Student physical health dynamic monitoring system based on deep learning

Abstract

The invention discloses a student physical health dynamic monitoring system based on deep learning. Student eye image data passing a quality threshold is used as qualified student eye image data; cutting according to the face key point positioning result to obtain student eye image data of the interested areas of the eyes and the sub-orbital area; generating CLAHE enhanced student eye image data; extracting an original image stream student eye feature map and a CLAHE enhanced stream student eye feature map; obtaining a fused student eye feature map; outputting a student eye fatigue area pixel-level mask map and a student eye fatigue area pixel-level confidence coefficient heat map; and generating a student eye fatigue sub-item index result, and outputting a student individualized eye fatigue index. According to the invention, large-scale, non-contact and fine-grained physical health monitoring for student groups is realized, and the refinement and traceability of health intervention are improved.

Description

Technical Field

[0001] This invention relates to the field of student physical health technology, and in particular to a student physical health dynamic monitoring system based on deep learning. Background Technology

[0002] With the increasing demand for student health management and the widespread use of electronic devices and close-range eye use habits, students' eye health has become an important focus in campus health monitoring.

[0003] Traditional methods for detecting eye fatigue in students often rely on overall facial expressions or single visual features for assessment, making it difficult to distinguish specific signs of fatigue in local eye areas. Furthermore, in practical applications, they mainly output a single overall fatigue score, which cannot support accurate regional identification and personalized intervention.

[0004] Existing visual algorithms are often limited by complex lighting, skin color diversity, and differences in device imaging, resulting in low accuracy in detecting fine-grained signs of fatigue such as conjunctival congestion and infraorbital hyperemia in student eye images. They are also easily affected by reflections, noise, and skin color differences in natural environments.

[0005] Existing methods often employ fixed parameters or simple algorithms in pixel-level segmentation, contrast enhancement, and feature extraction. These methods often suffer from drawbacks such as over-enhancement leading to noise and artifacts being misidentified as fatigued regions, or insufficient enhancement failing to improve the segmentation effect of key regions. Summary of the Invention

[0006] One objective of this invention is to propose a student physical health dynamic monitoring system based on deep learning. This invention enables large-scale, non-contact, and fine-grained physical health monitoring for student groups, improving the precision and traceability of health interventions.

[0007] A student physical health dynamic monitoring system based on deep learning according to an embodiment of the present invention includes: The quality assessment module collects student eye image data and performs quality assessment on the raw student eye image data. Based on occlusion detection, reflection detection, and sharpness detection, it determines whether the image meets the preset quality threshold. Student eye image data that passes the quality threshold is considered qualified student eye image data. The region of interest extraction module performs facial key point localization on qualified student eye image data, and crops the student eye image data of the eye eyes and infraorbital region based on the facial key point localization results; The learnable CLAHE module inputs student eye image data of the region of interest into the learnable CLAHE parameter prediction module, predicts key CLAHE parameters based on the brightness histogram and skin color features of the student eye image data of the region of interest, and performs CLAHE enhancement processing on the student eye image data of the region of interest using the key CLAHE parameters to generate CLAHE enhanced student eye image data. The H2Former feature encoding module is improved by inputting the student eye image data of the region of interest and the CLAHE-enhanced student eye image data into the improved H2Former encoder to extract the student eye feature map of the original image stream and the student eye feature map of the CLAHE-enhanced image stream. The gated feature fusion module performs cross-scale gated fusion processing on the student eye feature map of the original image stream and the student eye feature map of the CLAHE enhanced stream during the multi-scale feature fusion stage. The fusion weight is jointly controlled by the local contrast entropy and the texture consistency index to obtain the fused student eye feature map. The confidence generation module inputs the student's eye feature map into the semantic segmentation subnetwork and outputs a pixel-level mask map of the student's eye fatigue area and a pixel-level confidence heatmap of the student's eye fatigue area. The health early warning module calculates the proportion of conjunctival congestion area and the grayscale deviation of infraorbital dark area in the pixel-level mask image of the student's eye fatigue area, generating the results of the student's eye fatigue sub-indices. It then calculates the difference between the results of the student's eye fatigue sub-indices and the corresponding student eye indicators, outputs the student's individualized eye fatigue index, and determines the student's health early warning level based on the individualized eye fatigue index.

[0008] Optionally, the quality assessment module includes: The system uses cameras to capture raw, real-time images of students' eyes, including their entire faces. Calculate the occlusion rate of the student's eye image data in the region of interest. When the occlusion rate is less than or equal to the occlusion rate threshold, the occlusion detection is considered successful. Calculate the maximum brightness ratio of the student's eye image data in the region of interest. When the maximum brightness ratio is less than or equal to the reflection detection threshold, the reflection detection is considered passed. Calculate the image sharpness index of the student's eye image data in the region of interest. When the image sharpness index is greater than or equal to the image sharpness judgment threshold, the sharpness detection is considered passed. When the occlusion detection, reflection detection, and sharpness detection all meet their corresponding threshold conditions, the original student eye image data is judged as qualified student eye image data.

[0009] Optionally, the region of interest extraction module includes: Input qualified students' eye image data into the facial landmark detection model to obtain a set of facial landmark coordinates; The left eye key point index set and the right eye key point index set are determined based on the facial key point coordinate set, and the left eye bounding box and the right eye bounding box are calculated based on the left eye key point index set and the right eye key point index set, respectively. Calculate the infraorbital region bounding box based on the left and right eye bounding boxes; For qualified student eye image data, cropping operations are performed sequentially based on the left eye bounding box, the right eye bounding box, and the infraorbital region bounding box to obtain student eye image data for the left eye region of interest, the right eye region of interest, and the infraorbital region of interest, respectively. These images are then stitched together in pixel space to obtain student eye image data for the region of interest.

[0010] Optionally, the learnable CLAHE module includes: The number of pixels corresponding to each brightness level in the student eye image data of the region of interest is counted. The brightness probability distribution set is obtained by the ratio of the number of pixels to the total number of pixels in the student eye image data of the region of interest. The brightness entropy and brightness variance are calculated based on the brightness probability distribution set. The student eye image data of the region of interest is converted from the original color space to the YCbCr color space. The average value of the Cb channel and the average value of the Cr channel of the student eye image data of the region of interest in the YCbCr color space are calculated respectively. The ratio of the average value of the Cb channel to the average value of the Cr channel is calculated to obtain the skin color feature ratio. The brightness entropy, brightness variance and skin color feature ratio are input together into the learnable CLAHE parameter prediction network to obtain the CLAHE clipping constraint parameters and CLAHE grid size parameters. The student eye image data in the region of interest was divided into several image sub-blocks according to the CLAHE grid size parameter. Histogram equalization was performed on each image sub-block. The contrast enhancement range of the histogram was limited by the CLAHE clipping limit parameter. After equalization, the pixels of each image sub-block were recombined to obtain CLAHE-enhanced student eye image data.

