An AI vision-based wound automatic pairing and color calibration method and system

Abstract

The application relates to the field of artificial intelligence image processing, in particular to a method for realizing automatic color card pairing and color calibration by using AI vision technology, which combines self-defined color card design and AI vision technology, first locates a color card "hui" type mark based on a convolutional neural network, corrects a photo direction, then extracts color features of a color card area, adopts three-channel independent polynomial regression to construct a color mapping relationship, and completes overall color correction, finally introduces an adaptive color correction neural network to perform fine color calibration on a medical wound area, and the method significantly improves the precision, real-time performance and robustness of size estimation and color calibration in a medical image, is suitable for complex light, rotation and other scenes, and has important clinical application value.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence image processing, and more particularly to a wound automatic pairing and color calibration method and system based on AI vision. BACKGROUND

[0002] With the rapid development of computer vision and artificial intelligence technology, image analysis, color correction and real size estimation have shown wide application value in the fields of medical images, wound measurement, skin lesion recording, etc. In medical images or clinical photos, accurate measurement of the real size, color and position of the target has key significance for clinical diagnosis, efficacy evaluation and wound healing tracking. However, the traditional method usually relies on the combination of standard color cards and manual measurement ruler, and analyzes through artificial operation or simple image processing technology. This method has significant shortcomings; manual measurement is inefficient, tedious and time-consuming, and is difficult to meet the needs of efficient batch processing, and is easily affected by subjective factors, resulting in large errors; in addition, traditional image processing methods such as template matching and edge detection are sensitive to image rotation, tilt, scale change and changes in lighting conditions, and are prone to problems such as inaccurate positioning, insufficient precision and poor robustness, which seriously affect the actual application effect.

[0003] In recent years, some studies have tried to use feature point detection-based algorithms (such as SIFT, SURF) or template matching methods to realize color card positioning and size calibration, but these methods still have obvious limitations. First, feature point detection algorithms perform poorly when dealing with single texture, complex lighting or image rotation, with significantly reduced detection accuracy. Second, traditional template matching methods have poor robustness to scale changes and rotation transformations, resulting in low matching accuracy. Third, these methods have high computational complexity, and real-time performance and efficiency are difficult to meet the high requirements of medical image scenarios for computing speed and accuracy.

[0004] Based on the construction method of a wound grading model and a wound self-analysis system disclosed in CN120221096A with publication number CN202510593U, published on June 27, 2025, this patent specially designs the color card by combining the color card with the wound, and performs visual correction based on the designed color card, and applies the corrected color card and wound image to wound analysis. However, this technical solution does not involve the technical means of how to correct the color card and the wound, in addition, this technical solution does not involve intelligent correction and specific color correction combined with medical wounds, which may cause color deviation and affect the accuracy of clinical judgment. Therefore, there is an urgent need for a technical solution that can overcome the above-mentioned defects to improve the accuracy and real-time performance of photo size estimation and color correction in medical image and other actual scenarios. SUMMARY

[0005] In view of the deficiencies of the prior art, the purpose of the present application is to provide an AI vision-based wound automatic pairing and color calibration method and system, aiming to solve the problems of poor positioning accuracy, high computational complexity, insufficient real-time performance and the like of traditional methods in complex scenes such as rotation and inclination, by introducing an AI algorithm and a self-defined color card design, the present application not only significantly improves the positioning success rate and color correction accuracy, but also enhances the adaptability of the system to complex environments, provides more reliable technical support for medical image analysis and clinical diagnosis, and has important application value and technical advantages in improving efficiency, accuracy and robustness.

[0006] To achieve the above object, the present application provides the following technical scheme:

[0007] An AI vision-based wound automatic pairing and color calibration method, comprising the following steps:

[0008] An image acquisition step, obtaining an image including a color card and a wound taken by a vision camera as a to-be-processed image;

[0009] An overall image correction step, determining the deflection angle of the color card according to the marker points in the color card, and deflection the to-be-processed image to obtain a corrected image according to the deflection angle;

[0010] An image color difference processing step, extracting a color region in the corrected image, and mapping each color block color in the color region to each color block color in a preset standard color card to obtain a color difference, and performing color correction on the corrected image according to the color difference to obtain a to-be-analyzed image;

[0011] A wound area color correction step, extracting a wound area in the to-be-analyzed image, introducing an adaptive color correction neural network to construct a medical wound-specific color correction mapping function to finally calibrate the pixels of the wound area.

