Adenoid feature-based face volume comparison method and system

Abstract

The invention belongs to the technical field of image processing, and discloses a face volume comparison method and system based on adenoid features. Comprising the following steps: carrying out data acquisition on OSA disease related information of a child group to obtain corresponding face information; performing data preprocessing on the acquired face capacity information to obtain a corresponding initial face capacity image set and adenoid body evaluation index data; carrying out image feature extraction on the obtained initial face capacity image set to obtain a corresponding face capacity feature vector set; meanwhile, the data extraction module is also used for carrying out local feature extraction on the corresponding initial face capacity image set in combination with the obtained adenoid body evaluation data to obtain a corresponding face capacity representation vector set; constructing a corresponding feature recognition model based on the obtained face feature vector set and the face representation vector set, and performing face comparison on the to-be-compared children based on the feature recognition model to obtain a corresponding face comparison result; according to the invention, the quality of medical services can be improved.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and more specifically, to a facial comparison method and system based on adenoid features. Background Art

[0002] With the development of biometric technology, facial recognition technology has been widely used in security, finance, medical and other fields.

[0003] However, existing facial recognition technologies mainly rely on the overall features of the face, such as the position and shape of the eyes, nose, mouth, etc., while ignoring the features of the adenoids (such as earlobes, nose wings, etc.); and most facial recognition technologies are based on external facial features, such as eyes, nose and mouth, etc. for recognition; these methods ignore the internal structure of the face, especially the features of the adenoids, which may lead to limited recognition accuracy; and there are many limitations in the field of facial auxiliary diagnosis of OSA patients; in view of this, the present invention proposes a method to solve the above problems. Summary of the invention

[0004] In order to overcome the above-mentioned defects of the prior art and to achieve the above-mentioned purpose, the present invention provides the following technical solutions:

[0005] A facial comparison system based on adenoid features, comprising:

[0006] A data collection module, used to collect OSA disease-related information of a child group and obtain corresponding facial information, wherein the facial information includes a facial image set and facial medical data;

[0007] A data processing module is used to perform data preprocessing on the collected facial information to obtain a corresponding initial facial image set and adenoids evaluation index data;

[0008] A feature extraction module is used to extract image features from the obtained initial facial image set to obtain a corresponding facial feature vector set; at the same time, it is also used to extract local features from the corresponding initial facial image set in combination with the obtained adenoids evaluation data to obtain a corresponding facial representation vector set;

[0009] The facial comparison module is used to construct a corresponding feature recognition model based on the obtained facial feature vector set and facial representation vector set, and to perform facial comparison on the children to be compared based on the feature recognition model to obtain corresponding facial comparison results.

[0010] Furthermore, the process of collecting data on OSA disease-related information of the children group and obtaining corresponding facial information includes:

[0011] The data acquisition module is provided with a plurality of acquisition nodes; based on the acquisition nodes, facial images of a group of children aged 3-14 years who are diagnosed with OSA disease by standard polysomnography and a group of children aged 3-14 years who are not confirmed with OSA disease are respectively acquired to obtain a corresponding facial image set, wherein the facial image set is composed of facial images corresponding to the plurality of child groups;

[0012] At the same time, the collection node is also used to collect medical data corresponding to the group of children diagnosed with OSA disease to obtain corresponding facial medical data; summarize the obtained facial medical data and facial image collection to obtain corresponding facial information and store the data.

[0013] Furthermore, the process of performing data preprocessing on the collected facial information to obtain a corresponding initial facial image set and adenoids evaluation index data includes:

[0014] The data preprocessing module includes an image preprocessing unit and a medical preprocessing unit; the image preprocessing unit is used to perform image preprocessing on a facial image set in facial information, and the process of obtaining a corresponding initial facial image set includes:

[0015] Arbitrarily read a facial image and perform grayscale conversion on it to obtain a corresponding facial grayscale image, and obtain the grayscale value corresponding to each pixel in the corresponding facial grayscale image;

[0016] Decomposing the corresponding facial grayscale image by pixels to obtain high-frequency components and low-frequency components of the corresponding facial image in different decomposition directions;

[0017] Obtaining the high-frequency component corresponding to the corresponding facial image at the last decomposition, obtaining the grayscale value corresponding to each pixel point in the corresponding high-frequency component and calculating the average value, replacing the original pixel value in each pixel point in the corresponding high-frequency component based on the calculated average value, and obtaining the corresponding decomposed image; and recombining the high-frequency component and the low-frequency component in the pixel decomposition process and the decomposed image to obtain the corresponding enhanced grayscale image, and converting it into a color space to obtain the corresponding first facial image;

[0018] Identify the eye part in the corresponding first facial image based on the face detection algorithm, and mark the pixel points corresponding to the center of the left eye part and the right eye part as the left pixel point and the right pixel point respectively;

[0019] respectively obtaining pixel coordinates of a corresponding pixel left point and a pixel point in the first facial image, and constructing an image transformation matrix based thereon, performing image multiplication on the image transformation matrix and the corresponding first facial image, and marking the first facial image after the image multiplication as an enhanced facial image;

[0020] Performing face region recognition on the corresponding enhanced facial image, and performing region interception on the recognized face region to obtain the corresponding initial facial image;

[0021] Based on the acquisition process of the initial facial image, image preprocessing is performed on other facial images in the corresponding facial image set to obtain the corresponding initial facial image and summarize them to obtain the corresponding initial facial image set;

[0022] The medical preprocessing unit is used to preprocess the collected facial medical data to obtain the corresponding adenoids evaluation index, which includes:

[0023] Based on the collected medical facial data, the diagnostic indicators corresponding to the corresponding OSA diseases in different age groups within the corresponding child group are obtained, and the corresponding adenoids evaluation indicators are constructed based on the data; the adenoids evaluation indicators include lip-facial angle, nasal floor-maxillary sinus floor distance, SNA angle, SNB angle and other related indicators; among them, since the adenoids characteristics corresponding to the corresponding OSA diseases in different stages of the disease are different, therefore, there are some differences in the adenoids evaluation indicators corresponding to the OSA diseases in children of different age groups.

