A contactless automatic analysis method for red eye degree

Abstract

The present invention discloses a contactless automatic analysis method for the degree of eye redness, comprising the following steps: photographing a human eye from the front of the face, performing eye region detection and image preprocessing on the photographed image, performing scaling and preprocessing, and then using a U-Net++ model to detect the sclera region to obtain a binary image, performing an AND operation on the binary image and the scaled image to obtain a sclera region color map; performing image smoothing and adaptive histogram equalization to enhance image contrast, filtering the image enhancement image using a B-COSFIRE filter, and converting the image enhancement image into a LAB color model to obtain an image mask; finally, performing an AND operation on the filtered image and the image mask to obtain a bloodshot binary map; and calculating the number of non-zero pixels in the sclera region color map and the bloodshot binary map, respectively, and then performing a ratio calculation to obtain the eye redness ratio. The present invention provides a contactless automatic analysis method for the degree of eye redness, which has high adaptability and accuracy in automatic determination and can achieve intelligent auxiliary diagnosis of the degree of eye redness in patients.

Description

Technical Field

[0001] The invention relates to a contactless automatic analysis method for eye redness degree, belonging to the technical field of image processing and medical auxiliary diagnosis. Background Art

[0002] Eye redness analysis can determine the severity of ocular surface inflammation and, as a key indicator for dry eye detection, plays a crucial role in ophthalmic diagnosis. Traditionally, eye redness analysis relies on physicians' experience, which can lead to significant errors. Furthermore, despite the emergence of digital image processing methods, these methods offer limited adaptability and accuracy for automated assessment, and are unable to provide intelligent, assisted diagnosis of a patient's eye redness severity. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to overcome the defects of the prior art and provide a contactless automatic analysis method for the degree of redness of the eyes, which has high adaptability and accuracy in automatic determination and can realize intelligent auxiliary diagnosis of the degree of redness of the patients.

[0004] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0005] A non-contact automatic analysis method for red eye degree comprises the following steps:

[0006] Use a visible light imaging device to shoot human eyes from the front of the face, perform eye region detection and image preprocessing on the captured image, and obtain a preprocessed image;

[0007] The preprocessed image is scaled to obtain a scaled image, the scaled image is preprocessed and then the sclera area is detected using the U-Net++ model to obtain a binary image of the sclera area of ​​the image to be detected, and the binary image is ANDed with the scaled image to obtain a color image of the sclera area;

[0008] The sclera region color map is subjected to image smoothing and adaptive histogram equalization to enhance the image contrast, obtaining an image enhancement map. The image enhancement map is then filtered using a B-COSFIRE filter to obtain a filtered image. The image enhancement map is then converted into a LAB color model to obtain an image mask. Finally, the filtered image and the image mask are ANDed to obtain a bloodshot binary map.

[0009] The number of non-zero pixels in the sclera area color image and the bloodshot binary image were calculated respectively, and then the ratio of the number of non-zero pixels in the bloodshot binary image to the number of non-zero pixels in the sclera area color image was calculated to obtain the red eye ratio.

[0010] The human eye area detection comprises the following steps:

[0011] When detecting the captured image, the trained classifier is used to search each area in the captured image starting from the upper left corner, and the similarity criterion is used to determine whether it is a human eye;

[0012] If the result is a human eye, the eye area is selected and the size of the eye area is determined to be at least 256×256. If so, the process proceeds to the next step. If not, the system prompts that the human eye image is too small and requires re-shooting.

[0013] If the judgment result is non-human eye, it will prompt that the input image is wrong and ask you to retake the photo;

[0014] Among them, the human eye classifier training process is: using deep learning methods, first collect a large number of human eye and non-human eye samples from the Internet for pre-training, and obtain deep learning model parameters to build a human eye classifier.

[0015] The pre-processing of captured images includes the following steps:

[0016] Based on the eye area selection results, the length and width of the human eye area are expanded by 1.2 times the original ratio, so that the final selected eye area contains complete eye information from the corner of the eye. Based on the expanded length and the parallel center line as the baseline, the image is cropped with an aspect ratio of 5:4 to ensure that all eye images meet the size requirements and maximize the effective area.

[0017] The scaled image preprocessing includes the following steps: using the color threshold method, if the overall image color is judged to be reddish, the R channel of the RGB three-channel image is extracted and the sclera segmentation operation is continued; otherwise, after using the histogram equalization method to enhance the image, the sclera segmentation operation is still performed using the RGB three-channel color image.

