Non-contact blood pressure detection method and system based on deep learning

Abstract

The present invention discloses a non-contact blood pressure detection method and system based on deep learning, which relates to the field of computer vision technology, including the following steps: step 1: shooting a face video and collecting a data set; step 2: extracting frame-by-frame images from the face video and calculating the average pixel value to obtain an IPPG signal; step 3: constructing an IPPG-Mamba network model and training the IPPG-Mamba network model using the extracted IPPG signal data set to obtain a blood pressure detection model; step 4: collecting a face video file to be detected and extracting the IPPG signal; step 5: inputting the IPPG signal to be detected into the blood pressure detection model to obtain a blood pressure detection result. The network model used in the present invention combines HRV feature extraction and an improved CNN-Mamba Block. Through multi-feature fusion and a new network structure design, it can better extract the pulse wave characteristics of the IPPG signal and improve the accuracy of blood pressure prediction.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and more particularly to a non-contact blood pressure detection method and system based on deep learning. Background Art

[0002] Currently, blood pressure measurement mainly relies on traditional cuff or wrist blood pressure monitors, which require manual operation or specific hardware equipment. Although this type of method is mature in clinical application and has a certain degree of measurement accuracy, it has obvious limitations in daily use. Not only is the operation cumbersome, but it can also easily cause user discomfort during high-frequency or continuous measurements, especially at night or in work scenarios, making it difficult to achieve efficient and convenient monitoring. In addition, this contact measurement method is not user-friendly for certain special groups, such as children, the elderly, or those with sensitive skin, and it is difficult to meet the daily health needs of a wide range of people.

[0003] With the development of biosignal acquisition and intelligent sensing technologies, photoplethysmography (PPG) has become a widely used non-invasive method for collecting physiological signals. It can infer an individual's vital signs, such as heart rate, blood oxygen, and pulse rate, by detecting tiny changes in subcutaneous blood volume. However, traditional PPG signal acquisition relies on sensor devices that come into direct contact with the skin, such as finger clips and wristbands. This limits the flexibility of its application scenarios. Especially with the increasing demand for contactless services such as remote monitoring and mobile healthcare, the contact issues of existing PPG devices are gradually being exposed.

[0004] To overcome the inconvenience of contact sensors in practical applications, researchers have proposed image-based photoplethysmography (IPPG) technology. This technology uses ordinary cameras to capture facial videos and uses image processing and signal extraction techniques to analyze pulse signals reflected by subtle changes in skin color as blood flow, thereby achieving non-contact monitoring of vital signs. Compared with traditional PPG, IPPG is significantly more comfortable and scalable. Because its acquisition process does not require physical contact, it is more suitable for special groups such as children and the elderly, and also facilitates the implementation of intelligent applications such as telemedicine and home health management. Currently, IPPG technology has achieved high accuracy in non-contact heart rate and respiratory rate detection and has been verified in multiple experimental environments.

[0005] However, compared to vital signs such as heart rate and respiratory rate, estimating blood pressure is more challenging. Blood pressure is not only influenced by multiple physiological parameters such as pulse wave morphology, velocity, and amplitude, but also involves complex hemodynamic mechanisms, making its mapping to the IPPG signal more nonlinear and unstable. Traditional shallow learning models and feature engineering methods struggle to effectively establish a high-dimensional mapping between blood pressure and IPPG. Therefore, leveraging more powerful modeling capabilities to improve the accuracy of blood pressure estimation has become a key research issue.

[0006] In recent years, with the rapid development of deep learning technology, researchers have begun exploring the use of models such as convolutional neural networks (CNNs), recurrent neural networks (RNNs), and Transformers to model IPPG signals for more accurate blood pressure estimation. While these models offer advantages in feature extraction and sequence modeling, they still face limitations in processing long time series, low latency, and weak signal interference. In particular, in complex environments, such as those with varying lighting, skin color, facial expression, or posture, the models suffer from insufficient robustness and poor generalization, hindering their widespread adoption in practical applications.

[0007] Therefore, how to improve the accuracy and applicability of blood pressure estimation using IPPG signals is an urgent problem that needs to be solved by those skilled in the art. Summary of the Invention

[0008] In view of this, the present invention provides a non-contact blood pressure detection method and system based on deep learning to solve the problems of inconvenient contact, poor environmental adaptability and insufficient accuracy existing in traditional blood pressure detection methods.

