methods based on deep neural networks ignore the fact that the same area photographed at different angles may have RGB and brightness deviations when stitching images. If the images with deviations are directly stitched together, the stitched images will be unnatural in color and brightness. If the steps of eliminating the deviations of RGB values ​​and brightness values ​​are placed in the stitching process, it will lead to frequent cross-calling of the elimination steps and the stitching steps, resulting in reduced stitching efficiency and increased system computational complexity.

[0004] Therefore, the present invention discloses an ultra-high pixel image stitching method based on deep neural network to solve the above technical problems. Summary of the invention

[0005] The present invention aims to solve at least one of the technical problems existing in the prior art; to this end, the present invention proposes an ultra-high pixel image stitching method based on a deep neural network, which is used to solve the technical problem that the stitched images are unnatural in color and brightness due to the deviation of RGB and brightness of multiple images themselves. The present invention extracts similar areas of the target image, obtains the RGB deviation based on the RGB values ​​of each pixel in the two similar areas, and obtains the brightness deviation rate based on the brightness values ​​of each pixel in the two similar areas; issues an operation instruction based on the RGB deviation and the brightness deviation rate, and compensates the RGB value and brightness value of the image to be stitched with a larger number in the target image based on the operation instruction; and inputs the image set after the RGB value and brightness value compensation into the image stitching model to obtain a synthetic image to solve the above problem.

[0006] To achieve the above object, a first aspect of the present invention provides an ultra-high pixel image stitching method based on a deep neural network, comprising:

[0007] Get the number of the image to be stitched in the current image set, and get the RGB value and brightness value of each pixel in the image to be stitched;

[0008] Mark the two images to be stitched with the smallest numbers as target images, extract similar areas of the target images, obtain RGB deviation based on the RGB values ​​of each pixel in the two similar areas, and obtain brightness deviation rate based on the brightness values ​​of each pixel in the two similar areas; issue an operation instruction based on the RGB deviation and the brightness deviation rate, and compensate the RGB value and brightness value of the image to be stitched with a larger number in the target image based on the operation instruction; wherein the similar areas are obtained by manual selection;

[0009] Delete the numbers of the two images to be stitched that are to be compensated, mark the image to be output as the image to be stitched and mark the number as 1, re-mark the target image and perform RGB value and brightness value compensation operations on the target image, and stop the compensation operation when traversing to the last image to be stitched;

[0010] The image set after RGB value and brightness value compensation is input into the image stitching model to obtain a synthetic image; wherein the image stitching model is obtained by training a deep neural network model.

[0011] It is worth noting that in the acquisition of ultra-high pixel images, images of the same area acquired at different angles may have multiple images with different overall RGB values ​​and overall brightness values. Directly splicing such images will cause the spliced ​​images to be unnatural in color and brightness. If the step of eliminating the deviation of RGB value and brightness value is placed in the process of splicing, it will lead to frequent cross-calling of the elimination step and the splicing step, resulting in reduced splicing efficiency and increased system calculation amount. Therefore, the present invention extracts the RGB deviation and brightness deviation rate of the image to be spliced ​​before splicing the image, eliminates the differences in color and brightness of multiple images, and reduces the abruptness between the spliced ​​images.

[0012] Preferably, obtaining the RGB deviation based on the RGB value of each pixel in two similar areas includes:

[0013] Sequentially mark the RGB values ​​of corresponding pixels in two similar regions as corresponding RGB values, and obtain the difference between the intensity values ​​of the three colors R, G, and B in the corresponding RGB values; wherein the corresponding pixel points are the pixel points at the same position in the two similar regions;

[0014] The difference values ​​of the intensity values ​​of the three colors R, G, and B in each corresponding pixel point are divided into A1 difference group, A2 difference group, and A3 difference group according to the color type;

[0015] Extract the A1 difference group, the A2 difference group and the A3 difference group in sequence, obtain the variance of the data in the extracted difference group, and judge whether the variance is less than the corresponding variance threshold; if yes, retain the data of the difference group; if no, remove the data with the largest difference from the data mode in the difference group, and re-judge the variance until the variance of the difference group is less than the corresponding variance threshold, and retain the remaining data of the difference group; wherein the variance threshold is obtained through experience;