[0011] Optionally, the improved H2Former feature encoding module includes: The student eye image data of the region of interest and the student eye image data enhanced by CLAHE are divided into several patches of fixed size. For each patch, the local brightness difference and the blood color concentration index of the white of the eye are calculated. The fatigue-related patch weights are obtained by weighting the local brightness difference and the blood color concentration index of the white of the eye according to the preset weight coefficient. Normalize the weights of all fatigue-related patches to obtain fatigue significance coefficients. Extend the fatigue significance coefficients to the same dimension as the token vector through linear mapping to form a significance vector. Multiply the significance vector element-wise with the token vector corresponding to the patch to obtain the significance-weighted patch token. The saliency-weighted patch tokens are input into the H2Former encoding layers of the original image stream and the augmented image stream, respectively. In each H2Former encoding layer, a windowed multi-head self-attention mechanism is performed on the saliency-weighted patch tokens, and the cross-stream learnable fatigue gating coefficients are calculated based on the output features of the same position in the original image stream and the augmented image stream. Within each attention head, the saliency-weighted patch tokens of the original image stream and the CLAHE-enhanced stream are gated and fused using cross-stream learnable fatigue gating coefficients to obtain a gated update token. The obtained gated update token is then concatenated with the original token and transformed through a feedforward network to obtain the output of the current layer. A fatigue significance threshold is set. For all significant weighted patch tokens whose fatigue significance coefficient is less than the fatigue significance threshold, token sparsification is performed. Significant weighted patch tokens with a fatigue significance coefficient lower than the fatigue significance threshold are eliminated or their weights are reduced during feature calculation. Only significant weighted patch tokens with high significance are retained for the coding layer output. The coding layer output includes the original image stream student eye feature map and the CLAHE enhanced image stream student eye feature map.

[0012] Optionally, the gating feature fusion module includes: The original image stream student eye feature map and the CLAHE enhanced image stream student eye feature map are divided into scales under a multi-scale structure, and a total of several layers are set. Each layer yields a set of original image stream patch feature vectors and a set of CLAHE enhanced image stream patch feature vectors. Calculate the local contrast entropy of each patch in the CLAHE enhanced stream patch feature vector set; Calculate the texture consistency index of the original image flow patch feature vector and the CLAHE enhanced flow patch feature vector at the same spatial location in the corresponding layer; The gated fusion weights are jointly calculated based on the local contrast entropy and texture consistency index of each patch. For each patch, the feature vectors of the original image stream patch and the CLAHE enhanced stream patch are weighted and fused according to the gating fusion weights to obtain a fused feature vector. The fused feature vectors of all patches are concatenated in spatial order to form the corresponding layer fused student eye feature map. The fused student eye feature maps obtained at all scales are aggregated through upsampling and cross-layer connection operations to obtain the fused student eye feature map.

[0013] Optionally, the confidence generation module includes: The fused student eye feature map is input into a semantic segmentation subnetwork containing a multi-scale context decoding structure. Upsampling, skip connections, and feature reconstruction operations are performed on the fused student eye feature map to restore the fused feature tensor to the same spatial resolution as the student eye image data of the region of interest. The channel feature vector of each pixel position in the fused feature tensor is input into the pixel-level classifier to predict the fatigue region. The predicted probability of the corresponding pixel belonging to the fatigue region is output. The predicted probabilities of all pixel positions together form the pixel-level mask map of the student's eye fatigue region. An uncertainty modeling branch is introduced into the fusion feature tensor. The uncertainty modeling branch outputs the fatigue classification standard deviation for each pixel location. The confidence score of the pixel is obtained by subtracting the fatigue classification standard deviation. The confidence scores of all pixel locations together form a pixel-level confidence heatmap of the student's eye fatigue region.

[0014] Optionally, the health warning module includes: Calculate the pixel-level mask image of the student's eye fatigue area, mark the pixels of conjunctival congestion category as 1, and mark the rest of the pixels as 0, and calculate the area ratio of the conjunctival congestion area. Based on the number of pixels in the infraorbital dark area in the pixel-level mask image of the student's eye fatigue area, calculate the grayscale deviation of the infraorbital dark area; The individualized eye fatigue index of students is obtained by multiplying the difference in the area ratio of conjunctival hyperemia region and the difference in grayscale deviation of infraorbital dark region by preset weighting coefficients and then summing them. The system assesses students' individualized eye fatigue index to determine their health warning level and outputs different levels of health warning results.

[0015] Optionally, the determination of the student health warning level includes: When a student's individualized eye fatigue index is less than the first threshold, a normal health warning level is output. When a student's individualized eye fatigue index is greater than or equal to the first threshold and less than the second threshold, a health warning level is output. When a student's individualized eye fatigue index is greater than or equal to the second threshold and less than the maximum threshold, a health warning level is output.

[0016] The beneficial effects of this invention are: (1) This invention proposes to input student eye image data of the region of interest and CLAHE-enhanced student eye image data into a dual-stream H2Former encoder, and calculate fatigue-related patch weights through local brightness difference and blood color concentration index of the white of the eye, realizing fatigue salience weighted deep feature extraction. Combined with salience weighted gating mechanism and multi-layer sparsity, the model can adaptively focus on key areas with fatigue signs, suppress noise and interference from non-fatigue features, and solve the segmentation failure problem caused by blurred fatigue region boundaries and low contrast in natural scenes of traditional segmentation methods. Under complex lighting, skin color diversity and multi-terminal acquisition conditions, the IoU and Dice coefficients of region segmentation are improved by more than 10% compared with single-stream segmentation methods, improving segmentation accuracy and cross-scene generalization ability.

[0017] (2) This invention introduces local contrast entropy and texture consistency index to jointly regulate the gating fusion weight in the multi-scale feature fusion stage, thereby realizing dynamic weighting of the original image flow and CLAHE enhanced flow features. This not only effectively highlights the segmentation performance of fatigue-weak areas such as conjunctival congestion and infraorbital darkening, but also significantly suppresses false positive artifacts caused by over-enhancement, thereby improving the credibility and intervention reference value of the model in actual health monitoring.