[0012] Further, the wound area color correction step includes a network construction strategy, and the network construction strategy includes a double-channel feature input step, a feature extraction step and a channel processing step;

[0013] The double-channel feature input step constructs an ACCNet network with double-channel input, the first channel inputs the RGB features of 22 color blocks of the standard color card, and the second channel inputs the color distribution histogram features of the wound area;

[0014] The feature extraction step processes the double-channel features through a feature extraction convolution layer respectively to obtain feature maps;

[0015] The channel processing step weights the feature map through a channel attention mechanism to highlight wound-specific color channel information, and locates a color focus area through a spatial attention mechanism to strengthen the attention to the color of the wound tissue.

[0016] Further, the wound area color correction step includes a mapping function generation strategy, which includes a correction parameter output step and a loss function setting step.

[0017] The correction parameter output step outputs a color correction parameter vector of the wound area through the ACCNet network, and the correction parameter vector includes a correction weight of each color channel.

[0018] The loss function setting step sets the loss function of the ACCNet network as:

[0019] ;

[0020] Wherein, is the color card color correction error, is the clinical real loss, is a balance coefficient, and the includes a typical value of healthy tissue, a typical value of granulation tissue, a typical value of cellulose tissue, and a typical value of necrotic tissue.

[0021] Further, the image color difference processing step includes a color mapping strategy, which includes constructing a color mapping relationship between each color block color and each color block color in the preset standard color card through a three-channel independent polynomial regression method to obtain a color difference.

[0022] Further, the three-channel independent polynomial regression method includes a color channel normalization step, a polynomial model construction step, and a color mapping function generation step.

[0023] The color channel normalization step normalizes the RGB channel values of the color card in the correction image and the RGB channel values of the preset standard color card, respectively.

[0024] The polynomial model construction step constructs an n-order polynomial mapping model for each channel, respectively.

[0025] The color mapping function generation step solves the polynomial parameters of each channel by normal equation method with the objective of minimizing the loss function, and constructs a color mapping function according to the polynomial parameters of each channel.

[0026] Further, the image color difference processing step further includes a color region extraction strategy, which includes a color region cutting step and a grid center point positioning step.

[0027] In the color area cutting step, the coordinates of three marking points of the color card are identified in the corrected image, the coordinates of the four top corners of the color card area are calculated, and the color area of the color card is cut through the edge contour algorithm according to the coordinates of the four top corners;

[0028] In the grid center point positioning step, the color area is divided into grids, the size of each grid is calculated, and the pixel position of the center point of each grid is determined.

[0029] Furthermore, the color area extraction strategy further includes a marking point grid exclusion step and a color matching relationship establishment step;

[0030] In the marking point grid exclusion step, the grids where the three "hui" - shaped marking points in the color area fall are excluded;

[0031] In the color matching relationship establishment step, the RGB three - channel color values of the center point pixels of the remaining grids are extracted, the standard color values of the corresponding grids in the preset standard color card are obtained, and a one - to - one matching relationship between the two is established.

[0032] Furthermore, in the overall image correction step, there is a correction strategy, and the correction strategy includes a marking point recognition step, a center point coordinate detection step, and a rotation transformation step;

[0033] In the marking point recognition step, at least three marking points in the color card are recognized. The marking points are black - and - white "hui" - shaped patterns, which are respectively located at the upper - left, upper - right, and lower - left positions of the color card area;

[0034] In the center point coordinate detection step, the coordinate center points of the "hui" - shaped patterns are detected through a convolutional neural network to obtain three center point coordinates;

[0035] In the rotation transformation step, the deflection angle of the color card is calculated based on the line segment formed by the upper - left and upper - right marking points, and the待处理图像 (to - be - processed image) is transformed through the rotation transformation formula based on the deflection angle to obtain a corrected image.

[0036] Furthermore, in the wound area color correction step, the corrected color value of each pixel in the wound area of the待分析图像 (to - be - analyzed image) after image color difference processing is obtained, each pixel is calibrated with the calibration parameter vector, and then the calibrated color value is restored to the range of 0 - 255 to obtain the finally calibrated wound area image.

[0037] An AI - based vision - assisted wound automatic pairing and color calibration system includes:

[0038] An image acquisition module that acquires an image including a color card and a wound taken by a vision camera as a to - be - processed image;

[0039] An overall image correction module is configured to determine a deflection angle of the color card according to the marked points in the color card, and to obtain a corrected image by deflecting the to-be-processed image according to the deflection angle.