[0016] The maximum value ZD, the minimum value ZX and the mode ZS of the retained data in the difference group are obtained, and the RGB deviation Pi of the difference group corresponding to the two similar areas is obtained based on the formula Pi=(α1×ZD+α2×ZX+ZS) / 2; wherein α1 and α2 are both proportional adjustment coefficients greater than 0, and α1>α2, α1+α2=1; wherein the RGB deviation Pi includes the difference P1 of the R color intensity value, the difference P2 of the G color intensity value, and the difference P3 of the B color intensity value.

[0017] It is worth noting that the present invention performs a fluctuation value removal operation on each group when obtaining the RGB deviation. The fluctuation value that has a greater impact on the difference of each color intensity value is removed through the variance threshold, and a group of color intensity deviation groups that can represent the current similar area is obtained. Then, feature value extraction is performed on this deviation group to obtain the color intensity deviation representing the overall color of the two images to be stitched, which provides data support for subsequent RGB value compensation.

[0018] Preferably, obtaining the brightness deviation rate based on the brightness value of each pixel point in two similar areas includes:

[0019] Extract the brightness values ​​of the corresponding pixels in the two similar areas, mark the brightness values ​​of the corresponding pixels in the similar areas of the image to be stitched with the smaller number as LD1, mark the brightness values ​​of the corresponding pixels in the similar areas of the image to be stitched with the larger number as LD2, and subtract LD1 from LD2 to obtain the brightness change value of each corresponding pixel;

[0020] Get the mode of the brightness change value, mark the corresponding pixel point corresponding to the mode as the target pixel point, extract LD1 and LD2 of the target pixel point, and get the preliminary brightness deviation rate CL of each target pixel point based on the formula CL=(LD2-LD1) / LD1. Calculate the average value of several preliminary brightness deviation rates CL to get the brightness deviation rate of two similar areas.

[0021] It is worth noting that when the present invention compensates for the stitched image, the RGB value is compensated by RGB deviation, and the brightness value is compensated by brightness deviation rate. This is because when acquiring regional images at different angles, the deviation ratio of the brightness value caused by illumination is consistent at each pixel point, so the brightness value is compensated by the brightness deviation rate. Using the brightness deviation rate for compensation can simplify the processing process and improve calculation efficiency; when acquiring regional images at different angles, the angle change will change the reflection characteristics of the object surface, thereby affecting the RGB value of the image. The deviation ratio of the RGB value of each pixel point caused by this reflection characteristic is different, so the deviation ratio cannot be used to compensate the RGB value deviation of each pixel point of the entire image.

[0022] Preferably, issuing an operation instruction based on the RGB deviation and the brightness deviation rate includes:

[0023] Extract the difference P1 of the R color intensity value, the difference P2 of the G color intensity value, and the difference P3 of the B color intensity value; determine whether the difference P1 is within the defined range 1; if yes, do not perform any operation; if no, issue an operation instruction 1;

[0024] Determine whether the difference value P2 is within the second defined range; if yes, do not perform any operation; if no, issue the second operation instruction;

[0025] Determine whether the difference value P3 is within the defined range 3; if yes, do not perform any operation; if no, issue the third operation instruction; wherein the defined range 1, the defined range 2, the defined range 3 and the deviation rate defined range are all obtained through experience;

[0026] Extract the brightness deviation rate LP and determine whether the brightness deviation rate LP is within the deviation rate definition range; if yes, do not perform any operation; if no, issue operation instruction four; wherein the operation instructions include operation instruction one, operation instruction two, operation instruction three and operation instruction four.