[0018] (3) The CLAHE parameter learnable prediction and H2Former gating fusion strategy of the present invention realizes dynamic adaptive processing under different camera, different lighting and different student skin color conditions. The individual baseline mechanism compares the current fatigue area sub-indicators with the historical baseline, outputs the individualized eye fatigue index, and gives the health warning level in combination with the adaptive threshold. It realizes large-scale, non-contact, fine-grained sports health monitoring for students, and improves the refinement and traceability of health intervention. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a student physical health dynamic monitoring system based on deep learning proposed in this invention; Figure 2 This is a block diagram of the improved H2Former encoder in a student physical health dynamic monitoring system based on deep learning proposed in this invention. Detailed Implementation

[0020] Example 1: Reference Figures 1-2 A student physical health dynamic monitoring system based on deep learning includes: The quality assessment module collects student eye image data and performs quality assessment on the raw student eye image data. Based on occlusion detection, reflection detection, and sharpness detection, it determines whether the image meets the preset quality threshold. Student eye image data that passes the quality threshold is considered qualified student eye image data. In this embodiment, the quality assessment module includes: The system uses cameras to capture raw, real-time images of students' eyes, including their entire faces. Calculate the occlusion rate of the student's eye image data in the region of interest. When the occlusion rate is less than or equal to the occlusion rate threshold, the occlusion detection is considered successful. In Example 1, the occlusion rate is obtained by calculating the ratio of the number of pixels identified as occluded in the student's eye image data in the region of interest to the total number of pixels in the student's eye image data in the region of interest. The occlusion rate is used to measure the degree of occlusion caused by eyelids, eyelashes, hair or glasses reflection in the student's eye image data in the region of interest. The occlusion rate threshold is a preset threshold used to limit the occlusion tolerance range.

[0021] Calculate the maximum brightness ratio of the student's eye image data in the region of interest. When the maximum brightness ratio is less than or equal to the reflection detection threshold, the reflection detection is considered passed. In Example 1, the maximum brightness ratio is obtained by calculating the ratio of the maximum brightness value in the student eye image data of the region of interest to the average brightness value in the student eye image data of the region of interest. The maximum brightness ratio is used to measure whether there is strong reflection in the student eye image data of the region of interest. The reflection detection threshold is a preset threshold used to limit the reflection tolerance range.

[0022] Calculate the image sharpness index of the student's eye image data in the region of interest. When the image sharpness index is greater than or equal to the image sharpness judgment threshold, the sharpness detection is considered passed. In Example 1, the image sharpness index is obtained by taking the square of the difference between the response value of each pixel in the student eye image data of the region of interest after Laplacian filtering and the mean of all response values, and by averaging and taking the square root of the total number of pixels in the student eye image data of the region of interest. The image sharpness index is used to measure the sharpness of the student eye image data of the region of interest. The image sharpness judgment threshold is a preset threshold used to limit the blur tolerance range.

[0023] When the occlusion detection, reflection detection, and sharpness detection all meet their corresponding threshold conditions, the original student eye image data is judged as qualified student eye image data.

[0024] The region of interest extraction module performs facial landmark localization on qualified student eye image data, and crops out the student eye image data of the region of interest for both eyes and the infraorbital region based on the facial landmark localization results; In this embodiment, the region of interest extraction module includes: Input qualified students' eye image data into the facial landmark detection model to obtain a set of facial landmark coordinates; In Example 1, qualified student eye image data is input into a facial landmark detection model. The facial landmark detection model trained using a deep learning algorithm locates the position of each facial landmark in the qualified student eye image data, obtaining a set of facial landmark coordinates. The set of facial landmark coordinates refers to the set of horizontal and vertical pixel coordinates of all facial landmarks automatically identified and recorded by the facial landmark detection model in the two-dimensional pixel space of the qualified student eye image data. Each facial landmark uniquely corresponds to one horizontal pixel coordinate and one vertical pixel coordinate, and the total number of facial landmarks is a preset number.

[0025] The left eye key point index set and the right eye key point index set are determined based on the facial key point coordinate set, and the left eye bounding box and the right eye bounding box are calculated based on the left eye key point index set and the right eye key point index set, respectively. In Example 1, the left eye keypoint index set and the right eye keypoint index set refer to the sets of keypoint numbers that belong to the left eye or right eye region, identified by the face keypoint detection model. The left eye bounding box is obtained by extracting the minimum value of the horizontal pixel coordinates of all keypoints in the left eye keypoint index set as the left boundary, the minimum value of the vertical pixel coordinates as the upper boundary, the maximum value of the horizontal pixel coordinates as the right boundary, and the maximum value of the vertical pixel coordinates as the lower boundary, resulting in a quadruple bounding box. The calculation method of the right eye bounding box is the same as that of the left eye bounding box.

[0026] Calculate the infraorbital region bounding box based on the left and right eye bounding boxes; In Example 1, the infraorbital region bounding box is defined as follows: the minimum horizontal pixel coordinate of the left eye bounding box and the right eye bounding box is taken as the left boundary, the maximum horizontal pixel coordinate is taken as the right boundary, and the maximum vertical pixel coordinate of the lower boundary is taken as the upper boundary. The upper boundary is multiplied by a preset vertical scaling factor of the infraorbital region according to the eye height to obtain the lower boundary of the infraorbital region bounding box. All boundary coordinates of the infraorbital region bounding box are in pixels.

[0027] For qualified student eye image data, cropping operations are performed sequentially based on the left eye bounding box, the right eye bounding box, and the infraorbital region bounding box to obtain student eye image data for the left eye region of interest, the right eye region of interest, and the infraorbital region of interest, respectively. These images are then stitched together in pixel space to obtain student eye image data for the region of interest. The learnable CLAHE module inputs the student eye image data of the region of interest into the learnable CLAHE parameter prediction module. Based on the brightness histogram and skin color features of the student eye image data of the region of interest, the module predicts the key parameters of CLAHE and uses the key parameters of CLAHE to perform CLAHE enhancement processing on the student eye image data of the region of interest, generating CLAHE enhanced student eye image data.

[0028] In this implementation, the CLAHE module can be learned, including: The number of pixels corresponding to each brightness level in the student eye image data of the region of interest is counted. The brightness probability distribution set is obtained by the ratio of the number of pixels to the total number of pixels in the student eye image data of the region of interest. The brightness entropy and brightness variance are calculated based on the brightness probability distribution set. The luminance probability distribution set reflects the relative frequency of each luminance level in the student eye image data of the region of interest. The luminance entropy is used to measure the uniformity of the luminance distribution in the student eye image data of the region of interest. The luminance is obtained by taking the logarithm of the probability of each luminance level in the luminance probability distribution set and then summing the weighted values. The luminance variance is used to measure the luminance dispersion of the student eye image data of the region of interest. It is obtained by summing the squares of the differences between the luminance value of each pixel in the student eye image data of the region of interest and the mean luminance, and then normalizing and taking the square root of the summation.