[0040] An image color difference processing module is configured to extract a color region from the corrected image, to map each color block in the color region to each color block in a preset standard color card to obtain a color difference, and to perform color correction on the corrected image according to the color difference to obtain a to-be-analyzed image.

[0041] A wound area color correction module is configured to extract a wound area from the to-be-analyzed image, to introduce an adaptive color correction neural network to construct a medical wound-specific color correction mapping function, and to finally calibrate pixels in the wound area.

[0042] The present application has the following advantages: 1. By combining self-defined color card design with AI vision technology, the size estimation and color calibration technology in medical images and clinical photos are significantly improved. In particular, the method preliminarily calibrates the overall color difference through traditional color card pixel mapping, and then the AI model performs fine color calibration on the specific area of the medical wound, realizing multi-level accurate correction from the whole to the local. Among them, the typical color benchmarks of medical wound tissue (such as granulation, necrosis, and yellow tissue) are introduced as the training guide of the AI model, making the color calibration more suitable for the needs of clinical medical diagnosis. Further, the method effectively deals with complex factors such as light, angle, and tissue heterogeneity under different shooting conditions by adaptively learning the characteristics of the medical wound, greatly enhancing the system robustness. In particular, the AI model can adaptively adjust the correction parameters according to the medical color characteristics of the wound image itself, avoiding the defects of traditional single mapping method with fixed parameters and difficulty in real-time adjustment. The combination of traditional pixel correction and AI medical benchmark constraints makes the color of the corrected image closer to the real wound color expected by clinical experts, which is beneficial to improving the accuracy of diagnosis and treatment decisions.

[0043] 2. Under the RGB three-channel, the root mean square error (RMSE) between the color blocks of the photo color card and the corresponding blocks of the standard color card is reduced to an average of 2.6 pixel values or less by using a polynomial regression combined with an AI color correction method, which is more than 70% lower than the traditional manual or linear correction method (the error is usually in the range of 10-15 pixel values). The color structural similarity (SSIM) between the corrected photo color card and the standard color card is improved to more than 0.97, which is much higher than the traditional method (generally around 0.85), effectively improving the color authenticity of the photo. By introducing the AI algorithm and the self-defined color card positioning mark, the adaptability of the present application to rotation, inclination, and complex lighting environments is enhanced, and the positioning success rate reaches more than 98%, significantly reducing the risk of color correction failure caused by positioning errors. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 is a step diagram of the calibration method in the present application;

[0045] Figure 2 is a color correction flowchart of the image to be processed in the present application;

[0046] Figure 3 is a color card structure diagram in the present application;

[0047] Figure 4 is a cutting contour sample diagram in the present application. DETAILED DESCRIPTION

[0048] The present application will be further described in detail below in conjunction with the accompanying drawings and examples. Identical parts are denoted by identical reference numerals in the description below. It should be noted that the words "front", "back", "left", "right", "up" and "down" used in the following description refer to the directions in the drawings, and the words "bottom" and "top", "inner" and "outer" refer to the directions towards or away from the geometric center of a particular part.

[0049] The present application provides a method and system for realizing automatic color card pairing and color calibration using AI vision technology. Through technical means such as self-defined color card design, AI vision technology positioning and color feature extraction, two-stage color correction combining polynomial regression and adaptive neural network, the precision and real-time performance of size estimation and color calibration in medical images and clinical photos are significantly improved. The specific embodiments of the present application will be described in detail below in conjunction with the specific drawings in the description of the accompanying drawings.

[0050] As Figure 1 is a step diagram of the calibration method, the entire method is divided into multiple key steps, including self-defined color card design, image acquisition, color card positioning and correction, grid construction, color mapping relationship establishment, overall color correction and fine calibration of wound area, each step has been fully verified in actual application scenarios and has a clear technical implementation path.