[0024] Furthermore, the formula for pixel decomposition of the corresponding facial grayscale image is: k+1 (m,n)=α×∑ (i,j)∈ Z1 h i-2m ×h j-2n ×C k (i, j); where α represents the normalization factor, which is a fixed constant and is used to adjust the output value of the corresponding decomposition process to ensure that the grayscale value of the pixel after decomposition maintains an appropriate dynamic range; h represents the pre-selected filter, h i-2m and h j-2n Respectively represent the application of the filter in the horizontal and vertical directions; C k (i, j) represents the image information corresponding to the pixel (i, j) in the original facial grayscale image at the kth decomposition; C k+1 (m,n) represents the pixel component (i.e., grayscale value) corresponding to the pixel point (m,n) after k+1 decompositions; k = 1, 2, ..., P; P represents the total number of decompositions; Z1 represents the corresponding facial image;

[0025] The image transformation matrix

[0026] Wherein, α = scale × cosδ; β = scale × sinδ; where scale represents a scale factor, which is used in the image transformation process to determine the enlargement or reduction scale of the corresponding first facial image; (x, y) represents the pixel coordinates of the pixel point in the corresponding first facial image; (x, y) ∈ Z2; Z2 represents the first facial image;

[0027] Where δ represents the rotation angle; (x_ zy ,y_ zy ) and (x_ yy ,y_ yy ) represent the pixel coordinates of the corresponding pixel left point and pixel point in the first facial image respectively.

[0028] Furthermore, the process of performing image feature extraction on the obtained initial facial image set to obtain a corresponding facial feature vector set includes:

[0029] Constructing a wavelet filter bank, and performing image decomposition on each initial facial image in the corresponding initial facial image set based on the constructed wavelet filter bank to obtain response images in different scales and directions; and constructing a corresponding image energy matrix based on the response images;

[0030] Flatten the obtained pixel energy matrix to obtain the corresponding first eigenvector;

[0031] At the same time, feature points of the corresponding initial facial image are extracted to obtain facial feature points corresponding to the corresponding initial facial image;

[0032] Obtain pixel coordinates of pixel points corresponding to the facial feature points in the corresponding initial facial image, and based on the pixel coordinates, obtain the Euclidean distances between different facial feature points, and based on the Euclidean distances, obtain the multi-dimensional geometric features of the corresponding initial facial image;

[0033] Respectively obtain all first feature vectors and multidimensional geometric features corresponding to the corresponding initial facial image; and classify the corresponding first feature vectors and aligned geometric features based on a pre-built data classifier, and obtain weight parameters assigned by the data classifier to the corresponding first feature vectors and aligned geometric features during the corresponding classification process;

[0034] Based on the weight parameter, the first feature vector and multidimensional set feature ranked in the top P1 are obtained respectively, and they are marked as weighted feature vector and weighted geometric feature respectively; P1 is a fixed constant;

[0035] Normalizing the corresponding weight feature vector and weight geometric feature, and constructing corresponding row vector features based on the normalized weight feature vector and weight geometric feature, respectively, and performing feature fusion to obtain the corresponding facial feature vector;

[0036] The facial feature vectors corresponding to all the initial facial images are counted to obtain the desired facial feature vector set.

[0037] Furthermore, the process of constructing a wavelet filter bank includes:

[0038] Select the Gaussian function as the window function and define the kernel function of the corresponding filter based on it Where ω represents the standard deviation of the Gaussian factor corresponding to the Gaussian function; a and b represent the horizontal and vertical coordinates of the pixel point in the response image respectively;

[0039] Then, the scale and direction of the corresponding kernel function are changed to obtain filters of different scales and directions;

[0040] Kernel function of the corresponding filter after scale change In the formula, c represents the translation parameter; τ represents the scale parameter, which controls the width of the Gaussian function.

[0041] The kernel function of the corresponding filter after the direction change ε and γ represent spatial frequency parameters;

[0042] Obtain filters corresponding to different scales and directions, and obtain corresponding wavelet filter groups based on them;

[0043] The formula for image decomposition is: In the formula, represents the pixel value at the pixel position z in the response image corresponding to the scale μ and direction v; z = (a, b); In the formula, k μ,v k represents the wave vector of the plane corresponding to the corresponding scale μ and direction v; max Indicates the maximum frequency of the corresponding filter; f μ Represents the spatial factor in the spatial domain where the scale μ is located; σ refers to the scale factor of the window function.

[0044] Furthermore, the process of extracting local features from the corresponding initial facial image set in combination with the obtained adenoids evaluation data to obtain the corresponding facial representation vector set includes:

[0045] Reading the obtained adenoids evaluation index; and obtaining facial feature points related to the adenoids face of the target child based on the obtained adenoids evaluation index;

[0046] Based on the process of obtaining the multidimensional geometric features and the first feature vector, a local feature vector and a local geometric feature corresponding to the corresponding local facial image are obtained;

[0047] The local texture features and the local geometric features are subjected to feature fusion to obtain a corresponding facial representation vector; the facial representation vector is used to represent the image features of the corresponding local facial image;

[0048] All the obtained facial representation vectors are counted to obtain a corresponding facial representation vector set.

[0049] Furthermore, the process of constructing a corresponding feature recognition model based on the obtained facial feature vector set and facial representation vector set includes:

[0050] Obtaining facial feature vectors and facial representation vectors corresponding to the corresponding initial facial images, and constructing corresponding feature triples based thereon, and then obtaining feature triples corresponding to all initial facial images, and constructing corresponding training data sets based thereon;

[0051] The basic architecture of the feature recognition model is defined as an improved convolutional neural network, wherein the basic structure of the improved convolutional neural network is an input layer, a convolution layer, a pooling layer and an output layer;

[0052] The input layer is used to receive the initial facial image in the training data and perform image scaling processing on the corresponding initial facial image so that it meets the requirements of the improved convolutional neural network;

[0053] The convolution layer is used to perform convolution processing on the initial facial image after the image scaling processing based on the deployed convolution kernel; the pooling layer is used to perform pooling processing on the output feature map corresponding to the convolution layer; the output layer integrates the extracted features after the pooling processing of a fully connected layer and performs effective output;

[0054] The loss function of the feature recognition model is defined as: Where N represents the total number of training samples in the training data set input into the feature recognition model; d is the index variable of the training sample; F(t) represents the output result of the feature recognition model of the tth training sample; Represents the training label of the tth training sample;

[0055] The formula for the convolution process is: In the formula, cs represents the number of layers; and Represent the output feature maps corresponding to the l0th and l1th channels in the cs-th layer and cs-1 layer respectively; f() represents the pre-selected activation function; Represents the connection weight between the l0th and l1th channels;

[0056] The formula for the pooling process is: In the formula, and They respectively represent the output feature maps corresponding to the l0th channel of the first layer; Indicates the bias within the corresponding layer; dowm() represents the downsampling function operation;

[0057] The corresponding feature recognition model is iteratively trained based on the training data set until the loss function of the corresponding feature recognition model tends to converge, and the model parameters are saved, thus completing the training of the feature recognition model.