[0029] The student eye image data of the region of interest is converted from the original color space to the YCbCr color space. The average value of the Cb channel and the average value of the Cr channel of the student eye image data of the region of interest in the YCbCr color space are calculated respectively. The ratio of the average value of the Cb channel to the average value of the Cr channel is calculated to obtain the skin color feature ratio. The student eye image data in the region of interest is converted pixel by pixel from the original color space to the YCbCr color space. In the YCbCr color space, the Cb channel values ​​of all pixels in the student eye image data in the region of interest are summed and then divided by the total number of pixels in the student eye image data in the region of interest to obtain the average value of the Cb channel. The Cr channel values ​​of all pixels in the student eye image data in the region of interest are summed and then divided by the total number of pixels in the student eye image data in the region of interest. The skin color feature ratio indicates the degree of color shift in the student eye image data in the region of interest.

[0030] The brightness entropy, brightness variance and skin color feature ratio are input together into the learnable CLAHE parameter prediction network to obtain the CLAHE clipping constraint parameters and CLAHE grid size parameters. In Example 1, the luminance entropy, luminance variance, and skin color feature ratio are input into a learnable CLAHE parameter prediction network. The CLAHE parameter prediction network performs linear weighting and bias summation on the luminance entropy, luminance variance, and skin color feature ratio, respectively, and outputs CLAHE shearing constraint parameters and CLAHE grid size parameters. The CLAHE shearing constraint parameters are used to limit the maximum contrast of local histogram enhancement, and the CLAHE grid size parameters are used to determine the grid size used for local histogram equalization. The CLAHE shearing constraint parameters and CLAHE grid size parameters are obtained by multiplying the luminance entropy, luminance variance, and skin color feature ratio by their corresponding weight coefficients, which are learned by the CLAHE parameter prediction network. All weighted results are summed and then a bias term learned by the network is added to obtain the CLAHE shearing constraint parameters and CLAHE grid size parameters.

[0031] The student eye image data in the region of interest was divided into several image sub-blocks according to the CLAHE grid size parameter. Histogram equalization was performed on each image sub-block. The contrast enhancement range of the histogram was limited by the CLAHE clipping limit parameter. After equalization, the pixels of each image sub-block were recombined to obtain CLAHE-enhanced student eye image data. The H2Former feature encoding module is improved by inputting the student eye image data of the region of interest and the CLAHE-enhanced student eye image data into the improved H2Former encoder to extract the student eye feature map of the original image stream and the student eye feature map of the CLAHE-enhanced image stream. In this embodiment, the H2Former feature encoding module is improved, including: The student eye image data of the region of interest and the student eye image data enhanced by CLAHE are divided into several patches of fixed size. For each patch, the local brightness difference and the blood color concentration index of the white of the eye are calculated. The fatigue-related patch weights are obtained by weighting the local brightness difference and the blood color concentration index of the white of the eye according to the preset weight coefficient. In Example 1, the local brightness difference is used to reflect the magnitude of the brightness change in the local area after CLAHE enhancement. The local brightness difference is obtained by calculating the absolute value of the difference between the average brightness channel value of the patch in the CLAHE-enhanced student eye image data and the average brightness channel value of the patch in the region of interest student eye image data.

[0032] The sclera blood color intensity index is used to reflect the range of color change in a local area after CLAHE enhancement. It is obtained by subtracting the ratio of the average Cr channel value to the average Cb channel value of the patch in the CLAHE-enhanced student eye image data from the ratio of the average Cr channel value to the average Cb channel value of the patch in the region of interest student eye image data.

[0033] Fatigue-related patch weights are used to measure the contribution of patch regions to the identification of signs of eye fatigue; ; in, This represents the local brightness difference. This is an indicator of the concentration of blood color in the whites of the eyes. These are preset weighting coefficients.

[0034] Normalize the weights of all fatigue-related patches to obtain fatigue significance coefficients. Extend the fatigue significance coefficients to the same dimension as the token vector through linear mapping to form a significance vector. Multiply the significance vector element-wise with the token vector corresponding to the patch to obtain the significance-weighted patch token. In Example 1, the fatigue significance coefficient is obtained by subtracting the minimum weight among all patches from the fatigue-related patch weight of each patch, and then dividing by the difference between the maximum and minimum weights among all patches.

[0035] Each patch contains all pixels that are expanded in sequence and fed into a linear projection layer. The linear projection layer outputs a token vector by multiplying the matrix with the learnable parameters and adding the learnable bias. The token vector represents the deep features of the patch.

[0036] The saliency-weighted patch tokens are input into the H2Former encoding layers of the original image stream and the augmented image stream, respectively. In each H2Former encoding layer, a windowed multi-head self-attention mechanism is performed on the saliency-weighted patch tokens, and the cross-stream learnable fatigue gating coefficients are calculated based on the output features of the same position in the original image stream and the augmented image stream. ; ; in, Indicates the first The saliency-weighted patch token sequence output by the layer. Let ori represent the original image flow and cla represent the CLAHE-enhanced flow, reflecting the local detail features of the student eye image data and the CLAHE-enhanced student eye image data in the region of interest. W-MSA This is a hierarchical window multi-head self-attention operator used for local self-attention feature modeling of the input saliency-weighted patch token sequence at each level. It is used for adaptive modeling of features of different spatial locations in the student's eye region, including the conjunctiva and orbit. Indicates the first The feature matrix after multi-head self-attention contains multi-scale semantic features of the regions showing signs of eye fatigue in students. Indicates the first Layer The original image flow saliency weighted patch token and the first The cross-flow learnable fatigue gating coefficients between CLAHE-enhanced flow significance-weighted patch tokens This represents the Sigmoid activation function. This represents the learnable temperature coefficient in the fatigue gating mechanism, used to adjust the scale of the L1 distance, thus giving higher weight to the fatigue region in the gating interaction. This represents the L1 norm.

[0037] Within each attention head, the saliency-weighted patch tokens of the original image stream and the CLAHE-enhanced stream are gated and fused using cross-stream learnable fatigue gating coefficients to obtain a gated update token. The obtained gated update token is then concatenated with the original token and transformed through a feedforward network to obtain the output of the current layer.

[0038] In Example 1, for each patch in the original image stream, the saliency-weighted patch tokens of all patches in the corresponding CLAHE-enhanced stream at the same position are multiplied by the gating coefficients of the original image stream patch, and all product results are summed to obtain the gating update token of the current patch in the original image stream. The same method is used for the CLAHE-enhanced stream patch, calculating the gating update token of the CLAHE-enhanced stream using the current patch of the CLAHE-enhanced stream, the saliency-weighted patch tokens of all patches in the original image stream, and the corresponding gating coefficients. The gating update token of the current patch in the original image stream is concatenated with the original token of the current patch in the original image stream, and the same method is applied to the CLAHE-enhanced stream to obtain the concatenated patch features. The concatenated patch features are then input into the feedforward network to complete feature transformation, obtaining the output features of the current layer.