[0027] Preferably, the step of compensating the RGB value and brightness value of the image to be spliced ​​with a larger number in the target image based on the operation instruction includes:

[0028] Receive an operation instruction, extract the difference value P1 when receiving the operation instruction 1, and determine whether the difference value P1 is greater than zero; if yes, reduce the R color intensity value of each pixel point in the image to be spliced ​​with a larger number in the target image by P1 intensity value; if no, increase the R color intensity value of each pixel point in the image to be spliced ​​with a larger number in the target image by P1 intensity value;

[0029] When the second operation instruction is received, the difference value P2 is extracted, and it is determined whether the difference value P2 is greater than zero; if it is, the G color intensity value of each pixel point in the image to be spliced ​​with a larger number in the target image is reduced by P2 intensity values; if not, the G color intensity value of each pixel point in the image to be spliced ​​with a larger number in the target image is increased by P2 intensity values;

[0030] When the operation instruction three is received, the difference value P3 is extracted, and it is determined whether the difference value P3 is greater than zero; if yes, the B color intensity value of each pixel point in the image to be spliced ​​with a larger number in the target image is reduced by P3 intensity values; if no, the B color intensity value of each pixel point in the image to be spliced ​​with a larger number in the target image is increased by P3 intensity values;

[0031] When operation instruction 4 is received, the brightness deviation rate LP is extracted to determine whether the brightness deviation rate LP is greater than zero; if yes, the brightness value of each pixel in the target image to be spliced ​​with a larger number is multiplied by α1×(1-LP) to obtain the compensated brightness value of each pixel; if no, the brightness value of each pixel in the target image to be spliced ​​with a larger number is multiplied by α2×(1+|LP|) to obtain the compensated brightness value of each pixel; wherein α1 and α2 are both error adjustment coefficients, and the value range of α1 and α2 is (0,2).

[0032] Preferably, the image stitching model is obtained by training a deep neural network model, including:

[0033] Several image sets are extracted from the database, and the synthetic images corresponding to each image set are integrated into several groups of training data and test data; the artificial intelligence model is trained using the training data; the trained artificial intelligence model is tested using the test data, and the artificial intelligence model is adjusted according to the test results; finally, an image stitching model with image sets as input and synthetic images as output is obtained.

[0034] Specifically, the image set in the test data is input into the trained artificial intelligence model to obtain the corresponding synthetic image, which is compared with the synthetic image included in the test data. When the difference between the two meets the expert's expectations, there is no need to adjust the parameters, and the next set of test data is tested. If it does not meet the expert's expectations, the corresponding parameters are adjusted until the output synthetic image of the corresponding test data meets the expert's expectations, and then the next set of test data is tested. When the number of test data whose synthetic images meet the expert's expectations among all test data accounts for 90% or more of the total test data, an image stitching model with an input of an image set and an output of a synthetic image is obtained.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] 1. The present invention extracts similar areas of a target image, obtains RGB deviation based on the RGB value of each pixel in the two similar areas, and obtains the brightness deviation rate based on the brightness value of each pixel in the two similar areas; issues an operation instruction based on the RGB deviation and the brightness deviation rate, and compensates the RGB value and brightness value of the image to be spliced ​​with a larger number in the target image based on the operation instruction; inputs the image set after the RGB value and brightness value compensation into an image splicing model to obtain a composite image, thereby solving the technical problem that the color and brightness of the spliced ​​image are unnatural due to the deviation of RGB and brightness of multiple images themselves, and improving the quality and efficiency of image splicing.

[0037] 2. The present invention adopts a fluctuation value filtering method to optimize data quality when calculating RGB deviation. By setting a variance threshold, the fluctuation data that affects the significance of the color intensity difference is removed, thereby screening out a group of representative values ​​reflecting the color intensity deviation of the current similar area. Then, by performing characteristic value analysis on the deviation group, the overall color intensity deviation between the two images to be spliced ​​can be obtained, which provides reliable data support for subsequent RGB value compensation and ensures the accuracy of color correction during the splicing process. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0039] Figure 1 It is a schematic diagram of the operation steps of the present invention;

[0040] Figure 2 A schematic diagram of the operation steps for obtaining RGB deviations according to the present invention;

[0041] Figure 3 The figure is a schematic diagram of the operation steps of performing RGB value and brightness value compensation in the present invention. DETAILED DESCRIPTION

[0042] The technical solution of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0043] See also Figure 1-Figure 2The first aspect of the present invention provides an ultra-high pixel image stitching method based on a deep neural network, comprising:

[0044] Get the number of the image to be stitched in the current image set, and get the RGB value and brightness value of each pixel in the image to be stitched;

[0045] Mark the two images to be stitched with the smallest numbers as target images, extract similar areas of the target images, obtain RGB deviation based on the RGB values ​​of each pixel in the two similar areas, and obtain brightness deviation rate based on the brightness values ​​of each pixel in the two similar areas; issue an operation instruction based on the RGB deviation and the brightness deviation rate, and compensate the RGB value and brightness value of the image to be stitched with a larger number in the target image based on the operation instruction; wherein the similar areas are obtained by manual selection;

[0046] Delete the numbers of the two images to be stitched that are to be compensated, mark the image to be output as the image to be stitched and mark the number as 1, re-mark the target image and perform RGB value and brightness value compensation operations on the target image, and stop the compensation operation when traversing to the last image to be stitched;

[0047] The image set after RGB value and brightness value compensation is input into the image stitching model to obtain a synthetic image; wherein the image stitching model is obtained by training a deep neural network model.

[0048] It should be noted that an image set is a collection of images to be stitched.

[0049] It should be noted that the serial numbers of the images to be stitched are obtained by manual setting.

[0050] In this application, the RGB deviation is obtained based on the RGB values ​​of each pixel in two similar areas, including:

[0051] Sequentially mark the RGB values ​​of corresponding pixels in two similar regions as corresponding RGB values, and obtain the difference between the intensity values ​​of the three colors R, G, and B in the corresponding RGB values; wherein the corresponding pixel points are the pixel points at the same position in the two similar regions;

[0052] The difference values ​​of the intensity values ​​of the three colors R, G, and B in each corresponding pixel point are divided into A1 difference group, A2 difference group, and A3 difference group according to the color type;

[0053] Extract the A1 difference group, A2 difference group and A3 difference group in sequence, obtain the variance of the data in the extracted difference group, and judge whether the variance is less than the corresponding variance threshold; if yes, retain the data of the difference group; if no, remove the data with the largest difference from the data mode in the difference group, and re-judge the variance until the variance of the difference group is less than the corresponding variance threshold, and retain the remaining data of the difference group; wherein the variance threshold is obtained through experience;

[0054] The maximum value ZD, the minimum value ZX and the mode ZS of the retained data in the difference group are obtained, and the RGB deviation Pi of the difference group corresponding to the two similar areas is obtained based on the formula Pi=(α1×ZD+α2×ZX+ZS) / 2; wherein α1 and α2 are both proportional adjustment coefficients greater than 0, and α1>α2, α1+α2=1; wherein the RGB deviation Pi includes the difference P1 of the R color intensity value, the difference P2 of the G color intensity value, and the difference P3 of the B color intensity value.

[0055] It should be noted that the difference between the intensity values ​​of the three colors R, G, and B in the corresponding RGB values ​​is obtained by subtracting the RGB value corresponding to the image to be spliced ​​with a smaller number from the RGB value corresponding to the image to be spliced ​​with a larger number.

[0056] It should be noted that, when the difference in the intensity values ​​of the three colors R, G, and B corresponding to each pixel point is divided into A1 difference group, A2 difference group, and A3 difference group according to the color type, the A1 difference group is a difference group of the R color intensity value, the A2 difference group is a difference group of the G color intensity value, and the A3 difference group is a difference group of the B color intensity value.

[0057] In this application, the brightness deviation rate is obtained based on the brightness value of each pixel point in two similar areas, including:

[0058] Extract the brightness values ​​of the corresponding pixels in the two similar areas, mark the brightness values ​​of the corresponding pixels in the similar areas of the image to be stitched with the smaller number as LD1, mark the brightness values ​​of the corresponding pixels in the similar areas of the image to be stitched with the larger number as LD2, and subtract LD1 from LD2 to obtain the brightness change value of each corresponding pixel;

[0059] Get the mode of the brightness change value, mark the corresponding pixel point corresponding to the mode as the target pixel point, extract the target pixel point LD1 and LD2, and get the preliminary brightness deviation rate CL of each target pixel point based on the formula CL=(LD2-LD1) / LD1. Calculate the average value of several preliminary brightness deviation rates CL to get the brightness deviation rate of two similar areas.