[0058] Furthermore, the process of performing facial comparison on the child to be compared based on the feature recognition model and obtaining the corresponding facial comparison result includes:

[0059] Based on the collected facial information, the age of the corresponding child is obtained; and the facial representation vector and facial feature vector corresponding to the corresponding child are obtained by combining the facial representation vector set and the facial feature vector set; based on the age of the corresponding child, the corresponding facial representation vector and facial feature vector are arranged in time series;

[0060] Based on the facial representation vector and the facial feature vector arranged by the time series feature, a corresponding comparison is constructed to construct a corresponding tree bifurcation map, wherein the tree bifurcation map includes a first-level bifurcation point, a second-level bifurcation point and a third-level bifurcation point; the bifurcation point is used to store the time series information of the corresponding child; the second-level bifurcation point is used to store the facial representation vector; and the third-level bifurcation point is used to store the facial feature vector;

[0061] Then, a facial image of the child to be compared is obtained, and image preprocessing is performed on it to obtain an initial facial image of the corresponding child to be compared; and the initial facial image is input into the constructed feature recognition model to obtain a corresponding feature recognition result, wherein the feature recognition result includes a facial representation vector and a facial feature vector corresponding to the corresponding child to be compared;

[0062] At the same time, the age of the corresponding child to be compared is obtained, and based on the age, the corresponding feature recognition result is input into the first-level bifurcation point corresponding to the corresponding age and the adjacent age, and then the facial representation vector and facial feature vector in the corresponding feature recognition result are similarly calculated and sorted with the facial representation vector and facial feature vector stored in all the second-level bifurcation points and third-level bifurcation points connected to the corresponding first-level bifurcation point, and the second-level bifurcation point and third-level bifurcation point corresponding to the facial representation vector and facial feature vector with the highest similarity are selected as quasi-output nodes; whether the facial representation vector and facial feature vector in the corresponding quasi-output node are disease facial representation vectors and disease facial feature vectors is obtained; if not, it indicates that the corresponding child to be compared does not have abnormal disease characteristics related to OSA disease, and the output comparison result is normal; if so, it indicates that the corresponding child to be compared has abnormal disease characteristics related to OSA disease, and the output comparison result is abnormal;

[0063] At the same time, the facial representation vector and facial feature vector stored in the corresponding quasi-output node are fed back as data.

[0064] Further; a facial comparison method based on adenoid features includes:

[0065] Step 1: Collecting OSA disease-related information of the children group to obtain corresponding facial information, wherein the facial information includes a facial image set and facial medical data;

[0066] Step 2: preprocessing the collected facial information to obtain a corresponding initial facial image set and adenoids evaluation index data;

[0067] Step 3: extracting image features from the obtained initial facial image set to obtain a corresponding facial feature vector set; at the same time, extracting local features from the corresponding initial facial image set in combination with the obtained adenoids evaluation data to obtain a corresponding facial representation vector set;

[0068] Step 4: Construct a corresponding feature recognition model based on the obtained facial feature vector set and facial representation vector set, and perform facial comparison on the children to be compared based on the feature recognition model to obtain corresponding facial comparison results.

[0069] The technical effects and advantages of the present invention are as follows:

[0070] 1. It can arrange the facial representation vector and facial feature vector of the child in time series according to their age, and then construct a tree-like bifurcation map, which provides strong support for the early detection and intervention of OSA disease;

[0071] 2. Through the comprehensive processing of facial information and adenoids assessment data, the key features related to adenoids facial features can be extracted, and a feature recognition model can be constructed for facial comparison; compared with traditional diagnostic methods, this method can more accurately identify children with abnormal features related to OSA disease, reduce misdiagnosis and missed diagnosis, and thus improve the accuracy of diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 A schematic diagram of a facial comparison system based on adenoid features of the present invention;

[0073] Figure 2 It is a schematic diagram of a facial comparison method based on adenoid features of the present invention. DETAILED DESCRIPTION

[0074] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. 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 creative work are within the scope of protection of the present invention.

[0075] Example 1

[0076] See also Figure 1 As shown, this embodiment provides a facial comparison system based on adenoid features, including:

[0077] A data collection module is used to collect OSA disease-related information of the children group and obtain corresponding facial information, which includes facial image collection and facial medical data;

[0078] A data processing module is used to perform data preprocessing on the collected facial information to obtain a corresponding initial facial image set and adenoids evaluation index data;

[0079] A feature extraction module is used to extract image features from the obtained initial facial image set to obtain a corresponding facial feature vector set; at the same time, it is also used to extract local features from the corresponding initial facial image set in combination with the obtained adenoids evaluation data to obtain a corresponding facial representation vector set;

[0080] A facial comparison module is used to construct a corresponding feature recognition model based on the obtained facial feature vector set and facial representation vector set, and to perform facial comparison on the children to be compared based on the feature recognition model to obtain corresponding facial comparison results;

[0081] The modules are connected to each other via wired and / or wireless means to achieve data transmission between the modules.

[0082] It should be further explained that, in the specific implementation process, the process of collecting data related to OSA disease in children and obtaining corresponding facial information includes:

[0083] The data collection module is provided with a number of collection nodes, which are connected to the medical databases of multiple medical institutions based on network ports;

[0084] The collection node collects facial images of children aged 3-14 years who are diagnosed with OSA disease through standard polysomnography and children aged 3-14 years who are not diagnosed with OSA disease in the corresponding medical institutions based on the network port, and obtains a corresponding facial image set, which is composed of facial images corresponding to several groups of children, and the facial images include disease facial images and normal facial images; among which, the group of children aged 3-14 years who are diagnosed with OSA disease corresponds to disease facial images; the group of children aged 3-14 years who are not diagnosed with OSA disease corresponds to normal facial images;

[0085] At the same time, the collection node is used to collect medical data corresponding to the group of children diagnosed with OSA disease, and obtain corresponding facial medical data; summarize the obtained facial medical data and facial image collection, obtain corresponding facial information and store the data.