[0009] In order to achieve the above object, the present invention adopts the following technical solutions:

[0010] The non-contact blood pressure detection method based on deep learning includes the following steps:

[0011] Step 1: Collect face video samples, including face video and blood pressure data;

[0012] Step 2: Extract frame-by-frame images from the face video and calculate the pixel mean in the image to generate the IPPG signal;

[0013] Step 3: Construct an IPPG-Mamba network model and train it based on IPPG signals and blood pressure data to obtain a blood pressure detection model. The IPPG-Mamba network model includes an HRV feature extraction unit, a CNN-MambaBlock, a regression layer, and a feedforward neural network. The HRV feature extraction unit extracts the HRV feature sequence. The CNN-Mamba Block is an improved network module that combines the CNN network and the Mamba Block with wavelet transform convolution to extract fused features. The regression layer integrates the fused features to obtain deep features. The feedforward neural network fuses the HRV feature sequence and deep features to predict blood pressure detection results.

[0014] Step 4: Collect the face video to be detected, extract frame-by-frame images, calculate the pixel mean in the image, and generate the IPPG signal to be detected;

[0015] Step 5: Input the IPPG signal to be detected into the blood pressure detection model for detection to obtain the blood pressure detection result.

[0016] Preferably, the feedforward neural network includes a linear layer Linear 64, an activation function ELU, a linear layer Linear 32, an activation function ELU, and a linear layer Linear 1 connected in sequence.

[0017] Preferably, the CNN-Mamba Block includes a root mean square normalization layer RMS Norm, two groups of linear layers LinearLayer, a CNN network module, an activation function SILU, a selection state space model SSM and two groups of residual modules; the CNN network module includes two groups of transformation modules and a dropout layer Dropout, each group of transformation modules includes a one-dimensional wavelet transform convolution layer WaveletTransform Convolution 1D (WTConv1d), a batch normalization layer BatchNorm and a one-dimensional maximum pooling layer MaxPool1d; the first group of residual modules includes a linear layer and an activation function.

[0018] Preferably, the regression layer includes a root mean square normalization layer, a dropout layer, a linear layer and a global average pooling layer.

[0019] Preferably, the one-dimensional wavelet transform convolution layer WTConv1d includes a wavelet transform module, a one-dimensional convolution module and an inverse wavelet transform module connected in sequence.

[0020] Preferably, the blood pressure detection model performs the detection process as follows:

[0021] Step 31: The HRV feature extraction unit extracts the RR interval from the IPPG signal through peak detection, and calculates the HRV feature sequence based on the RR interval;

[0022] Step 32: The RMS normalization layer normalizes the IPPG signal to obtain a normalized input signal;

[0023] Step 33: The first group of linear layers performs a linear transformation on the normalized input signal to obtain linear features; the linear layers of the first group of residual modules performs a linear transformation on the normalized input signal to obtain linear features, and then introduces nonlinearity into the linear features through an activation function to obtain first residual features;

[0024] Step 34: The linear features output by the first set of linear layers are sequentially subjected to feature extraction by two sets of transformation modules and partial feature discarding by the discard layer to obtain multi-scale convolutional features. The feature extraction process of each set of transformation modules is as follows:

[0025] The wavelet transform module in the one-dimensional wavelet transform convolution layer performs wavelet transform on the linear features, decomposes the linear features into different frequency sub-bands, and obtains multi-scale features; the one-dimensional convolution module performs one-dimensional convolution operation on the multi-scale features to extract the local features of each frequency sub-band; the inverse wavelet transform module reconstructs the original signal space according to the local features of different frequency sub-bands, realizes the extraction of multi-scale deep features of linear features, and obtains multi-scale deep features;

[0026] The batch normalization layer normalizes multi-scale deep features to obtain multi-scale high-dimensional normalized features, reducing internal covariate shift and improving model training stability;

[0027] The one-dimensional maximum pooling layer reduces the dimensionality of multi-scale high-dimensional normalized features with a set pooling window and step size, reducing the amount of data while retaining key features, reducing the model calculation amount and the risk of overfitting, and obtaining multi-scale convolutional features;

[0028] Step 35: Introduce nonlinearity to the multi-scale convolution feature through the activation function to obtain nonlinear features;

[0029] Step 36: The state-space model is selected to selectively process information based on nonlinear features by recursion, thereby focusing on relevant data and discarding unimportant information to obtain selected features;

[0030] Step 37: Multiply the selected feature and the first residual feature, and obtain the multiplication result;

[0031] Step 38: The second set of linear layers performs a linear transformation on the multiplication results to obtain multiplication linear features, and performs an addition operation on the IPPG signals transmitted by the second set of residual modules to obtain fusion features;

[0032] Step 39: The fused features are sequentially integrated through the RMS normalization layer, the dropout layer, the linear layer, and the global average pooling layer of the regression layer to obtain deep features.