[0039] A fatigue significance threshold is set. For all significant weighted patch tokens whose fatigue significance coefficient is less than the fatigue significance threshold, token sparsification is performed. Significant weighted patch tokens with a fatigue significance coefficient lower than the fatigue significance threshold are eliminated or their weights are reduced during feature calculation. Only significant weighted patch tokens with high significance are retained for the coding layer output. The coding layer output includes the original image stream student eye feature map and the CLAHE enhanced image stream student eye feature map.

[0040] In Example 1, the fatigue significance threshold is used to distinguish between high and low significance weighted patch tokens. It is obtained by averaging the fatigue significance coefficients of all significance weighted patch tokens and multiplying the average fatigue significance coefficients of all significance weighted patch tokens by a preset downsampling ratio coefficient and then adding the average. The original student eye feature map and the CLAHE enhanced student eye feature map are both obtained by concatenating the output features of the remaining significance weighted patch tokens after sparsification.

[0041] The gated feature fusion module performs cross-scale gated fusion processing on the student eye feature map of the original image stream and the student eye feature map of the CLAHE enhanced stream during the multi-scale feature fusion stage. The fusion weight is jointly controlled by the local contrast entropy and the texture consistency index to obtain the fused student eye feature map. In this embodiment, the gating feature fusion module includes: The original image stream student eye feature map and the CLAHE enhanced image stream student eye feature map are divided into scales under a multi-scale structure, and a total of several layers are set. Each layer yields a set of original image stream patch feature vectors and a set of CLAHE enhanced image stream patch feature vectors. The original image stream patch feature vector set consists of the original image stream student eye feature vector of each patch in the current layer, and the CLAHE enhanced stream patch feature vector set consists of the CLAHE enhanced stream student eye feature vector of each patch in the current layer. The original image stream student eye feature vector and the CLAHE enhanced stream student eye feature vector of each patch have the same feature dimension, and the number of patches in each layer is the same.

[0042] Calculate the local contrast entropy of each patch in the CLAHE enhanced stream patch feature vector set; In Example 1, the feature vector of the corresponding patch is normalized in the channel dimension. After normalization, each channel obtains a probability value. The logarithm of the probability values ​​of all channels is taken and then multiplied by the probability value itself. The weighted results of all channels are summed and the negative value is taken to obtain the local contrast entropy of the corresponding patch. The local contrast entropy is used to reflect the texture intensity distribution of fatigue areas such as conjunctival congestion or infraorbital darkening.

[0043] Calculate the texture consistency index of the original image flow patch feature vector and the CLAHE enhanced flow patch feature vector at the same spatial location in the corresponding layer; In Example 1, the inner product of the original image flow patch feature vector and the CLAHE enhanced flow patch feature vector, and the L2 norm of the original image flow patch feature vector and the CLAHE enhanced flow patch feature vector are calculated respectively. The inner product result is divided by the product of the two L2 norms to obtain the texture consistency index. The texture consistency index is used to measure the consistency of the original image flow patch feature vector and the CLAHE enhanced flow patch feature vector in texture direction and energy. The larger the texture consistency index value, the more consistent the texture of the two patches is. ; in, Texture consistency metrics are used to identify whether false textures or distorted structures have been introduced into enhanced images. This represents the inner product operation. Describing the L2 norm, and They represent the first Layer The original student eye feature vector of the patched image and the student eye feature vector of the CLAHE enhanced image.

[0044] The gated fusion weights are jointly calculated based on the local contrast entropy and texture consistency index of each patch. In Example 1, the local contrast entropy is multiplied by a preset fusion control coefficient, and the texture consistency index is subtracted from one and then multiplied by another preset fusion control coefficient. The two are added to the bias term and then input into the Sigmoid activation function to obtain the gated fusion weight. The gated fusion weight is used to control the proportion of CLAHE enhanced flow patch features in the fusion. The higher the gated fusion weight value, the greater the proportion of the corresponding patch's CLAHE enhanced flow features in the fusion. ; in, For local contrast entropy, The preset fusion control coefficient, For bias terms, It is the Sigmoid activation function. For gating fusion weights.

[0045] For each patch, the feature vectors of the original image stream patch and the CLAHE enhanced stream patch are weighted and fused according to the gating fusion weights to obtain a fused feature vector. The fused feature vectors of all patches are concatenated in spatial order to form the corresponding layer fused student eye feature map. The fused student eye feature maps obtained at all scales are aggregated through upsampling and cross-layer connection operations to obtain the fused student eye feature map.

[0046] In Example 1, the CLAHE enhanced stream patch feature vector of the patch is multiplied by the gated fusion weight, and the original image stream patch feature vector of the patch is multiplied by one minus the gated fusion weight. The two results are added together to obtain the fusion feature vector of the patch. The current layer fused student eye feature map is composed of the fusion feature vector of each patch.

[0047] The confidence generation module inputs the student's eye feature map into the semantic segmentation subnetwork and outputs a pixel-level mask map of the student's eye fatigue area and a pixel-level confidence heatmap of the student's eye fatigue area.

[0048] In this embodiment, the confidence generation module includes: The fused student eye feature map is input into a semantic segmentation subnetwork containing a multi-scale context decoding structure. Upsampling, skip connections, and feature reconstruction operations are performed on the fused student eye feature map to restore the fused feature tensor to the same spatial resolution as the student eye image data of the region of interest. In Example 1, the semantic segmentation subnetwork upsamples the fused student eye feature map according to its spatial structure, restoring the low-resolution features to a higher resolution through the upsampling operation. After each upsampling, the features are concatenated with the fused student eye feature map at the same spatial scale in the encoding stage through skip connections. The concatenated features undergo feature reconstruction, which includes convolution transformation and activation function processing. After multiple upsampling, skip connections, and feature reconstruction operations, a fused feature tensor with the same spatial resolution as the student eye image data in the region of interest is finally obtained. Each pixel position of the fused feature tensor corresponds to a feature vector containing multi-channel fused features.

[0049] The channel feature vector of each pixel position in the fused feature tensor is input into the pixel-level classifier to predict the fatigue region. The predicted probability of the corresponding pixel belonging to the fatigue region is output. The predicted probabilities of all pixel positions together form the pixel-level mask map of the student's eye fatigue region. In Example 1, the channel feature vector of each pixel position in the fused feature tensor is input into a pixel-level classifier for fatigue region prediction. The pixel-level classifier performs a weighted summation by multiplying the channel feature vector of the pixel position with the corresponding elements of the preset classification weight vector. The weighted summation result is then added to a preset bias term and input into a Sigmoid activation function. The Sigmoid activation function outputs the predicted probability that the corresponding pixel belongs to the fatigue region. The predicted probabilities of all pixel positions together form a pixel-level mask of the student's eye fatigue region. Each pixel value in the pixel-level mask of the student's eye fatigue region is between 0 and 1. The larger the pixel value, the higher the probability of belonging to the fatigue region.