[0018] The U-Net++ model is an improved U-Net++ model. The input image is a scaled image, the output image is a single-channel binary image, and the input and output image shapes are both 512×512. A normalization operation is added between the convolution operation and the ReLU activation function operation in each layer of U-Net++. A dropout operation is added between the upper and lower layers, and an attention mechanism is added afterwards, followed by a pooling operation.

[0019] The specific operation of the binary image and the zoom image is as follows: all pixel points with a pixel value of 0 in the binary image are set to 0 in the corresponding pixel value in the zoom image, and only the sclera area in the image is retained to obtain the sclera area map.

[0020] The specific calculation method of red eye ratio is as follows:

[0021]

[0022] Among them, Degree represents the proportion of jealousy, B pixelsIndicates the number of non-zero pixels in the bloodshot binary image, S pixels Indicates the number of non-zero pixels in the sclera area image.

[0023] Multiply the calculated result of the redness ratio by 100 and round it down so that the final result is in the range of [0,100]. Finally, use the score grading system to obtain the corresponding redness level.

[0024] A non-contact automatic analysis device for red eye degree, comprising:

[0025] The eye image acquisition and processing module is used to use a visible light imaging device to capture the human eye from the front of the face, perform eye region detection and image preprocessing on the captured image, and obtain a preprocessed image;

[0026] The sclera region detection module is used to scale the preprocessed image to obtain a scaled image, preprocess the scaled image and perform sclera region detection on the scaled image using the U-Net++ model to obtain a binary image of the sclera region of the image to be detected, and perform an AND operation on the binary image and the scaled image to obtain a color image of the sclera region;

[0027] The red eye area extraction module is used to extract the sclera area color map by performing image smoothing and adaptive histogram equalization to enhance the image contrast, obtaining an image enhancement image. The image enhancement image is then filtered using a B-COSFIRE filter to obtain a filtered image. The image enhancement image is converted into a LAB color model to obtain an image mask. Finally, the filtered image and the image mask are ANDed to obtain a bloodshot binary image.

[0028] The red eye ratio calculation module is used to calculate the number of non-zero pixels in the sclera area color image and the bloodshot binary image respectively, and then calculate the ratio of the number of non-zero pixels in the bloodshot binary image to the number of non-zero pixels in the sclera area color image to obtain the red eye ratio.

[0029] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the method for automatically analyzing the degree of contact eye redness.

[0030] The beneficial effects of the present invention are as follows: the present invention provides a contactless automatic analysis method for the degree of redness of the eyes, which uses a visible light imaging device to aim at the front of the face to shoot the human eyes, and can be achieved by using the camera function of an ordinary smartphone or a camera to take pictures. The eye image is obtained contactlessly, without auxiliary light source, simple structure, and no discomfort. It can avoid the error caused by the redness presentation area and congestion around the eyes, accurately position and area of ​​the redness, provide accurate data, and automatically determine the adaptability and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1The present invention is a workflow diagram of a contactless automatic analysis method for red eye degree.

[0032] Figure 2 This is a flow chart of eye image acquisition and processing in the present invention;

[0033] Figure 3 This is a flow chart of sclera area detection in the present invention;

[0034] Figure 4 This is a flowchart for extracting red eye areas in the present invention;

[0035] Figure 5 This is a flow chart for calculating the red eye ratio and judging the red eye degree in the present invention. DETAILED DESCRIPTION

[0036] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention. Figure 1 As shown, the present invention discloses a contactless automatic analysis method of eye redness, comprising the following steps:

[0037] Step 1: Eye image acquisition and processing.

[0038] Use a visible light imaging device to photograph the human eye from the front of the face. The resulting image must include a complete single eye. Due to differences in imaging device resolution, the video memory occupied by the sclera region detection neural network training, the sclera region detection results, and the algorithm execution efficiency will be affected. Therefore, eye region detection and image preprocessing must be performed first. Use a visible light imaging device to photograph the human eye from the front of the face. This can be achieved using the camera function of an ordinary smartphone or a camera. The only requirement is that the image size of the eye region must be at least 256×256. No additional equipment or specialized ophthalmic instruments are required.