[0033] Step 310: The feedforward neural network efficiently fuses the HRV feature sequence and the deep features to predict the systolic blood pressure (SBP) and the diastolic blood pressure (DBP) to obtain the blood pressure detection results.

[0034] Preferably, a camera with a resolution of not less than 1080p and a frame rate of not less than 30fps is used to shoot facial dynamic image data of the person being tested for 10 seconds at a distance of 0.1-0.8 meters in an environment with uniform and soft lighting, and in a quiet and stable state as a facial video sample or a facial video to be detected.

[0035] Preferably, the HRV feature sequence includes time domain features and frequency domain features of heart rate variability (HRV), including the number NN50 of adjacent normal heartbeat intervals with a difference greater than 50ms, the standard deviation SDNN of all normal sinus beat RR intervals, the root mean square RMSSD of adjacent RR interval differences, the proportion PNN50 of adjacent normal heartbeat intervals with a difference greater than 50ms, low-frequency power LF, high-frequency power HF, and the ratio of low-frequency to high-frequency power LF / HF. The HRV features reflect the state of the cardiovascular system.

[0036] Preferably, the specific process of step 31 is:

[0037] Step 311: For a time series IPPG signal s={s1,s2,...,s w} Perform first-order difference operation to obtain the difference sequence , w represents the length of the IPPG signal; the differential sequence that satisfies and The i-th position is marked as peak P i , retain the peak value to construct the peak index sequence {P1, P2, ..., P n}, n represents the sequence length of the peak index sequence, combined with the sampling frequency Calculate the RR interval; the RR interval is expressed as:

[0038] ;

[0039] represents the RR interval at the i-th position;

[0040] Step 312: Calculate the time domain characteristics of HRV, assuming is the extracted RR interval sequence, n is the sequence length, and the calculation is:

[0041] ;

[0042] ;

[0043] ;

[0044] ;

[0045] Step 313: Calculate the frequency domain characteristics of HRV and analyze the RR interval sequence. Perform Fourier transform to obtain the frequency sequence frequencies=[f1,f2,...,f n ], and calculate the power spectral density of each frequency in the frequency sequence to obtain the corresponding PSD sequence power=[p1,p2,...,p n ];

[0046] Filter out low-frequency band frequencies from the frequency sequence to build a low-frequency index set , f represents frequency, m represents the number of low-frequency band frequencies; the frequency range of the low-frequency band is , calculate the low-frequency power according to the PSD sequence position corresponding to the low-frequency index set ; represents the power spectral density; represents the power spectral density of the jth position in the PSD sequence, represents the power spectral density at position j+1 in the PSD sequence; Represents the frequency of the j+1th position in the low-frequency index set; represents the frequency of the jth position in the low-frequency index set;

[0047] Filter out high-frequency band frequencies from the frequency sequence to build a high-frequency index set , v represents the number of high-frequency band frequencies, and the frequency range of the high-frequency band is , calculate the high frequency power according to the PSD sequence position corresponding to the high frequency index set ;

[0048] Calculate the ratio of low frequency to high frequency power LF / HF: .

[0049] The non-contact blood pressure detection system based on deep learning includes:

[0050] Face video acquisition module, which collects face video samples and face videos to be detected;

[0051] The IPPG signal extraction module extracts frame-by-frame images from the face video sample and the face video to be detected, and calculates the corresponding image-based photoplethysmography (IPPG) signal;

[0052] The model building module builds the IPPG-Mamba network model and trains it based on the IPPG signals corresponding to the face video samples to obtain the blood pressure detection model;

[0053] The blood pressure detection module loads the blood pressure detection model to detect the IPPG signal corresponding to the face video to be detected and obtain the blood pressure detection result.