[0086] It should be further explained that, in the specific implementation process, the process of preprocessing the collected facial information to obtain the corresponding initial facial image set and adenoids evaluation index data includes:

[0087] The data preprocessing module includes an image preprocessing unit and a medical preprocessing unit; the image preprocessing unit is used to perform image preprocessing on the facial image set in the facial information to obtain a corresponding initial facial image set; the medical preprocessing unit is used to perform data preprocessing on the collected facial medical data to obtain corresponding adenoids evaluation indicators;

[0088] It should be further explained that, in a specific implementation process, the image preprocessing unit is used to perform image preprocessing on the facial image set in the facial information to obtain the corresponding initial facial image set, which includes:

[0089] Taking any facial image in the facial image set as an example, grayscale conversion is performed on the corresponding facial image to obtain a corresponding facial grayscale image, and at the same time, the grayscale value corresponding to each pixel in the corresponding facial grayscale image is obtained;

[0090] Then, the corresponding facial grayscale image is pixel-decomposed to obtain the high-frequency component and the low-frequency component of the corresponding facial image in different decomposition directions; the formula for pixel-decomposing the corresponding facial grayscale image is: In the formula, represents the normalization factor, which is a fixed constant used to adjust the output value of the corresponding decomposition process to ensure that the grayscale value of the pixel after decomposition maintains an appropriate dynamic range; h represents the pre-selected filter, h i-2m and h j-2n Respectively represent the application of the filter in the horizontal and vertical directions; C k (i, j) represents the image information corresponding to the pixel (i, j) in the original facial grayscale image at the kth decomposition; C k+1 (m,n) represents the pixel component (i.e., grayscale value) corresponding to the pixel point (m,n) after k+1 decompositions; wherein, there is a corresponding correlation between (i,j) and (m,n), which is determined by the filter and downsampling operation of the wavelet transform; it is used to ensure that the information of the original facial grayscale image can be effectively transferred to the new image while maintaining important structural information; k = 1, 2, ..., P; P represents the total number of decompositions; Z1 represents the corresponding facial image;

[0091] Obtain the high-frequency component corresponding to the last decomposition of the corresponding facial image, obtain the grayscale value corresponding to each pixel point in the corresponding high-frequency component and calculate the average value, replace the original pixel value in each pixel point in the corresponding high-frequency component based on the calculated average value, and obtain the corresponding decomposed image; then, reconstruct the high-frequency component and low-frequency component corresponding to the previous P-1 decompositions and the decomposed image to obtain the corresponding enhanced grayscale image, and convert it into the color space to obtain the corresponding first facial image;

[0092] Then, based on the face detection algorithm, the eye part in the corresponding first facial image is identified, and the pixel points corresponding to the center of the left eye part and the right eye part are marked as the left pixel point and the right pixel point respectively;

[0093] respectively obtaining pixel coordinates of the corresponding pixel left point and the pixel point in the first facial image, and constructing an image transformation matrix based thereon, performing image multiplication on the image transformation matrix and the corresponding first facial image, and marking the first facial image after the image multiplication as an enhanced facial image, wherein the image multiplication on the image transformation matrix and the corresponding first facial image serves to perform image correction on the face region in the corresponding first facial image;

[0094] Performing face region recognition on the corresponding enhanced facial image, and performing region interception on the recognized face region to obtain the corresponding initial facial image;

[0095] Based on the acquisition process of the initial facial image, image preprocessing is performed on other facial images in the corresponding facial image set to obtain the corresponding initial facial image and summarize them to obtain the corresponding initial facial image set;

[0096] Among them, the image transformation matrix Wherein, α = scale × cosδ; β = scale × sinδ; where scale represents a scale factor, which is used in the image transformation process to determine the enlargement or reduction scale of the corresponding first facial image; (x, y) represents the pixel coordinates of the pixel point in the corresponding first facial image; (x, y) ∈ Z2; Z2 represents the first facial image; Where δ represents the rotation angle; (x_ zy ,y_ zy ) and (x_ yy ,y_ yy ) represent the pixel coordinates of the corresponding pixel left point and pixel right point in the first facial image respectively;

[0097] It should be further explained that, in the specific implementation process, the medical preprocessing unit is used to preprocess the collected facial medical data, and the process of obtaining the corresponding adenoids evaluation index includes:

[0098] Based on the collected medical facial data, the diagnostic indicators corresponding to the corresponding OSA diseases in different age groups within the corresponding child group are obtained, and the corresponding adenoids evaluation indicators are constructed based on them; the adenoids evaluation indicators include lip-facial angle, nasal floor-maxillary sinus floor distance, SNA angle, SNB angle and other related indicators; among them, since the adenoids characteristics corresponding to the corresponding OSA diseases in different stages of the disease are different, therefore, there are some differences in the adenoids evaluation indicators corresponding to the OSA diseases in children of different age groups.

[0099] It should be further explained that, in a specific implementation process, the process of extracting image features from the obtained initial facial image set and obtaining a corresponding facial feature vector set includes:

[0100] Constructing a wavelet filter group, and performing image decomposition on each initial facial image in the corresponding initial facial image set based on the constructed wavelet filter group to obtain response images at different scales and directions; and constructing a corresponding image energy matrix based on the response image, wherein each matrix element of the image energy represents the pixel energy distribution of the corresponding response image at the corresponding scale and direction;

[0101] The obtained pixel energy matrix is ​​matrix flattened to obtain a corresponding first eigenvector; wherein the first eigenvector is used to represent the texture features of the corresponding initial facial image corresponding to the corresponding image at different scales and directions; matrix flattening refers to converting a multidimensional matrix into a one-dimensional array, and each element in the corresponding one-dimensional array is the corresponding first eigenvector;

[0102] Obtain a pre-built DLIB library, the DLIB library is used to provide a pre-trained model, and the pre-stored model is used to extract key points of facial features in the corresponding initial facial image; wherein, the construction process of the DLIB library is a prior art, and the present invention will not elaborate on it in detail;