[0050] An uncertainty modeling branch is introduced into the fusion feature tensor. The uncertainty modeling branch outputs the fatigue classification standard deviation for each pixel location. The confidence score of the pixel is obtained by subtracting the fatigue classification standard deviation. The confidence scores of all pixel locations together form a pixel-level confidence heatmap of the student's eye fatigue region.

[0051] In Example 1, the uncertainty modeling branch inputs the channel feature vector of each pixel location into the prediction head. Through linear weighting and biasing, it outputs the fatigue classification mean and fatigue classification variance of the pixel location. The fatigue classification variance is squared to obtain the fatigue classification standard deviation, which is used to measure the uncertainty of the pixel in the fatigue region prediction. The confidence level of the pixel location is obtained by subtracting the fatigue classification standard deviation from 1. The confidence levels of all pixel locations together form the pixel-level confidence heatmap of the student's eye fatigue region. Each pixel value in the pixel-level confidence heatmap of the student's eye fatigue region is between 0 and 1. The larger the pixel value, the more stable the model judgment at the corresponding location.

[0052] The health early warning module calculates the proportion of conjunctival congestion area and the grayscale deviation of infraorbital dark area in the pixel-level mask image of the student's eye fatigue area, generating the results of the student's eye fatigue sub-indices. It then calculates the difference between the results of the student's eye fatigue sub-indices and the corresponding student eye indicators, outputs the student's individualized eye fatigue index, and determines the student's health early warning level based on the individualized eye fatigue index.

[0053] In this embodiment, the health warning module includes: Calculate the pixel-level mask image of the student's eye fatigue area, mark the pixels of conjunctival congestion category as 1, and mark the rest of the pixels as 0, and calculate the area ratio of the conjunctival congestion area. In Example 1, the area ratio of the conjunctival congestion region represents the proportion of the conjunctival congestion region in the total area of ​​the student's eye image data in the region of interest. It is obtained by summing the number of pixels belonging to the conjunctival congestion region among all pixels and then dividing by the total number of pixels in the student's eye image data in the region of interest.

[0054] Based on the number of pixels in the infraorbital dark area in the pixel-level mask image of the student's eye fatigue area, calculate the grayscale deviation of the infraorbital dark area; In Example 1, the grayscale deviation of the infraorbital dark region is used to measure the difference between the average brightness of the infraorbital dark region and the average brightness of the infraorbital region in the student's historical baseline image. It is obtained by subtracting the grayscale values ​​of all pixels belonging to the infraorbital dark region in the brightness channel from the average brightness value of the infraorbital region in the student's historical baseline image, taking the absolute values, summing them, and dividing by the number of pixels in the infraorbital dark region.

[0055] The individualized eye fatigue index of students is obtained by multiplying the difference in the area ratio of conjunctival hyperemia region and the difference in grayscale deviation of infraorbital dark region by preset weighting coefficients and then summing them. The individualized eye fatigue index for students is used to comprehensively reflect the overall deviation between the proportion of conjunctival congestion area and the grayscale deviation of the infraorbital dark area.

[0056] The system assesses students' individualized eye fatigue index to determine their health warning level and outputs different levels of health warning results.

[0057] In this implementation method, the determination of the student health warning level includes: When a student's individualized eye fatigue index is less than the first threshold, a normal health warning level is output. When a student's individualized eye fatigue index is greater than or equal to the first threshold and less than the second threshold, a health warning level is output. When a student's individualized eye fatigue index is greater than or equal to the second threshold and less than the maximum threshold, a health warning level is output.

[0058] Example 2: During a routine application of a student physical health dynamic monitoring system, the monitoring terminal automatically collected a batch of raw student eye image data. The system performed a quality assessment on this batch of raw student eye image data. In one of the raw student eye image data, the occlusion detection module detected 13 occluded pixels out of a total of 12,800 pixels, with an occlusion rate of 0.10%. Since the occlusion rate was lower than the set threshold of 0.2%, the image passed the occlusion detection. The reflection detection module detected a maximum brightness value of 238 and an average brightness value of 112 in the region of interest of the image, calculating a maximum brightness ratio of 2.12, which was lower than the reflection detection threshold of 2.5, thus the reflection detection was deemed successful. The sharpness detection module used the Laplacian operator to calculate the image sharpness index of 21.6, which was greater than the set threshold of 16.0, thus the sharpness detection was deemed successful. Finally, the raw student eye image data was automatically determined by the system to be qualified student eye image data.

[0059] The system inputs qualified student eye image data into the facial landmark detection model. The model automatically outputs a set of facial landmark coordinates, including 68 facial landmarks. The left eye landmark index set includes landmarks 37 to 42, and the right eye landmark index set includes landmarks 43 to 48. By taking the minimum horizontal pixel coordinate of 62 and the maximum vertical pixel coordinate of 91, and the minimum vertical pixel coordinate of 37 and the maximum vertical pixel coordinate of 51, the left eye bounding box is obtained as [62, 37, 91, 51]. The right eye bounding box is calculated similarly. The infraorbital region bounding box is obtained by taking the minimum horizontal pixel coordinate of 62 and the maximum vertical pixel coordinate of 123, and the maximum vertical lower boundary of 51, and setting the vertical scaling factor to 1.2. The lower boundary of the infraorbital region bounding box is 51 + 0.2 × (51 - 37) = 53.8. The system sequentially crops the student eye image data of the left eye region of interest, the right eye region of interest, and the infraorbital region of interest, and finally stitches them together in pixel space to obtain the student eye image data of the region of interest.

[0060] In the CLAHE enhancement stage, the system counts the number of pixels at each brightness level in the student eye image data of the region of interest, calculates the brightness probability distribution set, the image brightness entropy is 5.85, the brightness variance is 39.4, and converts to YCbCr space, the mean value of Cb channel is 107.3, the mean value of Cr channel is 143.2, and the skin color feature ratio is calculated to be 0.75. These three items are input into the CLAHE parameter prediction network, and the CLAHE clipping constraint parameter is dynamically predicted to be 2.0, and the CLAHE grid size parameter is 8. The system divides the student eye image data of the region of interest into 64 sub-blocks of 8×8, and performs histogram equalization on each sub-block separately. The histogram contrast enhancement is controlled by the CLAHE clipping constraint parameter 2.0. After pixel recombination, the CLAHE-enhanced student eye image data is output.

[0061] In the feature extraction stage, the system divides the student eye image data of the region of interest and the CLAHE-enhanced student eye image data into 48 fixed-size patches. For each patch, the local brightness difference (in Example 2, the average brightness of patch 17 increased from 102 to 135, a difference of 33) and the sclera blood color concentration index (in Example 2, the Cr / Cb ratio of patch 17 changed from 1.24 to 1.32, an index of 0.08) are calculated. The fatigue-related patch weights are obtained by weighting these values. After normalizing all patch weights, they are expanded into a saliency vector through a linear mapping. This saliency vector is then multiplied element-wise with the token vector corresponding to each patch to obtain the saliency-weighted patch token.