[0054] It can be seen from the above technical solution that compared with the existing technology, the present invention discloses a non-contact blood pressure detection method and system based on deep learning. By combining traditional HRV feature sequences (such as RR intervals, LF / HF, etc.) and deep-level features extracted by deep learning, it fully utilizes the time-frequency information of IPPG signals and effectively mines the key information related to blood pressure in IPPG signals. Compared with traditional non-contact detection models, it significantly improves the accuracy of blood pressure detection. Specifically, a CNN-Mamba Block with powerful feature extraction capabilities was designed. Combining the characteristics of CNN and Mamba networks, the WTConv1d wavelet transform convolution was introduced. This can perform detailed analysis of IPPG signals at different scales, effectively capturing subtle and critical feature changes in the signals, enhancing the model's multi-scale feature extraction capabilities for IPPG signals, and breaking through the limitations of the limited receptive field of traditional convolution. At the same time, multiple normalization, dimensionality reduction, and anti-overfitting operations within the module ensure the stability and efficiency of feature extraction. Non-contact convenience is improved. Non-contact detection is achieved by capturing facial videos with a camera, eliminating the reliance on professional operations of traditional contact detection, reducing patient discomfort, greatly improving detection convenience, and facilitating widespread application in scenarios such as home and mobile medical care. The environment is rationally standardized during the data collection process, and multiple advanced technologies such as HRV feature sequence extraction and CNN-Mamba Block are used in model construction to process signals. This can effectively cope with complex environmental conditions such as different lighting and postures, reduce the interference of environmental factors on the detection results, ensure stable and reliable blood pressure detection in diverse environments, and improve the environmental adaptability of non-contact blood pressure detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0056] Figure 1 A flow chart of the deep learning-based face video blood pressure detection method provided by the present invention;

[0057] Figure 2 A schematic diagram of the IPPG-Mamba network model structure provided by the present invention;

[0058] Figure 3 This is a schematic diagram of the CNN-Mamba Block structure provided by the present invention;

[0059] Figure 4 A schematic diagram of the one-dimensional wavelet transform convolutional layer structure provided by the present invention;

[0060] Figure 5 This is a schematic diagram of the feedforward neural network structure provided by the present invention. DETAILED DESCRIPTION

[0061] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0062] The embodiment of the present invention discloses a non-contact blood pressure detection method based on deep learning, such as Figure 1 As shown, the following steps are included:

[0063] S1: Collect face video samples, including face videos and blood pressure data;

[0064] S2: Extract frame-by-frame images from the face video, calculate the pixel mean in the image, and generate the IPPG signal;

[0065] S3: Construct an IPPG-Mamba network model and train it based on IPPG signals and blood pressure data to obtain a blood pressure detection model. The IPPG-Mamba network model includes an HRV feature extraction unit, a CNN-Mamba Block, a regression layer, and a feedforward neural network. The HRV feature extraction unit extracts the HRV (heart rate variability) feature sequence. The CNN-Mamba Block is an improved network module that combines a CNN network and wavelet transform convolution with the Mamba Block to extract deep features. The feedforward neural network fuses the HRV feature sequence and deep features to predict blood pressure detection results.

[0066] S4: Collect the face video to be detected, extract the frame-by-frame image, calculate the pixel mean in the image, and generate the IPPG signal to be detected;

[0067] S5: Input the IPPG signal to be detected into the blood pressure detection model for detection to obtain the blood pressure detection result.

[0068] In a specific embodiment, the specific process of performing blood pressure detection using the non-contact blood pressure detection method based on deep learning is as follows:

[0069] S1: Use a camera to shoot a 10-second face video and collect the blood pressure data of the subject;

[0070] A camera with a resolution of at least 1080p and a frame rate of at least 30fps should be used in a quiet and stable environment with even, soft lighting. The video should be shot for 10 seconds at a distance of 0.1-0.8 meters. During the recording, a blood pressure cuff should be used to measure the subject's blood pressure. After the measurement, the systolic blood pressure (SBP) and diastolic blood pressure (DBP) values ​​should be recorded and saved.

[0071] S2: Extract frame-by-frame images from the captured face video, calculate the pixel mean of each pixel in the image, and obtain a series of continuous values ​​as the IPPG signal;

[0072] S3: Construct IPPG-Mamba network model; the constructed IPPG-Mamba network model is as follows Figure 2 As shown, the IPPG-Mamba network model takes IPPG signals as input;

[0073] S4: Training the IPPG-Mamba network model:

[0074] Specifically, the extracted IPPG signals and the corresponding blood pressure data of the subjects were divided into 80% as the training set and 20% as the validation set. The Adam optimizer was used for training, with a fixed learning rate of 0.001, a batch size of 32, and 300 training rounds.