[0103] Based on the pre-trained model provided by the DL IB library, feature points of the corresponding initial facial image are extracted to obtain the facial feature points corresponding to the corresponding initial facial image;

[0104] Obtain pixel coordinates of corresponding pixels of facial feature points in the corresponding initial facial image, and based on the pixel coordinates, obtain the Euclidean distances between different facial feature points, and based on the pixel coordinates, obtain the multidimensional geometric features of the corresponding initial facial image; the multidimensional geometric features refer to the Euclidean distances and angles between different facial feature points;

[0105] All first feature vectors and multidimensional geometric features corresponding to the corresponding initial facial image are obtained respectively; and the corresponding first feature vectors and the corresponding geometric features are classified based on the pre-constructed data classifier, and at the same time, the weight parameters assigned by the data classifier to the corresponding first feature vectors and the corresponding geometric features in the corresponding classification process are obtained; wherein the data classifier refers to a classification model constructed based on the XGboost algorithm, and the corresponding construction process is a prior art, and the present invention will not elaborate on it in detail;

[0106] Based on the weight parameter, the first feature vector and multidimensional set feature ranked in the top P1 are obtained respectively, and they are marked as weighted feature vector and weighted geometric feature respectively; P1 is a fixed constant;

[0107] Furthermore, the corresponding weight feature vectors and weighted geometric features are normalized, and corresponding row vector features are constructed based on the normalized weight feature vectors and weighted geometric features, and feature fusion is performed to obtain corresponding facial feature vectors for characterizing the corresponding initial facial image; wherein each initial facial image corresponds to a plurality of facial feature vectors;

[0108] Counting facial feature vectors corresponding to all initial facial images to obtain a desired facial feature vector set;

[0109] It should be further explained that, in the specific implementation process, the process of constructing the wavelet filter bank includes:

[0110] Select the Gaussian function as the window function and define the kernel function of the corresponding filter based on it Where ω represents the standard deviation of the Gaussian factor corresponding to the Gaussian function; a and b represent the horizontal and vertical coordinates of the pixel point in the response image respectively;

[0111] Then, the scale and direction of the corresponding filter kernel function are changed to obtain filters of different scales and directions;

[0112] Kernel function of the corresponding filter after scale change In the formula, c represents the translation parameter; τ represents the scale parameter, which controls the width of the Gaussian function;

[0113] The kernel function of the corresponding filter after the direction change ε and γ represent spatial frequency parameters;

[0114] Obtain filters corresponding to different scales and directions, and obtain corresponding wavelet filter groups based on them;

[0115] It should be further explained that, in the specific implementation process, the process of extracting local features from the corresponding initial facial image set in combination with the obtained adenoids evaluation data to obtain the corresponding facial representation vector set includes:

[0116] The obtained adenoids evaluation index is read; and based on the obtained adenoids evaluation index, facial feature points related to the adenoids face of the target child are obtained; it should be further explained that, in the specific implementation process, since the adenoids themselves are not part of the face, but lymphatic tissue located in the nasopharynx; therefore, the facial features of the adenoids are usually characterized by the correlation between the jaw, nose, lip and other facial parts of the corresponding child's face; for example: the SNA angle is the angle formed by the center of the sella turcica, the root of the nose and the upper alveolar seat; this angle reflects the front-back position relationship of the maxilla relative to the skull; therefore, the facial feature points to be obtained are the center of the sella turcica, the root of the nose and the upper alveolar seat;

[0117] Then, based on the facial feature points corresponding to the adenoids evaluation index, a corresponding initial facial image is partially intercepted to obtain a corresponding partial facial image;

[0118] Based on the process of obtaining the multidimensional geometric features and the first feature vector, the local feature vector and the local geometric features corresponding to the corresponding local facial image are obtained; it should be further explained that, in the specific implementation process, the local geometric features are determined by the adenoids evaluation index involved in the corresponding local facial image. For example, if the adenoids evaluation index corresponding to the local facial image is the SNA angle, the corresponding local geometric features are jointly determined by the Euclidean distance and the angle between the three facial feature points of the corresponding sella turcica center, the nasion point and the upper alveolar seat point;

[0119] Then, the local texture features and the local geometric features are fused to obtain a corresponding facial representation vector; the facial representation vector is used to represent the image features of the corresponding local facial image;

[0120] Counting all the obtained facial representation vectors to obtain a corresponding facial representation vector set;

[0121] It should be further explained that, in the specific implementation process, the facial representation vector includes a disease facial representation vector and a normal facial representation vector; the facial feature vector includes a disease facial feature vector and a normal facial feature vector.

[0122] It should be further explained that, in a specific implementation process, the process of constructing a corresponding feature recognition model based on the obtained facial feature vector set and facial representation vector set includes:

[0123] Obtaining facial feature vectors and facial representation vectors corresponding to the corresponding initial facial images, and constructing corresponding feature triples based thereon, and then obtaining feature triples corresponding to all initial facial images, and constructing corresponding training data sets based thereon;

[0124] The basic architecture of the feature recognition model is defined as an improved convolutional neural network, and the basic structure of the improved convolutional neural network is an input layer, a convolution layer, a pooling layer, and an output layer;

[0125] The input layer is used to receive the initial facial image in the training data and perform image scaling processing on the corresponding initial facial image so that it meets the requirements of the improved convolutional neural network;

[0126] The convolution layer is used to perform convolution processing on the initial facial image after image scaling based on the deployed convolution kernel; the pooling layer is used to perform pooling processing on the output feature map corresponding to the convolution layer; the output layer is used to synthesize the extracted features after the pooling processing of a fully connected layer and output them effectively;

[0127] The loss function of the feature recognition model is defined as: Where N represents the total number of training samples in the training data set input into the feature recognition model; d is the index variable of the training sample; F(t) represents the output result of the feature recognition model of the tth training sample; Represents the training label of the tth training sample;

[0128] The formula for convolution processing is: In the formula, cs represents the number of layers; and They represent the output feature maps corresponding to the l0th and l1th channels in the cs-th layer and cs-1 layer respectively; f() represents the pre-selected activation function, which is used to transform the linear fit; Represents the connection weight between the l0th and l1th channels;