[0051] First, based on the disclosure number CN120221096AThe application discloses a construction method of a wound grading model and a wound self-analysis system, and designs a self-defined color card which comprises a color card area and a handheld area. The color card area adopts a 5*5 square layout and comprises 25 blocks, wherein three blocks are black-and-white "Hui" type positioning marks which are respectively located at upper left, upper right and lower left corner positions, and the remaining 22 blocks are standard color blocks which are used for color feature extraction. The handheld area is convenient for user handheld operation and avoids pollution or shielding of the color card area. The uniqueness of the color card design lies in that the "Hui" type positioning marks are used for realizing accurate positioning and angle correction of the color card in an image, and the 22 standard color blocks are used for providing accurate reference for subsequent color correction. In actual application, a user is required to shoot a photo containing a target wound and the self-defined color card, and ensure that the photo resolution is wherein represents a width (pixel), represents a height (pixel), and the color card and the wound are required to be in the same plane and have no obvious shielding during shooting, so as to improve the accuracy of subsequent processing.

[0052] Then, an image acquisition and preliminary processing stage is entered, an image including the color card and the wound shot by a visual camera is acquired as a to-be-processed image. In order to realize accurate positioning and direction correction of the color card, a convolutional neural network model is used to detect the positions of three "Hui" type positioning marks in the color card area, as shown in the drawing. Figure 3 The convolutional neural network model can effectively identify the center point coordinates of the "Hui" type pattern under complex lighting conditions by training a large amount of labeled data.

[0053]

[0054] According to the three points, the rotation angle of the color card in the photo is determined . Assuming that a line segment formed by the "Hui" pattern 1 and 2 is a reference line, the angle calculation formula is: . According to the calculated rotation angle , the to-be-processed image is rotated and transformed to obtain a corrected image .

[0055]

[0056] The core of this step lies in improving the robustness and accuracy of color card positioning through an AI algorithm, and the positioning success rate is still high under complex lighting conditions.

[0057] After the color card positioning is completed, an accurate cutting and grid construction stage of the color card area is entered. According to the rotation angle and the "Hui" coordinates, four top corner coordinates of the color card are calculated . The area of the color card is determined according to the coordinates.​​​

[0058] ,

[0059] At the same time, the color card area is divided into a 5x5 grid, and the size of each square is: Then the center pixel position of the square in the row and the column is:

[0060] ,

[0061] As Figure 4 shown, the cutting contour is obtained by an edge contour algorithm, which ensures accurate cutting of the color card area, calculates the size of each grid, and determines the center pixel position of each grid.

[0062] In particular, in the image color card area , the upper left, upper right, and lower left three "hui" type positioning marks are excluded, and the RGB three-channel color values of the remaining 22 standard color square center points are extracted, wherein the center pixel coordinates of the remaining 22 standard color squares are:

[0063] { },

[0064] The RGB three-channel color values are extracted from based on the center pixel coordinates of each square, denoted as:

[0065] ,

[0066] The image color card color vector is formed, and the standard color card image is divided in the same way as described above. Specifically, first define the standard grid in the standard color card image, and the grid size is also 5x5. The center point coordinate calculation formula is also:

[0067] ,

[0068] wherein is the top-left corner coordinate of the standard color card area, , are the width and height of each grid in the standard color card, respectively.

[0069] After excluding the upper left, upper right, and lower left three positioning squares from the standard color card area, the center point color pixel values of the remaining 22 color squares in the standard color card image are extracted:

[0070] ,

[0071] Obtain the standard color card color vector: ,

[0072] Then match the image color card color vector with the standard color card color vector one by one to establish a color mapping relationship, a position matching relationship, and the photo color card Color block position Color block position ;

[0073] Finally, a strict one-to-one color matching is obtained: In the color feature extraction process, the RGB value is normalized to the interval [0, 1] to eliminate the brightness difference that may exist between different devices.

[0074] Subsequently, enter the color mapping correction relationship construction stage, and the present application adopts a three-channel independent polynomial regression method to establish a color mapping relationship and generate a color correction function. Normalize each color channel value to the interval [0, 1], and the photo color card RGB channel normalized representation is: ,

[0075] The standard color card RGB channel normalized representation is: ,

[0076] The normalized color set is denoted as:

[0077] Photo color: ,

[0078] Standard color: ,

[0079] Define the color mapping model as an n-order polynomial form, for example, the mapping model of the R channel is:

[0080] ,

[0081] Where =[ ] is the parameter set to be solved, and similarly, the G and B channel mapping functions are defined as:

[0082] ,

[0083] ,

[0084] Solve the polynomial regression model parameters by least squares method, define the optimization target as , and the specific loss function is:

[0085] ,

[0086] ,

[0087] ,

[0088] Minimizing the loss function above respectively, the parameters are solved explicitly by normal equation method, taking the R channel as an example, the input vector and the parameter vector are defined as:

[0089] , ,

[0090] Wherein, the parameter solution is: Similarly, the parameters and are solved;

[0091] Finally, the color mapping functions of the three channels are obtained ,

[0092] ,

[0093] For subsequent color correction, the color value of any pixel point in the photo is normalized, the color correction function is applied, and then the corrected color value is restored to the range of 0~255. This step calibrates the RGB value of each pixel point one by one, ensuring the consistency and authenticity of the overall image color.