[0060] In this application, operation instructions are issued based on RGB deviation and brightness deviation rate, including:

[0061] Extract the difference P1 of the R color intensity value, the difference P2 of the G color intensity value, and the difference P3 of the B color intensity value; determine whether the difference P1 is within the defined range 1; if yes, do not perform any operation; if no, issue an operation instruction 1;

[0062] Determine whether the difference value P2 is within the defined range 2; if yes, do not perform any operation; if no, issue the second operation instruction;

[0063] Determine whether the difference value P3 is within the defined range 3; if yes, do not perform any operation; if no, issue the operation instruction 3; wherein, the defined range 1, the defined range 2, the defined range 3 and the defined range of the deviation rate are all obtained through experience;

[0064] Extract the brightness deviation rate LP and determine whether the brightness deviation rate LP is within the deviation rate definition range; if yes, do not perform any operation; if no, issue operation instruction four; wherein the operation instructions include operation instruction one, operation instruction two, operation instruction three and operation instruction four.

[0065] It should be noted that when each difference value Pi or brightness deviation rate LP is not within the corresponding defined range, it indicates that the deviation is large and a compensation operation is required.

[0066] See also Figure 3 In this application, based on the operation instruction, the RGB value and brightness value compensation of the image to be stitched with a larger number in the target image is performed, including:

[0067] Receive an operation instruction, extract the difference value P1 when receiving the operation instruction 1, and determine whether the difference value P1 is greater than zero; if yes, reduce the R color intensity value of each pixel point in the image to be spliced ​​with a larger number in the target image by P1 intensity value; if no, increase the R color intensity value of each pixel point in the image to be spliced ​​with a larger number in the target image by P1 intensity value;

[0068] When the second operation instruction is received, the difference value P2 is extracted to determine whether the difference value P2 is greater than zero; if yes, the G color intensity value of each pixel point in the target image with a larger number to be stitched is reduced by P2 intensity values; if no, the G color intensity value of each pixel point in the target image with a larger number to be stitched is increased by P2 intensity values;

[0069] When the operation instruction three is received, the difference value P3 is extracted to determine whether the difference value P3 is greater than zero; if yes, the B color intensity value of each pixel point in the image to be spliced ​​with a larger number in the target image is reduced by P3 intensity values; if no, the B color intensity value of each pixel point in the image to be spliced ​​with a larger number in the target image is increased by P3 intensity values;

[0070] When operation instruction 4 is received, the brightness deviation rate LP is extracted to determine whether the brightness deviation rate LP is greater than zero; if yes, the brightness value of each pixel in the target image to be spliced ​​with a larger number is multiplied by α1×(1-LP) to obtain the compensated brightness value of each pixel; if no, the brightness value of each pixel in the target image to be spliced ​​with a larger number is multiplied by α2×(1+|LP|) to obtain the compensated brightness value of each pixel; wherein α1 and α2 are both error adjustment coefficients, and the value range of α1 and α2 is (0,2).

[0071] The image stitching model in this application is trained by a deep neural network model, including:

[0072] Several image sets are extracted from the database, and the synthetic images corresponding to each image set are integrated into several groups of training data and test data; the artificial intelligence model is trained using the training data; the trained artificial intelligence model is tested using the test data, and the artificial intelligence model is adjusted according to the test results; finally, an image stitching model with image sets as input and synthetic images as output is obtained.

[0073] Specifically, the image set in the test data is input into the trained artificial intelligence model to obtain the corresponding synthetic image, which is compared with the synthetic image included in the test data. When the difference between the two meets the expert's expectations, there is no need to adjust the parameters, and the next set of test data is tested. If it does not meet the expert's expectations, the corresponding parameters are adjusted until the output synthetic image of the corresponding test data meets the expert's expectations, and then the next set of test data is tested. When the number of test data whose synthetic images meet the expert's expectations among all test data accounts for 90% or more of the total test data, an image stitching model with an input of an image set and an output of a synthetic image is obtained.