[0039] Eye region detection uses pre-training on a large number of collected human eye and non-human eye samples to obtain model parameters for constructing an eye classifier. When detecting captured images, the trained classifier is used to search each region of the image, starting from the upper left corner, using a similarity criterion to determine whether an image is a human eye. Region-by-region search involves dividing the input image into several small rectangular regions, each of which is then individually tested for eye detection. The similarity criterion uses the eye classifier model to calculate similarity features for objects detected within each region during the region-by-region search. The result indicates the degree of similarity between the detected object and the human eye.

[0040] If the judgment result is a human eye, it is necessary to select the range of the eye area. The eye area range must be no less than 256×256, and then proceed to the next step of image preprocessing. Otherwise, it will prompt that the human eye image is too small and require reshooting. If the judgment result is not a human eye, it will prompt that the input image is incorrect and require reshooting. Image preprocessing is to expand the length and width of the human eye area by 1.2 times the original ratio based on the eye area selection result, so that the final selected eye area contains complete eye information such as the corner of the eye. Based on the expanded length, the parallel center line is used as the baseline and the aspect ratio of 5:4 is used for cropping, so that all eye images meet the size requirements and the effective area is maximized.

[0041] The eye classifier is a deep learning model pre-trained on a large number of human eye and non-eye image samples. It distinguishes between human eyes and non-eyes in input images. It performs eye region selection on images detected as human eyes. For selected images that meet the eye region requirements, the algorithm proceeds to image preprocessing, including sclera region detection, red eye region extraction, and red eye ratio calculation and analysis. For images where the eye region selection results do not meet the requirements, or where the input is detected as non-human eyes, the algorithm terminates and displays a corresponding warning message. Pre-training refers to initial training of the model using a large-scale dataset before the target task, allowing the model to learn some common features and representations. Eye region selection uses a bounding box to select the area identified as human eyes based on human visual judgment. The model then returns the image coordinates and aspect ratio of the selected area. The selected area should include complete eye information, such as the inner and outer canthi, while minimizing other information in the input image that is not relevant to the human eye. The eye region range requires that the selected eye region have an aspect ratio of no less than 256×256 to facilitate subsequent steps such as sclera region detection and red eye region extraction. If the selected eye region has an aspect ratio of no less than 256×256, the input eye image meets the requirements and can proceed to subsequent steps. Otherwise, a message will be displayed indicating that the eye image is too small and a retake will be requested.

[0042] Step 2: sclera area detection.

[0043] The actual eye image obtained after cropping the input image is scaled to 512×512. After preprocessing, the improved U-Net++ model is used to detect the sclera area. The model detection result is a binary image of the sclera area in the image to be tested. The binary image is ANDed with the scaled image to obtain a color image of the sclera area.

[0044] Image scaling involves cropping the eye area and preprocessing the input image. Images smaller than 512×512 are then super-resolved, while images larger than 512×512 are down-sampled to reduce the loss of image detail caused by scaling. Super-resolution involves increasing the resolution of a target image whose actual resolution is lower than the required resolution through interpolation, thereby improving image detail and quality. Downsampling involves averaging some pixels in a target image whose actual resolution is higher than the required resolution, thereby reducing the image's spatial resolution.

[0045] Image preprocessing uses a color thresholding method. If the overall image color is judged to be reddish, the R channel of the RGB three-channel image is extracted and the sclera segmentation operation is continued. Otherwise, after image enhancement using histogram equalization, the sclera segmentation operation is still performed on the RGB three-channel color image.

[0046] The color threshold method calculates the average value of the red channel in the image to determine whether the image needs to be extracted from the red channel before sclera segmentation. When the average value of the color channel exceeds the set threshold, it is considered necessary to extract the red channel of the image; otherwise, it is not necessary.

[0047] Histogram equalization calculates the frequency of occurrence of each pixel value in the image to obtain a histogram of the pixel values. After normalizing the histogram, the cumulative distribution function (CDF) of the histogram is calculated to represent the cumulative probability of occurrence of each pixel value. Based on the CDF, the mapping value of each pixel value in the new histogram is calculated, and the original pixel is mapped to the new pixel value. The new pixel value replaces the original pixel value to obtain the equalized image. The cumulative distribution function CDF is expressed using the following formula:

[0048]

[0049] Where x is the pixel value and P(i) is the frequency of pixel value i in the normalized image.

[0050] Histogram normalization is to standardize the frequency values ​​in the histogram so that the frequency values ​​are between 0 and 1. The normalization method used here is to divide each frequency value by the total number of pixels to ensure that the sum of the normalized frequency values ​​is 1. It is expressed using the following formula:

[0051]

[0052] Among them, N(i) is the normalized frequency value, H(i) is the frequency value of pixel value i in the original histogram, and N pixels is the total number of pixels. The normalized frequency value represents the relative probability of pixel value i appearing.