[0075] The HRV feature extraction unit of the IPPG-Mamba network model serves as an input branch. It extracts the RR interval (the time interval between adjacent peaks) from the IPPG signal through peak detection and, combined with the sampling frequency, calculates a series of HRV features, such as the number of adjacent normal heartbeat intervals with a difference greater than 50ms (NN50), the standard deviation (SDNN) of all normal sinus beat RR intervals, the root mean square (RMSSD) of the adjacent RR interval differences, the proportion (PNN50) of adjacent normal heartbeat intervals with a difference greater than 50ms (low-frequency power LF), the high-frequency power (HF), and the ratio of low-frequency to high-frequency power (LF / HF).

[0076] The core part of the other input branch is CNN-Mamba Block, which is improved from Mamba network. Figure 3As shown in the figure, it includes a root mean square normalization layer RMS Norm, two groups of linear layers Linear Layer, a CNN network module, an activation function SILU, a selection state space model SSM and two groups of residual modules; the first group of residual modules includes a linear layer Linear Layer and an activation function SILU; the CNN network module contains a unit composed of two groups of one-dimensional wavelet transform convolution layers WTConv1d, a batch normalization layer BatchNorm and a one-dimensional maximum pooling layer MaxPool1d, as well as a dropout layer Dropout 0.3, where the WTConv1d model module is as shown in the figure. Figure 4 As shown in the figure, it includes the wavelet transform module WT, the one-dimensional convolution module Conv1d and the inverse wavelet transform module IWT; after the IPPG signal is processed by the CNN-Mamba Block, the features will enter the regression layer, which includes RMS Norm (root mean square normalization), Dropout (random inactivation and discarding to prevent overfitting), Linear (linear layer) and other operations in sequence. Finally, the features are integrated through Global Average Pooling to obtain deep features.

[0077] The features extracted by the two input branches, namely HRV features and depth features, are finally input into the FNN (feedforward neural network). The structure of the feedforward neural network is as follows: Figure 5 As shown in the figure, through its internal linear layers (such as Linear 64, Linear 32, Linear 1) and activation function ELU (exponential linear unit), these features with different properties are fused. After processing by FNN, the model finally outputs SBP (systolic blood pressure) and DBP (diastolic blood pressure) values;

[0078] During training, the hyperparameters are adjusted based on the SBP and DBP predicted by the validation set and the network model, and the final trained model weights are saved as the blood pressure detection model;

[0079] S5: Use a camera to collect a 10-second facial video file of the person to be detected, then extract the IPPG signal from the facial video of the person to be detected in the facial video file, input the extracted IPPG signal into the blood pressure detection model, and output the blood pressure detection results including SBP and DBP.

[0080] In a specific embodiment, the process of extracting the RR interval from the IPPG signal by peak detection and calculating the HRV feature in combination with the sampling frequency is as follows:

[0081] S41: For IPPG signal s={s1,s2,...,s w} Perform first-order difference operation to obtain the difference sequence , w represents the length of the IPPG signal; the differential sequence that satisfies and The i-th position is marked as peak P i , retain the peak value to construct the peak index sequence {P1, P2, ..., P n}, n represents the sequence length of the peak index sequence, combined with the sampling frequency Calculate the RR interval; the RR interval is expressed as:

[0082] ;

[0083] represents the RR interval at the i-th position;

[0084] S42: Calculate the time domain characteristics of HRV, assuming is the extracted RR interval sequence, n is the sequence length, and the calculation is:

[0085] ;

[0086] ;

[0087] ;

[0088] ;

[0089] S43: Calculate the frequency domain characteristics of HRV and the RR interval sequence Perform Fourier transform to obtain the frequency sequence frequencies=[f1,f2,...,f n ], and calculate the power spectral density to get the corresponding PSD sequence power=[p1,p2,...,p n ];

[0090] Filter out low-frequency band frequencies from the frequency sequence to build a low-frequency index set , f represents frequency, m represents the number of low-frequency band frequencies; the frequency range of the low-frequency band is , calculate the low-frequency power according to the PSD sequence position corresponding to the low-frequency index set ; represents the power spectral density; represents the power spectral density of the jth position in the PSD sequence, represents the power spectral density at position j+1 in the PSD sequence; Represents the frequency of the j+1th position in the low-frequency index set; represents the frequency of the jth position in the low-frequency index set;

[0091] Filter out high-frequency band frequencies from the frequency sequence to build a high-frequency index set , v represents the number of high-frequency band frequencies, and the frequency range of the high-frequency band is , calculate the high frequency power according to the PSD sequence position corresponding to the high frequency index set ;

[0092] Calculate the ratio of low frequency to high frequency power LF / HF: .