[0094] Specifically, for the original color value of any pixel point in the photo , the following steps are corrected:

[0095] First, the original color value is normalized: ,

[0096] Then, the color correction function is applied: ,

[0097] Next, the corrected color value is restored to the range of 0~255:

[0098] ,

[0099] Finally, the above steps are performed on all pixels in the photo, and the corrected image with accurate color correction is obtained .

[0100] After the overall color correction is completed, an adaptive color correction neural network (ACCNet) is further introduced to fine-tune the wound picture that has completed the color correction based on the color card, the network adopts a double-channel feature input structure: the first channel inputs the standard color card feature (color reference), including the RGB features of the 22 color blocks of the color card; the second channel inputs the wound feature, and the color histogram feature of the wound picture is used, and the color distribution histogram is: , , wherein represents the RGB color value of the i-th pixel of the wound picture, and N is the effective pixel number of the picture; the medical tissue-specific feature reflecting the wound color, the two channel features are fused through a feature extraction convolution layer and an attention module: the channel attention module is used to highlight the wound-specific color channel information, and the attention weight of the wound feature is automatically improved; the spatial attention module is used to accurately locate the wound color space area which should be focused on by the color correction model; the ACCNet outputs an adaptive color correction parameter vector, including the correction weight of each color channel (RGB), and introduces the color authenticity constraint of the medical wound image through the loss function design.

[0101] The wound area color correction includes a mapping function generation strategy, specifically, the ACCNet network outputs an adaptive color correction parameter vector , including the correction weight of each color channel (RGB): ;

[0102] The loss function design of the ACCNet introduces the color authenticity constraint of the medical wound image:

[0103] ,

[0104] , wherein is the color card color correction error (the error between the wound picture and the standard color card), is the clinical medical wound color correction realness loss, and the wound real color reference (such as red tissue, yellow tissue and black necrotic tissue and other typical medical colors) confirmed by the pre-labeled clinical medical experts is used to guide the network correction.

[0105] In the present application, , considering that the wound color correction faces various wounds, the following are respectively set:

[0106] the typical value of healthy tissue =(210, 165, 155),

[0107] the typical value of granulation tissue =(170, 55, 55),

[0108] the typical value of cellulose tissue = (195, 160, 55),

[0109] Necrotic tissue typical values = (45, 35, 30),

[0110] is the balance coefficient, and in the present application, the value is 0.48.

[0111] Accurate color correction of wound picture implementation: AI-based accurate calibration is performed on the original RGB color values (R, G, B) of each pixel of the medical wound image,

[0112] ,

[0113] Finally, the wound area image after final calibration is obtained .

[0114] In actual application, the present method can be embedded in various hardware device end side deployment, such as portable medical image analyzer, intelligent terminal, etc. Taking an embedded device as an example, the device end needs to be configured with a high-performance processor (such as GPU or TPU) to support the fast inference of the convolutional neural network, and at the same time needs to have sufficient storage space to save the pre-trained model and intermediate data. The training set scale is recommended to be more than 2000 high-quality medical images, covering various shooting conditions (such as different light, angle, resolution, etc.). In terms of hyperparameter setting, the learning rate is preferably in the range of 0.001-0.01, the batch size is preferably in the range of 32-128, and the network layer number can be adjusted according to the specific task requirement. In the actual running process, the system first loads the pre-trained model and initializes the parameters, and then receives the user uploaded image to be processed. Through the execution of the above steps, the system can complete the whole process processing from image acquisition to final color correction within milliseconds, meeting the real-time requirement.

[0115] Based on the wound color calibration method, a set of calibration system is designed, including:

[0116] An image acquisition module acquires an image including a color card and a wound obtained by a visual camera as a to-be-processed image;

[0117] An overall image correction module determines the deflection angle of the color card according to the marker points in the color card, and deflection is performed on the to-be-processed image according to the deflection angle to obtain a corrected image;

[0118] An image color difference processing module extracts a color region in the corrected image, and maps each color block color in the color region to each color block color in a preset standard color card to obtain a color difference, and performs color correction on the corrected image according to the color difference to obtain an analyzed image;

[0119] The wound area color correction module extracts a wound area in the image to be analyzed, introduces an adaptive color correction neural network to construct a medical wound-specific color correction mapping function, and finally calibrates the pixels of the wound area.