[0053] The RGB channels refer to the red (R), green (G), and blue (B) color channels of the image. Each pixel in the image is composed of the values ​​of the three color channels, representing the intensity or brightness of red, green, and blue respectively. The value of each color channel is usually an integer between 0 and 255. The higher the value, the larger the color component of the channel, and the brighter the single channel.

[0054] The U-Net++ model is a deep learning model for image segmentation tasks. It expands and improves the classic U-Net model. U-Net++ enhances the model's expressiveness and improves image segmentation performance by introducing a recursive network structure, making it widely used in the field of medical image segmentation. The improved U-Net++ model in this paper primarily adjusts the input and output components of the original U-Net++ network, the convolution operations within each layer, and some of the structures between upper and lower layers. In the input and output components, the input image is a preprocessed image, and the output image is a single-channel binary image. Both the input and output images are 512×512 pixels in size. A normalization operation is added between the convolution operation and the ReLU activation function in each U-Net++ layer. A dropout operation is added between the upper and lower layers, followed by an attention mechanism and a pooling operation. Following these improvements, accuracy improved by 2% to 3% on test data.

[0055] A single-channel binary image is a binary image that contains only a single channel (the grayscale channel). Each pixel has only two possible values: black and white. A grayscale value of 0 is usually used to represent a black pixel, and a grayscale value of 255 (or 1) is used to represent a white pixel.

[0056] The ReLU activation function, short for Rectified Linear Unit (RLU), is a commonly used activation function in neural networks. It generally refers to the ramp function in mathematics.

[0057] Dropout is a strategy widely used in deep learning to solve the problem of model overfitting. Dropout solves the co-adaption problem and makes it possible to train wider networks.

[0058] The co-adaptation problem refers to the co-adaptation problem. Some nodes in the network have stronger representational capabilities than others. As the network continues to train, the nodes with stronger representational capabilities are continuously strengthened, while the weaker nodes are gradually weakened until their contribution to the network becomes negligible. In this case, only some nodes in the network are trained, wasting the width and depth of the network, which in turn limits the effectiveness of the model.

[0059] The attention mechanism is a special structure embedded in machine learning models that automatically learns and calculates the contribution of input data to output data. By introducing the attention mechanism, neural networks can automatically learn and selectively focus on important information in the input, improving the model's performance and generalization capabilities.

[0060] The pooling operation imitates the human visual system to reduce the dimensionality of data, mainly to extract the key information of a certain area.

[0061] Model detection uses the trained improved U-Net++ model to detect the input image. The input image needs to be detected by the human eye classifier in the first step. After the human eye image is cropped, the image size is adjusted to 512×512 as the input of the model. The model can accurately detect the sclera area in the input image and provide a binary image of the detection image result.

[0062] The binary image and the scaled image are ANDed, and all pixels with a pixel value of 0 (black part) in the binary image are set to 0 in the corresponding scaled image, retaining only the sclera area in the image to obtain the sclera area map.

[0063] Step 3: Extract the red eye area.

[0064] The color image of the sclera region obtained in the previous step is used as the input image. Image smoothing and adaptive histogram equalization are performed on the input image to enhance image contrast. The B-COSFIRE filter is used to extract bloodshot from the processed image. A mask is obtained from the enhanced color image. The RGB color model of the color image is converted to the LAB color model. A threshold is set based on the L (brightness) component to regenerate an appropriate mask. The mask of the low-brightness area (i.e., the dark gray background area) is set to 0, and the mask of the rest is set to 1. This effectively prevents the intersection of black and color from being identified as bloodshot by the algorithm. The B-COSFIRE filter is used to extract bloodshot from the processed image, and the resulting image is a binary image of bloodshot.

[0065] Image smoothing uses Gaussian filtering to perform convolution operations on the image using a Gaussian kernel function to achieve a smoothing effect, remove high-frequency noise from the image, smooth the image, and maintain the overall details and edges of the image.

[0066] Adaptive histogram equalization divides the original image into uniform blocks or segments it according to a specific algorithm. Histogram equalization is then applied to each block. Brightness correction is performed based on the histogram distribution of adjacent blocks, ultimately recombining all blocks into an enhanced image. The equalized image improves local contrast by preventing excessive noise amplification in relatively uniform areas.