[0129] The formula for pooling is: In the formula, and They respectively represent the output feature maps corresponding to the l1th channel of the first layer; Indicates the bias of the channel in the corresponding layer; dowm() represents the downsampling function operation;

[0130] The training data set is divided into P3 training data subsets, and the training data subsets are planned as sequence data according to the order of historical data collection time points; AdaGrad is defined as the optimizer to continuously optimize the parameters of the feature recognition model during the training process, and the sequence data is input into the feature recognition model in batches according to time, and the value of the corresponding loss function is recorded. When the value of the loss function of consecutive L batches no longer decreases or changes, the parameters of the feature recognition model at this time are saved, and the training of the feature recognition model is completed. Among them, L and P3 are fixed constants;

[0131] It should be further explained that, in the specific implementation process, the process of performing facial comparison on the children to be compared based on the feature recognition model and obtaining the corresponding facial comparison results includes:

[0132] Based on the collected facial information, the age of the corresponding child and the corresponding facial image are obtained; the facial representation vector set and the facial feature vector set are combined to obtain the facial representation vector and the facial feature vector corresponding to the corresponding child; and based on the age of the corresponding child, the corresponding facial representation vector and the facial feature vector are arranged in time series;

[0133] It should be further explained that when the corresponding facial feature vectors and facial representation vectors are arranged in time series, the age of the corresponding child is used as the first constraint condition. The younger the age, the earlier the time series of the corresponding facial representation vector and facial feature vector. If different facial representation vectors are in the same time period, they are classified.

[0134] Based on the facial representation vector and the facial feature vector arranged by the time series feature, a corresponding comparison is constructed to construct a corresponding tree bifurcation map, wherein the tree bifurcation map includes a first-level bifurcation point, a second-level bifurcation point and a third-level bifurcation point; and the bifurcation point is used for the time series information of the corresponding child (i.e., the child's age); the second-level bifurcation point is used to store the facial representation vector; and the third-level bifurcation point is used to store the facial feature vector;

[0135] Then, a facial image of the child to be compared is obtained, and image preprocessing is performed on it to obtain an initial facial image of the corresponding child to be compared; and the initial facial image is input into the constructed feature recognition model to obtain a corresponding feature recognition result, which includes a facial representation vector and a facial feature vector corresponding to the corresponding child to be compared;

[0136] At the same time, the age of the corresponding child to be compared is obtained, and based on the age, the corresponding feature recognition result is input into the first-level bifurcation point corresponding to the corresponding age and the adjacent age, and then the facial representation vector and facial feature vector in the corresponding feature recognition result are similarly calculated and sorted with the facial representation vector and facial feature vector stored in all the second-level bifurcation points and third-level bifurcation points connected to the corresponding first-level bifurcation point, and the second-level bifurcation point and third-level bifurcation point corresponding to the facial representation vector and facial feature vector with the highest similarity are selected as quasi-output nodes; whether the facial representation vector and facial feature vector in the corresponding quasi-output node are disease facial representation vectors and disease facial feature vectors is obtained; if not, it indicates that the corresponding child to be compared does not have abnormal disease characteristics related to OSA disease, and the output comparison result is normal; if so, it indicates that the corresponding child to be compared has abnormal disease characteristics related to OSA disease, and the output comparison result is abnormal;

[0137] At the same time, the facial representation vector and facial feature vector stored in the corresponding quasi-output node are fed back as data;

[0138] The present invention achieves efficient and accurate identification of OSA in children by integrating advanced data collection, processing, feature extraction and comparison technologies. This not only helps to improve the quality of medical services, but also may provide valuable data support for related research and clinical practice.

[0139] Example 2

[0140] See also Figure 2 As shown, the part not described in detail in this embodiment is described in Example 1, which provides a face comparison method based on adenoid features, including:

[0141] Step 1: Collect data related to OSA disease in children to obtain corresponding facial information, which includes facial image collection and facial medical data;

[0142] Step 2: preprocessing the collected facial information to obtain a corresponding initial facial image set and adenoids evaluation index data;

[0143] Step 3: extracting image features from the obtained initial facial image set to obtain a corresponding facial feature vector set; at the same time, extracting local features from the corresponding initial facial image set in combination with the obtained adenoids evaluation data to obtain a corresponding facial representation vector set;

[0144] Step 4: Construct a corresponding feature recognition model based on the obtained facial feature vector set and facial representation vector set, and perform facial comparison on the children to be compared based on the feature recognition model to obtain corresponding facial comparison results.

[0145] Example 3

[0146] The present embodiment discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the operation mode of the adenoid feature-based facial comparison method and system provided above is implemented.

[0147] Since the electronic device introduced in this embodiment is an electronic device used to implement a facial comparison method and system based on adenoid features in the embodiment of the present application, based on the facial comparison method and system based on adenoid features introduced in the embodiment of the present application, a person skilled in the art can understand the specific implementation of the electronic device of the present embodiment and its various variations, so how the electronic device implements the method in the embodiment of the present application is not described in detail here. As long as a person skilled in the art implements the electronic device used by a facial comparison method and system based on adenoid features in the embodiment of the present application, it belongs to the scope of protection of this application.

[0148] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters and thresholds in the formula are set by technicians in this field according to actual conditions.

[0149] The above are only preferred embodiments of the present invention. The protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technical users in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. A facial comparison system based on adenoid features, comprising: A data collection module, used to collect OSA disease-related information of a child group and obtain corresponding facial information, wherein the facial information includes a facial image set and facial medical data; A data processing module is used to perform data preprocessing on the collected facial information to obtain a corresponding initial facial image set and adenoids evaluation index data; A feature extraction module is used to extract image features from the obtained initial facial image set to obtain a corresponding facial feature vector set; at the same time, it is also used to extract local features from the corresponding initial facial image set in combination with the obtained adenoids evaluation data to obtain a corresponding facial representation vector set; The facial comparison module is used to construct a corresponding feature recognition model based on the obtained facial feature vector set and facial representation vector set, and to perform facial comparison on the children to be compared based on the feature recognition model to obtain corresponding facial comparison results.