[0062] The system inputs saliency-weighted patch tokens into the H2Former encoding layers of the original image stream and the CLAHE-enhanced stream. In the second-layer windowed multi-head self-attention mechanism, the L1 distance between the feature vector of patch 17 in the original image stream and the feature vector of patch 17 in the CLAHE-enhanced stream is 0.18, the fatigue gating temperature coefficient is 0.75, and the calculated cross-flow fatigue gating coefficient is 0.62. The system weights and fuses the patch features according to the above gating coefficient and inputs them in series into the feedforward network to complete the feature transformation. All tokens of patches below the fatigue saliency threshold (mean 0.41 plus standard deviation 0.08 multiplied by 1.5 equals 0.53 in Example 2) are weighted or removed in the next step, and only highly saliency patches are retained. Finally, the encoding layer outputs 38 student eye features from the original image stream and 38 student eye features from the CLAHE-enhanced stream.

[0063] In the multi-scale feature fusion stage, the system calculates the local contrast entropy (the normalized probability values ​​of each channel of patch 11 in Example 2 are 0.16, 0.11, 0.08..., and the entropy is 2.41) and texture consistency index (0.73 for patch 11 in Example 2) for each layer of original image stream and CLAHE enhancement stream patch feature vector set, and sets the fusion control coefficients to 0.6, 0.7, and bias to -0.1. After Sigmoid activation, the gated fusion weight is 0.58. The final fused feature vector of patch 11 is the original image stream feature vector multiplied by 0.42 plus the CLAHE enhancement stream feature vector multiplied by 0.58. After all patches are fused, they are spliced ​​in spatial order and then upsampled and connected across layers to form a complete fused student eye feature map.

[0064] The fused feature map is input into the semantic segmentation subnetwork, and the feature tensor is upsampled to a spatial resolution of 128×64. A pixel-level classifier outputs the fatigue region probability for each pixel (in Example 2, the probability of pixel (25,40) is 0.93), forming a pixel-level mask image of the student's eye fatigue region. This image contains 2430 pixels of conjunctival hyperemia and 1521 pixels of infraorbital hyperpigmentation, for a total of 8192 pixels in the mask image. The system simultaneously generates a confidence heatmap; in Example 2, the confidence score for pixel (25,40) is 0.87, indicating stable classification.

[0065] In the indicator quantification and health warning stage, the system first calculates the area ratio of conjunctival hyperemia (2430 / 8192 = 0.297) and the grayscale deviation of infraorbital hypopigmentation (0.14, baseline 0.08), with a current deviation of 0.06. The sub-indices are then compared with the student's historical baseline (conjunctival hyperemia area ratio 0.09, infraorbital hypopigmentation grayscale deviation 0.08), with preset weighting coefficients of 0.5 and 0.5, resulting in an individualized eye fatigue index of 0.5 × (0.297 - 0.09) + 0.5 × (0.14 - 0.08) = 0.1335. Based on the health warning level thresholds (first threshold 0.08, second threshold 0.16), the current index falls within the attention range, outputting a health warning level, and the system automatically pushes a notification to the physical education teacher and parents.

[0066] To verify the difference between the present invention and traditional methods, the system simultaneously randomly selected 500 samples for comparative evaluation.

[0067] The method of this invention achieves an average Dice coefficient of 0.88 for conjunctival hyperemia region segmentation, compared to only 0.77 for the traditional U-Net+CLAHE method. The false detection rate for infraorbital hypopigmentation regions is 4.9%, compared to 15.7% for the traditional method. The mean confidence heatmap value is improved by 14%, and the mean square error for boundary localization is reduced to 2.0 pixels (compared to 5.2 for the traditional method). In the same batch of samples, the traditional method resulted in misclassification of 35 mask image regions due to uneven lighting and skin color deviation, leading to segmentation result offsets and discrepancies between health warning results and actual human assessments.

[0068] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A student physical health dynamic monitoring system based on deep learning, characterized in that, include: The quality assessment module collects students' eye image data and performs quality assessments, and uses the eye image data of students that pass the quality threshold as qualified student eye image data. The region of interest extraction module performs facial key point localization on qualified student eye image data, and crops the student eye image data of the region of interest for both eyes and the infraorbital region based on the facial key point localization results; The learnable CLAHE module inputs student eye image data of the region of interest into the learnable CLAHE parameter prediction module. Based on the brightness histogram and skin color features of the student eye image data of the region of interest, the module predicts the key parameters of CLAHE and performs CLAHE enhancement processing on the student eye image data of the region of interest to generate CLAHE enhanced student eye image data. The H2Former feature encoding module is improved by inputting the student eye image data of the region of interest and the CLAHE-enhanced student eye image data into the improved H2Former encoder to extract the student eye feature map of the original image stream and the student eye feature map of the CLAHE-enhanced image stream. The gated feature fusion module performs cross-scale gated fusion processing on the student eye feature map of the original image stream and the student eye feature map of the CLAHE enhanced stream during the multi-scale feature fusion stage to obtain a fused student eye feature map. The confidence generation module will input the student's eye feature map into the semantic segmentation subnetwork and output a pixel-level mask map of the student's eye fatigue area and a pixel-level confidence heatmap of the student's eye fatigue area. The health early warning module calculates the percentage of conjunctival congestion area and the grayscale deviation of infraorbital dark area as sub-indices of student eye fatigue based on pixel-level mask images of student eye fatigue areas, and then determines the student's health early warning level.

2. The student physical health dynamic monitoring system based on deep learning according to claim 1, characterized in that, The quality assessment module includes: The system uses cameras to capture raw, real-time images of students' eyes, including their full faces. Calculate the occlusion rate of the student's eye image data in the region of interest. When the occlusion rate is less than or equal to the occlusion rate threshold, the occlusion detection is considered successful. Calculate the maximum brightness ratio of the student's eye image data in the region of interest. When the maximum brightness ratio is less than or equal to the reflection detection threshold, the reflection detection is considered passed. Calculate the image sharpness index of the student's eye image data in the region of interest. When the image sharpness index is greater than or equal to the image sharpness judgment threshold, the sharpness detection is considered passed. When the occlusion detection, reflection detection, and sharpness detection all meet their corresponding threshold conditions, the original student eye image data is judged as qualified student eye image data.