[0067] The RGB color model is a model used to describe colors. It is based on a combination of three basic color channels: red, green, and blue. The value of each color channel is usually between 0 and 255, where 0 represents the minimum brightness and 255 represents the maximum brightness.

[0068] The LAB color model is a model used to describe color, consisting of two channels: brightness (L) and color opposition (A and B). The A channel represents opposition from green to red, and the B channel represents opposition from blue to yellow.

[0069] B-COSFIRE stands for Bar-selective Combination of Shifted Filter Responses. B-COSFIRE achieves directional selectivity by computing the weighted geometric mean of the outputs of a set of Difference of Gaussian filters whose regions of support are linearly aligned. By summing the responses of two rotationally invariant B-COSFIRE filters and applying a threshold, ocular blood vessel segmentation can be achieved.

[0070] Rotation invariance means that when processing an image, the algorithm or method has the same recognition or analysis results for the object at different rotation angles. In other words, no matter how the object is rotated, the algorithm can correctly recognize or process it without being affected by the rotation.

[0071] The B-COSFIRE filter response function is:

[0072]

[0073] In the above formula, σ is the standard deviation of the Gaussian function that determines the response range.

[0074] Step 4: Calculate the percentage of redness and determine the degree of redness.

[0075] Based on the above steps, the number of nonzero pixels in the obtained bloodshot binary image and sclera area image was calculated, and then the ratio was calculated. The ratio of the two was the redness percentage. The calculated redness percentage was multiplied by 100 and rounded down to a value within the range of [0, 100]. Finally, the corresponding redness level was determined using the following scoring system, as shown in Table 1.

[0076]

[0077] Among them, Degree indicates the degree of jealousy, B pixels Indicates the number of non-zero pixels in the bloodshot binary image, S pixels Indicates the number of non-zero pixels in the sclera area image.

[0078] Table 1 Red eye degree level

[0079] Degree 81～100 61～80 41～60 21～40 0～20 grade A B C D E

[0080] The present invention also discloses a non-contact automatic analysis device for red eye degree, comprising:

[0081] The eye image acquisition and processing module is used to use a visible light imaging device to capture the human eye from the front of the face, perform eye region detection and image preprocessing on the captured image, and obtain a preprocessed image;

[0082] The sclera region detection module is used to scale the preprocessed image to obtain a scaled image, preprocess the scaled image and perform sclera region detection on the scaled image using the U-Net++ model to obtain a binary image of the sclera region of the image to be detected, and perform an AND operation on the binary image and the scaled image to obtain a color image of the sclera region;

[0083] The red eye area extraction module is used to extract the sclera area color map by performing image smoothing and adaptive histogram equalization to enhance the image contrast, obtaining an image enhancement image. The image enhancement image is then filtered using a B-COSFIRE filter to obtain a filtered image. The image enhancement image is converted into a LAB color model to obtain an image mask. Finally, the filtered image and the image mask are ANDed to obtain a bloodshot binary image.

[0084] The red eye ratio calculation module is used to calculate the number of non-zero pixels in the sclera area color image and the bloodshot binary image respectively, and then calculate the ratio of the number of non-zero pixels in the bloodshot binary image to the number of non-zero pixels in the sclera area color image to obtain the red eye ratio.

[0085] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the automatic analysis method for the degree of contact eye redness of the present invention.

[0086] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A non-contact automatic analysis method for red eye degree, characterized by: The following steps are involved: Use a visible light imaging device to shoot a human eye from the front of the face, perform human eye region detection and image preprocessing on the captured image, and obtain a preprocessed image; The preprocessed image is scaled to obtain a scaled image, the scaled image is preprocessed and then the sclera area is detected using the U-Net++ model to obtain a binary image of the sclera area of ​​the image to be detected, and the binary image is ANDed with the scaled image to obtain a color image of the sclera area; The sclera region color map is subjected to image smoothing and adaptive histogram equalization to enhance the image contrast, obtaining an image enhancement map. The image enhancement map is then filtered using a B-COSFIRE filter to obtain a filtered image. The image enhancement map is then converted into a LAB color model to obtain an image mask. Finally, the filtered image and the image mask are ANDed to obtain a bloodshot binary map. The number of non-zero pixels in the sclera area color image and the bloodshot binary image were calculated respectively, and then the ratio of the number of non-zero pixels in the bloodshot binary image to the number of non-zero pixels in the sclera area color image was calculated to obtain the red eye ratio.