[0074] It should be noted that the synthetic images included in the test data in the image stitching model are obtained through manual stitching.

[0075] Part of the data in the above formula is calculated by removing the dimension and taking its numerical value. The formula is a formula closest to the actual situation obtained by software simulation of a large amount of collected data; the preset parameters and preset thresholds in the formula are set by technical personnel in this field according to actual conditions or obtained through simulation of a large amount of data.

[0076] Working principle of the present invention:

[0077] The numbers of the images to be stitched in the current image set are obtained, and the RGB value and brightness value of each pixel in the images to be stitched are obtained; the two images to be stitched with the smallest numbers are marked as target images, similar areas of the target images are extracted, RGB deviations are obtained based on the RGB values ​​of each pixel in the two similar areas, and brightness deviation rates are obtained based on the brightness values ​​of each pixel in the two similar areas; operation instructions are issued based on the RGB deviations and brightness deviation rates, and RGB and brightness value compensation is performed on the image to be stitched with a larger number in the target images based on the operation instructions; the numbers of the two images to be stitched that are subjected to the compensation operation are deleted, the image to be output is marked as the image to be stitched and the number is marked as 1, the target image is re-marked, and the RGB and brightness value compensation operation is performed on the target image, until the compensation operation is stopped when the last image to be stitched is traversed; the image set after the RGB and brightness value compensation is performed is input into the image stitching model to obtain a composite image.

[0078] The above embodiments are only used to illustrate the technical method of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical method of the present invention.

Claims

1. A super-high pixel image stitching method based on deep neural network, characterized in that: include: Get the number of the image to be stitched in the current image set, and get the RGB value and brightness value of each pixel in the image to be stitched; Mark the two images to be stitched with the smallest numbers as target images, extract similar areas of the target images, obtain RGB deviation based on the RGB values ​​of each pixel in the two similar areas, and obtain brightness deviation rate based on the brightness values ​​of each pixel in the two similar areas; issue an operation instruction based on the RGB deviation and the brightness deviation rate, and compensate the RGB value and brightness value of the image to be stitched with a larger number in the target image based on the operation instruction; Delete the numbers of the two images to be stitched that are to be compensated, mark the image to be output as the image to be stitched and mark the number as 1, re-mark the target image and perform RGB value and brightness value compensation operations on the target image, and stop the compensation operation when traversing to the last image to be stitched; The image set after RGB value and brightness value compensation is input into the image stitching model to obtain a synthetic image; wherein the image stitching model is obtained by training a deep neural network model.

2. The ultra-high pixel image stitching method based on deep neural network according to claim 1 is characterized in that: The obtaining the RGB deviation based on the RGB value of each pixel point in two similar areas includes: Sequentially mark the RGB values ​​of corresponding pixels in two similar regions as corresponding RGB values, and obtain the difference between the intensity values ​​of the three colors R, G, and B in the corresponding RGB values; wherein the corresponding pixel points are the pixel points at the same position in the two similar regions; The difference values ​​of the intensity values ​​of the three colors R, G, and B in each corresponding pixel point are divided into A1 difference group, A2 difference group, and A3 difference group according to the color type; Extract the A1 difference group, the A2 difference group, and the A3 difference group in sequence, obtain the variance of the data in the extracted difference group, and judge whether the variance is less than the corresponding variance threshold; if yes, retain the data in the difference group; if no, remove the data in the difference group that differs the most from the data mode, and re-judge the variance until the variance of the difference group is less than the corresponding variance threshold, and retain the remaining data in the difference group; The maximum value ZD, the minimum value ZX and the mode ZS of the retained data in the difference group are obtained, and the RGB deviation Pi of the difference group corresponding to the two similar areas is obtained based on the formula Pi=(α1×ZD+α2×ZX+ZS) / 2; wherein α1 and α2 are both proportional adjustment coefficients greater than 0, and α1>α2, α1+α2=1; wherein the RGB deviation Pi includes the difference P1 of the R color intensity value, the difference P2 of the G color intensity value, and the difference P3 of the B color intensity value.