3. The student physical health dynamic monitoring system based on deep learning according to claim 1, characterized in that, The region of interest extraction module includes: Input qualified students' eye image data into the facial landmark detection model to obtain a set of facial landmark coordinates; The left eye key point index set and the right eye key point index set are determined based on the facial key point coordinate set, and the left eye bounding box and the right eye bounding box are calculated based on the left eye key point index set and the right eye key point index set, respectively. Calculate the infraorbital region bounding box based on the left and right eye bounding boxes; For qualified student eye image data, cropping operations are performed sequentially based on the left eye bounding box, the right eye bounding box, and the infraorbital region bounding box to obtain student eye image data for the left eye region of interest, the right eye region of interest, and the infraorbital region of interest, respectively. These images are then stitched together in pixel space to obtain student eye image data for the region of interest.

4. The student physical health dynamic monitoring system based on deep learning according to claim 1, characterized in that, The learnable CLAHE module includes: The number of pixels corresponding to each brightness level in the student eye image data of the region of interest is counted. The brightness probability distribution set is obtained by the ratio of the number of pixels to the total number of pixels in the student eye image data of the region of interest. The brightness entropy and brightness variance are calculated based on the brightness probability distribution set. The student eye image data of the region of interest is converted from the original color space to the YCbCr color space. The average value of the Cb channel and the average value of the Cr channel of the student eye image data of the region of interest in the YCbCr color space are calculated respectively. The ratio of the average value of the Cb channel to the average value of the Cr channel is calculated to obtain the skin color feature ratio. The brightness entropy, brightness variance and skin color feature ratio are input together into the learnable CLAHE parameter prediction network to obtain the CLAHE clipping constraint parameters and CLAHE grid size parameters. The student eye image data in the region of interest was divided into several image sub-blocks according to the CLAHE grid size parameter. Histogram equalization was performed on each image sub-block. The contrast enhancement range of the histogram was limited by the CLAHE clipping limit parameter. After equalization, the pixels of each image sub-block were recombined to obtain CLAHE-enhanced student eye image data.

5. The student physical health dynamic monitoring system based on deep learning according to claim 1, characterized in that, The improved H2Former feature encoding module includes: The student eye image data of the region of interest and the student eye image data enhanced by CLAHE are divided into several patches of fixed size. For each patch, the local brightness difference and the blood color concentration index of the white of the eye are calculated. The fatigue-related patch weights are obtained by weighting the local brightness difference and the blood color concentration index of the white of the eye according to the preset weight coefficient. Normalize the weights of all fatigue-related patches to obtain fatigue significance coefficients. Extend the fatigue significance coefficients to the same dimension as the token vector through linear mapping to form a significance vector. Multiply the significance vector element-wise with the token vector corresponding to the patch to obtain the significance-weighted patch token. The saliency-weighted patch tokens are input into the H2Former encoding layers of the original image stream and the augmented image stream, respectively. In each H2Former encoding layer, a windowed multi-head self-attention mechanism is performed on the saliency-weighted patch tokens, and the cross-stream learnable fatigue gating coefficients are calculated based on the output features of the same position in the original image stream and the augmented image stream. Within each attention head, the saliency-weighted patch tokens of the original image stream and the CLAHE-enhanced stream are gated and fused using cross-stream learnable fatigue gating coefficients to obtain a gated update token. The obtained gated update token is then concatenated with the original token and transformed through a feedforward network to obtain the output of the current layer. A fatigue significance threshold is set. For all significant weighted patch tokens whose fatigue significance coefficient is less than the fatigue significance threshold, token sparsification is performed. Significant weighted patch tokens with a fatigue significance coefficient lower than the fatigue significance threshold are eliminated or their weights are reduced during feature calculation. Only significant weighted patch tokens with high significance are retained for the coding layer output. The coding layer output includes the original image stream student eye feature map and the CLAHE enhanced image stream student eye feature map.

6. The student physical health dynamic monitoring system based on deep learning according to claim 1, characterized in that, The gated feature fusion module includes: The original image stream student eye feature map and the CLAHE enhanced image stream student eye feature map are divided into scales under a multi-scale structure, and a total of several layers are set. Each layer yields a set of original image stream patch feature vectors and a set of CLAHE enhanced image stream patch feature vectors. Calculate the local contrast entropy of each patch in the CLAHE enhanced stream patch feature vector set; Calculate the texture consistency index of the original image flow patch feature vector and the CLAHE enhanced flow patch feature vector at the same spatial location in the corresponding layer; The gated fusion weights are jointly calculated based on the local contrast entropy and texture consistency index of each patch. For each patch, the feature vectors of the original image stream patch and the CLAHE enhanced stream patch are weighted and fused according to the gating fusion weights to obtain a fused feature vector. The fused feature vectors of all patches are concatenated in spatial order to form the corresponding layer fused student eye feature map. The fused student eye feature maps obtained at all scales are aggregated through upsampling and cross-layer connection operations to obtain the fused student eye feature map.

7. The student physical health dynamic monitoring system based on deep learning according to claim 1, characterized in that, The confidence generation module includes: The fused student eye feature map is input into a semantic segmentation subnetwork containing a multi-scale context decoding structure. Upsampling, skip connections, and feature reconstruction operations are performed on the fused student eye feature map to restore the fused feature tensor to the same spatial resolution as the student eye image data of the region of interest. The channel feature vector of each pixel position in the fused feature tensor is input into the pixel-level classifier to predict the fatigue region. The predicted probability of the corresponding pixel belonging to the fatigue region is output. The predicted probabilities of all pixel positions together form the pixel-level mask map of the student's eye fatigue region. An uncertainty modeling branch is introduced into the fusion feature tensor. The uncertainty modeling branch outputs the fatigue classification standard deviation for each pixel location. The confidence score of the pixel is obtained by subtracting the fatigue classification standard deviation. The confidence scores of all pixel locations together form a pixel-level confidence heatmap of the student's eye fatigue region.

8. The student physical health dynamic monitoring system based on deep learning according to claim 1, characterized in that, The health warning module includes: Calculate the pixel-level mask image of the student's eye fatigue area, mark the pixels of conjunctival congestion category as 1, and mark the rest of the pixels as 0, and calculate the area ratio of the conjunctival congestion area. Based on the number of pixels in the infraorbital dark area in the pixel-level mask image of the student's eye fatigue area, calculate the grayscale deviation of the infraorbital dark area; The individualized eye fatigue index of students is obtained by multiplying the difference in the area ratio of conjunctival hyperemia region and the difference in grayscale deviation of infraorbital dark region by preset weighting coefficients and then summing them. The system assesses students' individualized eye fatigue index to determine their health warning level and outputs different levels of health warning results.

9. A student physical health dynamic monitoring system based on deep learning according to claim 8, characterized in that, The determination of the student health warning level includes: When a student's individualized eye fatigue index is less than the first threshold, a normal health warning level is output. When a student's individualized eye fatigue index is greater than or equal to the first threshold and less than the second threshold, a health warning level is output. When a student's individualized eye fatigue index is greater than or equal to the second threshold and less than the maximum threshold, a health warning level is output.

Images