2. The non-contact automatic analysis method for red eye degree according to claim 1, characterized in that: Visible light imaging devices, including smartphones and cameras, must meet the requirement that the image size of the eye area is no less than 256×256.

3. The non-contact automatic analysis method for red eye degree according to claim 1, characterized in that: The human eye area detection comprises the following steps: When detecting the captured image, the trained classifier is used to search each area in the captured image starting from the upper left corner, and the similarity criterion is used to determine whether it is a human eye; If the result is a human eye, the eye area is selected and the size of the eye area is determined to be at least 256×256. If so, the process proceeds to the next step. If not, the system prompts that the human eye image is too small and requires re-shooting. If the judgment result is non-human eye, it will prompt that the input image is wrong and ask you to retake the photo; Among them, the human eye classifier training process is: using deep learning methods, first collect a large number of human eye and non-human eye samples from the Internet for pre-training, and obtain deep learning model parameters to build a human eye classifier.

4. The non-contact automatic analysis method of eye redness according to claim 3, characterized in that: The pre-processing of captured images includes the following steps: Based on the eye area selection results, the length and width of the human eye area are expanded by 1.2 times the original ratio, so that the final selected eye area contains complete eye information from the corner of the eye. Based on the expanded length and the parallel center line as the baseline, the image is cropped with an aspect ratio of 5:4 to ensure that all eye images meet the size requirements and maximize the effective area.

5. The non-contact automatic analysis method of eye redness according to claim 1, characterized in that: The scaled image preprocessing includes the following steps: using the color threshold method, if the overall image color is judged to be reddish, the R channel of the RGB three-channel image is extracted and the sclera segmentation operation is continued; otherwise, after using the histogram equalization method to enhance the image, the sclera segmentation operation is still performed using the RGB three-channel color image.

6. The non-contact automatic analysis method for red eye degree according to claim 1, characterized in that: The U-Net++ model is an improved U-Net++ model. The input image is a scaled image, the output image is a single-channel binary image, and the input and output image shapes are both 512×512. A normalization operation is added between the convolution operation and the ReLU activation function operation in each layer of U-Net++. A dropout operation is added between the upper and lower layers, and an attention mechanism is added afterwards, followed by a pooling operation.

7. The non-contact automatic analysis method for red eye degree according to claim 1, characterized in that: The specific operation of the binary image and the zoom image is as follows: all pixel points with a pixel value of 0 in the binary image are set to 0 in the corresponding pixel value in the zoom image, and only the sclera area in the image is retained to obtain the sclera area map.

8. The non-contact automatic analysis method for red eye degree according to claim 1, characterized in that: The specific calculation method of red eye ratio is as follows: Among them, Degree represents the proportion of jealousy, B pixels Indicates the number of non-zero pixels in the bloodshot binary image, S pixels Indicates the number of non-zero pixels in the sclera area map.

9. The non-contact automatic analysis method of eye redness according to claim 8, characterized in that: Multiply the calculated result of the redness ratio by 100 and round it down so that the final result is in the range of [0,100]. Finally, use the score grading system to obtain the corresponding redness level.

10. A non-contact automatic analysis device for red eye degree, characterized by: include: The eye image acquisition and processing module is used to use a visible light imaging device to capture the human eye from the front of the face, perform eye region detection and image preprocessing on the captured image, and obtain a preprocessed image; The sclera region detection module is used to scale the preprocessed image to obtain a scaled image, preprocess the scaled image and perform sclera region detection on the scaled image using the U-Net++ model to obtain a binary image of the sclera region of the image to be detected, and perform an AND operation on the binary image and the scaled image to obtain a color image of the sclera region; The red eye area extraction module is used to extract the sclera area color map by performing image smoothing and adaptive histogram equalization to enhance the image contrast, obtaining an image enhancement image. The image enhancement image is then filtered using a B-COSFIRE filter to obtain a filtered image. The image enhancement image is converted into a LAB color model to obtain an image mask. Finally, the filtered image and the image mask are ANDed to obtain a bloodshot binary image. The red eye ratio calculation module is used to calculate the number of non-zero pixels in the sclera area color image and the bloodshot binary image respectively, and then calculate the ratio of the number of non-zero pixels in the bloodshot binary image to the number of non-zero pixels in the sclera area color image to obtain the red eye ratio.

Images