[0120] The present application preliminarily calibrates the overall color difference through traditional color card pixel mapping, and then the AI model performs fine color calibration on the specific area of the medical wound, realizes multi-level accurate correction from the whole to the local, introduces typical color benchmarks of medical wound tissue (such as granulation, necrosis, and yellow tissue) as the training guide of the AI model, makes the color calibration more suitable for the needs of clinical medical diagnosis, effectively deals with complex factors such as light, angle, and tissue heterogeneity under different shooting conditions through adaptive learning of medical wound features, and greatly enhances the system robustness.

[0121] The AI model can adaptively adjust the correction parameters according to the medical color features of the wound image itself, avoiding the defects of fixed parameters and difficulty in real-time adjustment of traditional single mapping method. The combination of traditional pixel correction and AI medical benchmark constraint makes the color of the corrected image closer to the real wound color expected by clinical experts, which is beneficial to improve the accuracy of diagnosis and treatment decision.

[0122] The above is only the preferred embodiment of the present application, and the protection scope of the present application is not limited to the above-mentioned embodiments. Any technical solution falling within the scope of the present application shall be considered as falling within the protection scope of the present application. It should be noted that for ordinary skilled persons in the art, some improvements and refinements without departing from the principles of the present application shall also be considered as falling within the protection scope of the present application.

Claims

1. An AI vision-based wound automatic pairing and color calibration method, characterized in that: The method comprises the following steps: An image acquisition step, obtaining an image including a color card and a wound surface taken by a visual camera as a to-be-processed image; An overall image correction step, determining a deflection angle of the color card according to a mark point in the color card, and deflection of the to-be-processed image according to the deflection angle to obtain a corrected image; An image color difference processing step, extracting a color region in the corrected image, and mapping each color block in the color region to each color block in a preset standard color card to obtain a color difference, and performing color correction on the corrected image according to the color difference to obtain a to-be-analyzed image; A wound surface region color correction step, extracting a wound surface region in the to-be-analyzed image, introducing an adaptive color correction neural network to construct a medical wound surface specific color correction mapping function to finally calibrate pixels of the wound surface region; the wound surface region color correction step comprises a network construction strategy, and the network construction strategy comprises a double-channel feature input step, a feature extraction step and a channel processing step; The double-channel feature input step constructs an ACCNet network with double-channel input, the first channel inputs RGB features of 22 color blocks of the standard color card, and the second channel inputs color distribution histogram features of the wound surface region; The feature extraction step processes the double-channel features through a feature extraction convolution layer respectively to obtain feature maps; The channel processing step weights the feature maps through a channel attention mechanism to highlight wound surface specific color channel information, and positions a color key region of the wound surface region through a spatial attention mechanism to strengthen attention to the color of the wound surface tissue; The wound surface region color correction step comprises a mapping function generation strategy, and the mapping function generation strategy comprises a correction parameter output step; the correction parameter output step outputs a color correction parameter vector of the wound surface region through the ACCNet network, and the correction parameter vector comprises correction weights of each color channel.

2. The AI vision-based wound surface automatic pairing and color calibration method according to claim 1, characterized in that: The wound surface region color correction step comprises a loss function setting step; the loss function setting step sets the loss function of the ACCNet network as: Loss = Loss MSE (C acc， C std )+λLoss MED (C acc +C clinical ); Wherein, Loss MSE is the color card color correction error, Loss MED is the clinical medical wound color correction real loss, λ is a balance coefficient, and the C clinical includes a healthy tissue typical value, a granulation tissue typical value, a fibrous tissue typical value, and a necrotic tissue typical value.

3. The AI vision-based wound automatic pairing and color calibration method according to claim 1 or 2, characterized in that: The image color difference processing step comprises a color mapping strategy, and the color mapping strategy comprises constructing a color mapping relationship through a three-channel independent polynomial regression method to obtain a color difference between each color block and each color block in a preset standard color card.