2. The facial comparison system based on adenoid features according to claim 1, characterized in that: The process of collecting data related to OSA disease in children and obtaining corresponding facial information includes: The data acquisition module is provided with a plurality of acquisition nodes; based on the acquisition nodes, facial images of a group of children aged 3-14 years who are diagnosed with OSA disease by standard polysomnography and a group of children aged 3-14 years who are not confirmed with OSA disease are respectively acquired to obtain a corresponding facial image set, wherein the facial image set is composed of facial images corresponding to the plurality of child groups; At the same time, the collection node is also used to collect medical data corresponding to the group of children diagnosed with OSA disease to obtain corresponding facial medical data; summarize the obtained facial medical data and facial image collection to obtain corresponding facial information and store the data.

3. The facial comparison system based on adenoid features according to claim 2, characterized in that: The process of preprocessing the collected facial information to obtain a corresponding initial facial image set and adenoids evaluation index data includes: The data preprocessing module includes an image preprocessing unit and a medical preprocessing unit; the image preprocessing unit is used to perform image preprocessing on a facial image set in facial information, and the process of obtaining a corresponding initial facial image set includes: Arbitrarily read a facial image and perform grayscale conversion on it to obtain a corresponding facial grayscale image, and obtain the grayscale value corresponding to each pixel in the corresponding facial grayscale image; Decomposing the corresponding facial grayscale image by pixels to obtain high-frequency components and low-frequency components of the corresponding facial image in different decomposition directions; Obtaining the high-frequency component corresponding to the corresponding facial image at the last decomposition, obtaining the grayscale value corresponding to each pixel point in the corresponding high-frequency component and calculating the average value, replacing the original pixel value in each pixel point in the corresponding high-frequency component based on the calculated average value, and obtaining the corresponding decomposed image; and recombining the high-frequency component and the low-frequency component in the pixel decomposition process and the decomposed image to obtain the corresponding enhanced grayscale image, and converting it into a color space to obtain the corresponding first facial image; Identify the eye part in the corresponding first facial image based on the face detection algorithm, and mark the pixel points corresponding to the center of the left eye part and the right eye part as the left pixel point and the right pixel point respectively; respectively obtaining pixel coordinates of a corresponding pixel left point and a pixel point in the first facial image, and constructing an image transformation matrix based thereon, performing image multiplication on the image transformation matrix and the corresponding first facial image, and marking the first facial image after the image multiplication as an enhanced facial image; Performing face region recognition on the corresponding enhanced facial image, and performing region interception on the recognized face region to obtain the corresponding initial facial image; Based on the acquisition process of the initial facial image, image preprocessing is performed on other facial images in the corresponding facial image set to obtain the corresponding initial facial image and summarize them to obtain the corresponding initial facial image set; The medical preprocessing unit is used to preprocess the collected facial medical data to obtain the corresponding adenoids evaluation index, which includes: Based on the collected medical facial data, the diagnostic indicators corresponding to the corresponding OSA diseases in different age groups within the corresponding children's group are obtained, and the corresponding adenoids evaluation indicators are constructed based on them.

4. The facial comparison system based on adenoid features according to claim 3, characterized in that: The formula for pixel decomposition of the corresponding facial grayscale image is: In the formula, represents the normalization factor; h represents the pre-selected filter, h i-2m and h j-2n Respectively represent the application of the filter in the horizontal and vertical directions; C k (i, j) represents the image information corresponding to the pixel (i, j) in the original facial grayscale image at the kth decomposition; C k+1 (m,n) represents the pixel component corresponding to the pixel point (m,n) after k+1 decompositions; k = 1, 2, ..., P; P represents the total number of decompositions; Z1 represents the corresponding facial image; The image transformation matrix Wherein, α = scale × cosδ; β = scale × sinδ; where scale represents a scale factor; (x, y) represents the pixel coordinates of the corresponding pixel point in the first facial image; (x, y) ∈ Z2; Z2 represents the first facial image; Where δ represents the rotation angle; (x_ zy ,y_ zy ) and (x_ yy ,y_ yy ) represent the pixel coordinates of the corresponding pixel left point and pixel point in the first facial image respectively.

5. The facial comparison system based on adenoid features according to claim 3, characterized in that: The process of extracting image features from the obtained initial facial image set and obtaining a corresponding facial feature vector set includes: Constructing a wavelet filter bank, and performing image decomposition on each initial facial image in the corresponding initial facial image set based on the constructed wavelet filter bank to obtain response images in different scales and directions; and constructing a corresponding image energy matrix based on the response images; Flatten the obtained pixel energy matrix to obtain the corresponding first eigenvector; At the same time, feature points of the corresponding initial facial image are extracted to obtain facial feature points corresponding to the corresponding initial facial image; Obtain pixel coordinates of pixel points corresponding to the facial feature points in the corresponding initial facial image, and based on the pixel coordinates, obtain the Euclidean distances between different facial feature points, and based on the Euclidean distances, obtain the multi-dimensional geometric features of the corresponding initial facial image; Respectively obtain all first feature vectors and multidimensional geometric features corresponding to the corresponding initial facial image; and classify the corresponding first feature vectors and aligned geometric features based on a pre-built data classifier, and obtain weight parameters assigned by the data classifier to the corresponding first feature vectors and aligned geometric features during the corresponding classification process; Based on the weight parameter, the first feature vector and multidimensional set feature ranked in the top P1 are obtained respectively, and they are marked as weighted feature vector and weighted geometric feature respectively; P1 is a fixed constant; Normalizing the corresponding weight feature vector and weight geometric feature, and constructing corresponding row vector features based on the normalized weight feature vector and weight geometric feature, respectively, and performing feature fusion to obtain the corresponding facial feature vector; The facial feature vectors corresponding to all the initial facial images are counted to obtain the desired facial feature vector set.