[0093] On the other hand, in a specific embodiment, a non-contact blood pressure detection system based on deep learning includes:

[0094] Face video acquisition module, which collects face video samples and face videos to be detected;

[0095] The IPPG signal extraction module extracts frame-by-frame images from the face video sample and the face video to be detected, and calculates the corresponding IPPG signal;

[0096] The model building module builds the IPPG-Mamba network model and trains it based on the IPPG signals corresponding to the face video samples to obtain the blood pressure detection model;

[0097] The blood pressure detection module loads the blood pressure detection model to detect the IPPG signal corresponding to the face video to be detected and obtain the blood pressure detection result.

[0098] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0099] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A non-contact blood pressure detection method based on deep learning, characterized in that: The following steps are involved: Step 1: Collect face video samples, including face video and blood pressure data; Step 2: Extract frame-by-frame images from the face video and calculate the pixel mean in the image to generate the IPPG signal; Step 3: Construct an IPPG-Mamba network model and train it based on IPPG signals and blood pressure data to obtain a blood pressure detection model. The IPPG-Mamba network model includes an HRV feature extraction unit, a CNN-MambaBlock, a regression layer, and a feedforward neural network. The HRV feature extraction unit extracts the HRV feature sequence. The CNN-Mamba Block is an improved network module that combines the CNN network and the Mamba Block with wavelet transform convolution to extract fused features. The regression layer integrates the fused features to obtain deep features. The feedforward neural network fuses the HRV feature sequence and deep features to predict blood pressure detection results. Step 4: Collect the face video to be detected, extract the frame-by-frame image, and calculate the pixel mean in the image to generate the IPPG signal to be detected; Step 5: Input the IPPG signal to be detected into the blood pressure detection model for detection to obtain the blood pressure detection result; The CNN-Mamba Block includes a root mean square normalization layer, two sets of linear layers, a CNN network module, an activation function, a selected state space model, and two sets of residual modules. The CNN network module includes two sets of transformation modules and a dropout layer. Each set of transformation modules includes a one-dimensional wavelet transform convolution layer, a batch normalization layer, and a one-dimensional maximum pooling layer. The first set of residual modules includes a linear layer and an activation function. The blood pressure detection model performs the following detection process: Step 31: The HRV feature extraction unit extracts the RR interval from the IPPG signal through peak detection, and calculates the HRV feature sequence based on the RR interval; Step 32: The RMS normalization layer normalizes the IPPG signal to obtain a normalized input signal; Step 33: The first set of linear layers performs linear transformation on the normalized input signal to obtain linear features; The linear layer of the first group of residual modules performs linear transformation on the normalized input signal and introduces nonlinearity to obtain the first residual feature; Step 34: The linear features output by the first set of linear layers are passed through the CNN network module to obtain multi-scale convolution features, and nonlinearity is introduced through the activation function to obtain nonlinear features; The linear features output by the first set of linear layers are sequentially extracted by two sets of transformation modules and some features are discarded by the discard layer to obtain multi-scale convolutional features. The feature extraction process of each set of transformation modules is as follows: The wavelet transform module in the one-dimensional wavelet transform convolution layer performs wavelet transform on the linear features, decomposes the linear features into different frequency sub-bands, and obtains multi-scale features; the one-dimensional convolution module performs one-dimensional convolution operation on the multi-scale features to extract the local features of each frequency sub-band; The inverse wavelet transform module reconstructs the original signal space according to the local features of different frequency sub-bands, realizes the extraction of multi-scale deep features of linear features, and obtains multi-scale deep features; The batch normalization layer normalizes multi-scale deep features to obtain multi-scale high-dimensional normalized features, reducing internal covariate shift and improving model training stability; The one-dimensional maximum pooling layer reduces the dimensionality of multi-scale high-dimensional normalized features with a set pooling window and step size, reducing the amount of data while retaining key features, reducing the model calculation amount and the risk of overfitting, and obtaining multi-scale convolutional features; Step 35: The state space model is selected to obtain the selected feature by recursively processing the information selectively according to the nonlinear feature, and the selected feature is multiplied by the first residual feature to obtain the multiplication result; Step 36: The second set of linear layers performs a linear transformation on the multiplication results to obtain multiplication linear features, and performs an addition operation on the IPPG signals transmitted by the second set of residual modules to obtain fusion features, which are then integrated through the regression layer to obtain deep features; Step 37: The feedforward neural network fuses the HRV feature sequence and the deep features to predict systolic and diastolic blood pressure to obtain blood pressure test results; The specific process of step 31 is: Step 311: IPPG signal s={s1,s2,...,s w } Perform the first-order difference operation to obtain the difference sequence Δs={s2-s1, s3-s2, s w -s w-1 }, w represents the length of the IPPG signal; the i-th position in the differential sequence that satisfies Δs[i-1]>0 and Δs[i-1]< is marked as the peak P i , retain the peak value to construct the peak index sequence {P1, P2, ..., P n }, n represents the sequence length of the peak index sequence, combined with the sampling frequency f s Calculate the RR interval; the RR interval is expressed as: r i represents the RR interval at the i-th position; Step 312: Calculate the time domain characteristics and frequency domain characteristics of HRV; Calculate the frequency domain characteristics of HRV, and for the RR interval sequence rr_intervals = [r1, r2, ..., r n ] Perform Fourier transform to obtain the frequency sequence frequencies = [f1, f2, ..., fn], and calculate the power spectral density of each frequency in the frequency sequence to obtain the corresponding PSD sequence power = [p1, p2, ..., pn]; Filter out low-frequency band frequencies from the frequency sequence to construct a low-frequency index set I lf ={j:0.04≤f j <0.15, j=1,…,m}, f represents frequency, m represents the number of low-frequency band frequencies; the frequency range of the low-frequency band is 0.04≤f<0.15, and the low-frequency power is calculated according to the PSD sequence position corresponding to the low-frequency index set p represents the power spectral density; p j represents the power spectral density of the jth position in the PSD sequence, p j+1 represents the power spectral density of the j+1th position in the PSD sequence; f j+1 represents the frequency of the j+1th position in the low-frequency index set; f j represents the frequency of the jth position in the low-frequency index set; Filter out high-frequency band frequencies from the frequency sequence to build a high-frequency index set I hf ={j:0.15≤f j <0.4,j=1,…,v}, v represents the number of high frequency bands, the frequency range of the high frequency band is 0.15≤f<0.4, and the high frequency power is calculated according to the PSD sequence position corresponding to the high frequency index set. Calculate the ratio of low frequency to high frequency power LF / HF:

2. The non-contact blood pressure detection method based on deep learning according to claim 1, characterized in that: A feedforward neural network consists of a linear layer and an activation function.

3. The non-contact blood pressure detection method based on deep learning according to claim 1, characterized in that: The regression layers include RMS normalization, dropout, linear, and global average pooling layers.

4. The non-contact blood pressure detection method based on deep learning according to claim 1, characterized in that: Use a camera with a resolution of not less than 1080p and a frame rate of not less than 30fps, and shoot facial dynamic image data of the person being tested for 10 seconds at a distance of 0.1-0.8 meters as the face video sample or the face video to be detected.

5. The non-contact blood pressure detection method based on deep learning according to claim 1, characterized in that: The HRV characteristic sequence includes the time domain characteristics and frequency domain characteristics of HRV. The time domain characteristics include the number of adjacent normal heartbeat intervals with a difference greater than 50ms, the standard deviation of all normal sinus beat RR intervals, the root mean square of the differences between adjacent RR intervals, and the proportion of adjacent normal heartbeat intervals with a difference greater than 50ms. The frequency domain characteristics include low-frequency power, high-frequency power, and the ratio of low-frequency to high-frequency power.

6. A non-contact blood pressure detection system based on deep learning, characterized in that: The non-contact blood pressure detection method based on deep learning according to any one of claims 1 to 5 comprises: Face video acquisition module, which collects face video samples and face videos to be detected; The IPPG signal extraction module extracts frame-by-frame images from the face video sample and the face video to be detected, and calculates the corresponding IPPG signal; The model building module builds the IPPG-Mamba network model and trains it based on the IPPG signals corresponding to the face video samples to obtain the blood pressure detection model; The blood pressure detection module loads the blood pressure detection model to detect the IPPG signal corresponding to the face video to be detected and obtain the blood pressure detection result.

Images