4. The AI vision-based wound surface automatic pairing and color calibration method according to claim 3, characterized in that: The three-channel independent polynomial regression method comprises a color channel normalization step, a polynomial model construction step and a color mapping function generation step; The color channel normalization step respectively normalizes RGB channel values of the color card in the corrected image and RGB channel values of the preset standard color card; The polynomial model construction step respectively constructs an n-order polynomial mapping model for each channel; The color mapping function generation step solves polynomial parameters of each channel through a normal equation method with the objective of minimizing a loss function, and constructs a color mapping function according to the polynomial parameters of each channel.

5. The AI vision-based wound automatic pairing and color calibration method according to claim 4, characterized in that: The image color difference processing step further comprises a color region extraction strategy, and the color region extraction strategy comprises a color region cutting step and a grid center point positioning step; In the color area cutting step, the coordinates of three marker points of the color card are identified in the corrected image, the coordinates of the four top corners of the color card area are calculated, and the color area of the color card is cut by the edge contour algorithm according to the four top corner coordinates; In the grid center point positioning step, the color area is divided into grids, the size of each grid is calculated, and the pixel position of the center point of each grid is determined.

6. The AI vision-based wound surface automatic pairing and color calibration method according to claim 5, characterized in that: The color area extraction strategy further includes a marker point grid exclusion step and a color matching relationship establishment step; In the marker point grid exclusion step, the grids into which the three "回”-shaped marker points in the color area fall are excluded; In the color matching relationship establishment step, the RGB three-channel color values of the center point pixels of the remaining grids are extracted, the standard color values of the corresponding grids in the preset standard color card are obtained, and a one-to-one matching relationship between the two is established.

7. The AI vision-based wound automatic pairing and color calibration method according to claim 6, characterized in that: In the overall image correction step, there is a correction strategy, and the correction strategy includes a marker point recognition step, a center point coordinate detection step, and a rotation transformation step; In the marker point recognition step, at least three marker points in the color card are recognized. The marker points are black-and-white "回”-shaped patterns, which are respectively located at the upper left, upper right, and lower left positions of the color card area; In the center point coordinate detection step, the coordinate center points of the "回”-shaped patterns are detected by a convolutional neural network to obtain three center point coordinates; In the rotation transformation step, the deflection angle of the color card is calculated based on the line segment formed by the upper left and upper right marker points, and the to-be-processed image is transformed by the rotation transformation formula based on the deflection angle to obtain a corrected image.

8. The AI vision-based wound automatic pairing and color calibration method according to claim 7, characterized in that: In the wound area color correction step, the corrected color value of each pixel in the wound area of the to-be-analyzed image after image color difference processing is obtained, each pixel is calibrated with the calibration parameter vector, and then the calibrated color value is restored to the range of 0-255 to obtain the finally calibrated wound area image.

9. An AI vision-based wound automatic pairing and color calibration system, characterized in that: including: An image acquisition module that obtains an image including a color card and a wound taken by a vision camera as a to-be-processed image; An overall image correction module that determines the deflection angle of the color card according to the marker points in the color card, and deflects the to-be-processed image according to the deflection angle to obtain a corrected image; An image color difference processing module that extracts the color area in the corrected image, maps the color of each color block in the color area to the color of each color block in the preset standard color card to obtain a color difference, and performs color correction on the corrected image according to the color difference to obtain a to-be-analyzed image; A wound area color correction module that extracts the wound area in the to-be-analyzed image, introduces an adaptive color correction neural network to construct a medical wound-specific color correction mapping function to finally calibrate the pixels in the wound area; in the wound area color correction module, there is a network construction strategy, and the network construction strategy includes a dual-channel feature input step, a feature extraction step, and a channel processing step; In the dual-channel feature input step, an ACCNet network with dual-channel input is constructed. The first channel inputs the RGB features of 22 color blocks of the standard color card, and the second channel inputs the color distribution histogram features of the wound area; The feature extraction step is that the double-channel features are respectively processed through a feature extraction convolutional layer to obtain feature maps; The channel processing step is that the feature maps are weighted through a channel attention mechanism to highlight wound-specific color channel information, and the wound area is positioned in a color key area through a spatial attention mechanism to strengthen the attention to the color of the wound tissue. The wound area color correction module includes a mapping function generation strategy, and the mapping function generation strategy includes a correction parameter output step. The correction parameter output step outputs a color correction parameter vector of the wound area through an ACCNet network, and the correction parameter vector includes a correction weight of each color channel.

Images