6. The facial comparison system based on adenoid features according to claim 5, characterized in that: The process of constructing a wavelet filter bank includes: Select the Gaussian function as the window function and define the kernel function of the corresponding filter based on it Where ω represents the standard deviation of the Gaussian factor corresponding to the Gaussian function; a and b represent the horizontal and vertical coordinates of the pixel point in the response image respectively; Change the scale and direction of the corresponding kernel function to obtain filters of different scales and directions; Kernel function of the corresponding filter after scale change In the formula, c represents the translation parameter; τ represents the scale parameter; The kernel function of the corresponding filter after the direction change is ε and γ represent spatial frequency parameters; Obtain filters corresponding to different scales and directions, and obtain corresponding wavelet filter groups based on them; The formula for image decomposition is: In the formula, represents the pixel value at the pixel position z in the response image corresponding to the scale μ and direction v; z = (a, b); In the formula, k μ,v k represents the wave vector of the plane corresponding to the corresponding scale μ and direction v; max Indicates the maximum frequency of the corresponding filter; f μ represents the spatial factor in the spatial domain where the scale μ is located; σ refers to the scale factor of the window function; μ and v refer to the scale and direction corresponding to the corresponding filter.

7. The facial comparison system based on adenoid features according to claim 3, characterized in that: The process of extracting local features from the corresponding initial facial image set in combination with the obtained adenoids evaluation data to obtain the corresponding facial representation vector set includes: Reading the obtained adenoids evaluation index; and obtaining facial feature points related to the adenoids face of the target child based on the obtained adenoids evaluation index; Based on the process of obtaining the multidimensional geometric features and the first feature vector, a local feature vector and a local geometric feature corresponding to the corresponding local facial image are obtained; The local texture features and the local geometric features are subjected to feature fusion to obtain a corresponding facial representation vector; the facial representation vector is used to represent the image features of the corresponding local facial image; All the obtained facial representation vectors are counted to obtain a corresponding facial representation vector set.

8. The facial comparison system based on adenoid features according to claim 3, characterized in that: The process of constructing a corresponding feature recognition model based on the obtained facial feature vector set and facial representation vector set includes: Obtaining facial feature vectors and facial representation vectors corresponding to the corresponding initial facial images, and constructing corresponding feature triples based thereon, obtaining feature triples corresponding to all initial facial images, and constructing corresponding training data sets based thereon; The basic architecture of the feature recognition model is defined as an improved convolutional neural network, wherein the basic structure of the improved convolutional neural network is an input layer, a convolution layer, a pooling layer and an output layer; The input layer is used to receive the initial facial image in the training data and perform image scaling processing on the corresponding initial facial image; The convolution layer is used to perform convolution processing on the initial facial image after the image scaling processing based on the deployed convolution kernel; the pooling layer is used to perform pooling processing on the output feature map corresponding to the convolution layer; the output layer integrates the extracted features after the pooling processing of a fully connected layer and performs effective output; The loss function of the feature recognition model is defined as: Where N represents the total number of training samples in the training data set input into the feature recognition model; d is the index variable of the training sample; F(t) represents the output result of the feature recognition model of the tth training sample; Represents the training label of the tth training sample; The formula for the convolution process is: In the formula, cs represents the number of layers; and denote the output feature maps corresponding to the l0th and l1th channels in the cs-th layer and cs-1 layer respectively; f() denotes the pre-selected activation function; Represents the connection weight between the l0th and l1th channels; The formula for the pooling process is: In the formula, and They respectively represent the output feature maps corresponding to the l1th channel of the first layer; Indicates the bias within the corresponding layer; dowm() represents the downsampling function operation; The corresponding feature recognition model is iteratively trained based on the training data set until the loss function of the corresponding feature recognition model tends to converge, and the model parameters are saved, thus completing the training of the feature recognition model.

9. The facial comparison system based on adenoid features according to claim 8, characterized in that: The process of performing facial comparison on the child to be compared based on the feature recognition model and obtaining the corresponding facial comparison result includes: Based on the collected facial information, the age of the corresponding child is obtained; and the facial representation vector and facial feature vector corresponding to the corresponding child are obtained by combining the facial representation vector set and the facial feature vector set; based on the age of the corresponding child, the corresponding facial representation vector and facial feature vector are arranged in time series; Based on the facial representation vector and the facial feature vector arranged by the time series feature, a corresponding comparison is constructed to construct a corresponding tree bifurcation map, wherein the tree bifurcation map includes a first-level bifurcation point, a second-level bifurcation point and a third-level bifurcation point; the bifurcation point is used to store the time series information of the corresponding child; the second-level bifurcation point is used to store the facial representation vector; and the third-level bifurcation point is used to store the facial feature vector; Acquire a facial image of the child to be compared, perform image preprocessing on it, and obtain an initial facial image of the corresponding child to be compared; and input the image into the constructed feature recognition model to obtain a corresponding feature recognition result, wherein the feature recognition result includes a facial representation vector and a facial feature vector corresponding to the corresponding child to be compared; At the same time, the age of the corresponding child to be compared is obtained, and based on the age, the corresponding feature recognition result is input into the first-level bifurcation point corresponding to the corresponding age and the adjacent age, and the facial representation vector and facial feature vector in the corresponding feature recognition result are similar to the facial representation vector and facial feature vector stored in all the second-level bifurcation points and third-level bifurcation points connected to the corresponding first-level bifurcation point. The facial representation vector and facial feature vector stored in the third-level bifurcation points are selected as the quasi-output nodes; whether the facial representation vector and facial feature vector in the corresponding quasi-output node are disease facial representation vectors and disease facial feature vectors is obtained; if not, it indicates that the corresponding child to be compared does not have abnormal disease characteristics related to OSA disease, and the output comparison result is normal; if so, it indicates that the corresponding child to be compared has abnormal disease characteristics related to OSA disease, and the output comparison result is abnormal; At the same time, the facial representation vector and facial feature vector stored in the corresponding quasi-output node are fed back as data.

10. A facial comparison method based on adenoid features, which is implemented based on the facial comparison system based on adenoid features according to any one of claims 1 to 9, characterized in that: include: Step 1: Collecting OSA disease-related information of the children group to obtain corresponding facial information, wherein the facial information includes a facial image set and facial medical data; Step 2: preprocessing the collected facial information to obtain a corresponding initial facial image set and adenoids evaluation index data; Step 3: extracting image features from the obtained initial facial image set to obtain a corresponding facial feature vector set; at the same time, extracting local features from the corresponding initial facial image set in combination with the obtained adenoids evaluation data to obtain a corresponding facial representation vector set; Step 4: Construct a corresponding feature recognition model based on the obtained facial feature vector set and facial representation vector set, and perform facial comparison on the children to be compared based on the feature recognition model to obtain corresponding facial comparison results.

Images