3. The ultra-high pixel image stitching method based on deep neural network according to claim 2 is characterized in that: The step of obtaining the brightness deviation rate based on the brightness value of each pixel point in two similar areas includes: Extract the brightness values ​​of the corresponding pixels in the two similar areas, mark the brightness values ​​of the corresponding pixels in the similar areas of the image to be stitched with the smaller number as LD1, mark the brightness values ​​of the corresponding pixels in the similar areas of the image to be stitched with the larger number as LD2, and subtract LD1 from LD2 to obtain the brightness change value of each corresponding pixel; Get the mode of the brightness change value, mark the corresponding pixel point corresponding to the mode as the target pixel point, extract LD1 and LD2 of the target pixel point, and get the preliminary brightness deviation rate CL of each target pixel point based on the formula CL=(LD2-LD1) / LD1. Calculate the average value of several preliminary brightness deviation rates CL to get the brightness deviation rate of two similar areas.

4. The ultra-high pixel image stitching method based on deep neural network according to claim 3 is characterized in that: The issuing of an operation instruction based on the RGB deviation and the brightness deviation rate includes: Extract the difference P1 of the R color intensity value, the difference P2 of the G color intensity value, and the difference P3 of the B color intensity value; determine whether the difference P1 is within the defined range 1; if yes, do not perform any operation; if no, issue an operation instruction 1; Determine whether the difference value P2 is within the second defined range; if yes, do not perform any operation; if no, issue the second operation instruction; Determine whether the difference value P3 is within the defined range three; if yes, do not perform any operation; if no, issue the operation instruction three; Extract the brightness deviation rate LP and determine whether the brightness deviation rate LP is within the deviation rate definition range; if yes, do not perform any operation; if no, issue operation instruction four; wherein the operation instructions include operation instruction one, operation instruction two, operation instruction three and operation instruction four.

5. The ultra-high pixel image stitching method based on deep neural network according to claim 4 is characterized in that: The method of compensating the RGB value and brightness value of the image to be spliced ​​with a larger number in the target image based on the operation instruction includes: Receive an operation instruction, extract the difference value P1 when receiving the operation instruction 1, and determine whether the difference value P1 is greater than zero; if yes, reduce the R color intensity value of each pixel point in the image to be spliced ​​with a larger number in the target image by P1 intensity value; if no, increase the R color intensity value of each pixel point in the image to be spliced ​​with a larger number in the target image by P1 intensity value; When the second operation instruction is received, the difference value P2 is extracted, and it is determined whether the difference value P2 is greater than zero; if it is, the G color intensity value of each pixel point in the image to be spliced ​​with a larger number in the target image is reduced by P2 intensity values; if not, the G color intensity value of each pixel point in the image to be spliced ​​with a larger number in the target image is increased by P2 intensity values; When the operation instruction three is received, the difference value P3 is extracted, and it is determined whether the difference value P3 is greater than zero; if yes, the B color intensity value of each pixel point in the image to be spliced ​​with a larger number in the target image is reduced by P3 intensity values; if no, the B color intensity value of each pixel point in the image to be spliced ​​with a larger number in the target image is increased by P3 intensity values; When operation instruction 4 is received, the brightness deviation rate LP is extracted to determine whether the brightness deviation rate LP is greater than zero; if yes, the brightness value of each pixel in the target image to be spliced ​​with a larger number is multiplied by α1×(1-LP) to obtain the compensated brightness value of each pixel; if no, the brightness value of each pixel in the target image to be spliced ​​with a larger number is multiplied by α2×(1+|LP|) to obtain the compensated brightness value of each pixel; wherein α1 and α2 are both error adjustment coefficients, and the value range of α1 and α2 is (0,2).

6. The ultra-high pixel image stitching method based on deep neural network according to claim 1, characterized in that: The image stitching model is obtained by training a deep neural network model, including: Extracting a number of image sets from a database and integrating the synthetic images corresponding to each image set into a number of sets of training data and test data; using the training data to train the artificial intelligence model; using the test data to test the trained artificial intelligence model, and adjusting the artificial intelligence model according to the test results; Finally, we obtain an image stitching model whose input is an image set and whose output is a synthetic image.

Images