WO2021120007A1 - 一种基于红外图像序列的睡眠质量评估系统和方法 - Google Patents

一种基于红外图像序列的睡眠质量评估系统和方法 Download PDF

Info

Publication number
WO2021120007A1
WO2021120007A1 PCT/CN2019/126038 CN2019126038W WO2021120007A1 WO 2021120007 A1 WO2021120007 A1 WO 2021120007A1 CN 2019126038 W CN2019126038 W CN 2019126038W WO 2021120007 A1 WO2021120007 A1 WO 2021120007A1
Authority
WO
WIPO (PCT)
Prior art keywords
infrared image
respiratory
tensor
image sequence
sleep quality
Prior art date
Application number
PCT/CN2019/126038
Other languages
English (en)
French (fr)
Inventor
王书强
游森榕
吴国宝
陆一乾
苗芬
张炽堂
Original Assignee
深圳先进技术研究院
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 深圳先进技术研究院 filed Critical 深圳先进技术研究院
Priority to PCT/CN2019/126038 priority Critical patent/WO2021120007A1/zh
Priority to US17/786,840 priority patent/US20230022206A1/en
Publication of WO2021120007A1 publication Critical patent/WO2021120007A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/0002—Inspection of images, e.g. flaw detection
    • G06T7/0012—Biomedical image inspection
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/0059—Measuring for diagnostic purposes; Identification of persons using light, e.g. diagnosis by transillumination, diascopy, fluorescence
    • A61B5/0082—Measuring for diagnostic purposes; Identification of persons using light, e.g. diagnosis by transillumination, diascopy, fluorescence adapted for particular medical purposes
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/08—Detecting, measuring or recording devices for evaluating the respiratory organs
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/48—Other medical applications
    • A61B5/4806—Sleep evaluation
    • A61B5/4815—Sleep quality
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
    • A61B5/7235—Details of waveform analysis
    • A61B5/7264—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems
    • A61B5/7267—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems involving training the classification device
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
    • G06V10/7753—Incorporation of unlabelled data, e.g. multiple instance learning [MIL]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/20—Movements or behaviour, e.g. gesture recognition
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/30—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for calculating health indices; for individual health risk assessment
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10016—Video; Image sequence
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10048—Infrared image
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20081—Training; Learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20084—Artificial neural networks [ANN]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30196—Human being; Person
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30232—Surveillance

Definitions

  • This application belongs to the field of deep learning technology, and in particular relates to a sleep quality assessment system and method based on infrared image sequences.
  • Sleep as a complex life behavior, is closely related to human health. Sleep occupies about one-third of human life. However, the pace of modern life is generally accelerating, the pressure of life and work continues to increase, and sleep loss is becoming more common, which seriously affects the quality of people's daily life and even their health. Therefore, it is of great significance to detect sleep quality, evaluate people's sleep status, and promptly guide treatment for poor sleep.
  • the subject needs to use the relevant equipment to monitor for about 8 hours in the hospital, which is expensive, time cost and price
  • the cost is relatively high; when the wearable device captures physiological data, when analyzing the quality of sleep, it needs to directly contact the human body, which brings inconvenience and psychological burden to the testee, interferes with the sleep process of the testee, and affects the testee’s sleep quality. Sleeping habits ultimately affect the accuracy of the assessment of the sleep quality of the subjects.
  • the first aspect of the embodiments of the present application provides a sleep quality assessment method based on an infrared image sequence, and the sleep quality assessment method includes:
  • the image sequence acquisition module is used to acquire multiple respiratory infrared image sequences to be evaluated, and one respiratory infrared image sequence to be evaluated includes multiple frames of respiratory infrared images to be evaluated;
  • the sleep quality evaluation module is configured to evaluate the sleep quality of each of the plurality of respiratory infrared image sequences to be evaluated by using a classifier to obtain the sleep quality corresponding to each of the respiratory infrared image sequences to be evaluated evaluation result;
  • the third aspect of the embodiments of the present application provides a terminal device, including a memory, a processor, and a computer program stored in the memory and running on the processor.
  • the processor executes the computer program, The steps of the sleep quality evaluation method as described in the first aspect above are implemented.
  • a fourth aspect of the embodiments of the present application provides a computer-readable storage medium, the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the sleep quality evaluation as described in the first aspect is implemented. Method steps.
  • the fifth aspect of the present application provides a computer program product, which when the computer program product runs on a terminal device, causes the terminal device to execute the steps of the sleep quality assessment method described in the first aspect.
  • this solution uses the infrared camera device to obtain multiple infrared image sequences of the user’s breathing to be evaluated during sleep, which can realize the non-contact sleep monitoring of the user and reduce the monitoring cost at the same time.
  • Evaluate sleep quality as a whole for each of the multiple respiratory infrared image sequences to be evaluated which can effectively retain the temporal and spatial continuity between the respiratory infrared images to be evaluated in each respiratory infrared image sequence to be evaluated sexual information, improve the accuracy of the sleep quality evaluation results for each breathing infrared image sequence to be evaluated, and use the sleep quality evaluation result that accounts for the most among multiple sleep quality evaluation results as the user’s sleep quality evaluation result, which improves The accuracy of the assessment of the user's sleep quality.
  • FIG. 2 is a schematic diagram of the implementation process of the method for evaluating sleep quality based on infrared image sequences provided in the second embodiment of the present application;
  • FIG. 4 is a schematic structural diagram of a sleep quality assessment system based on infrared image sequences provided in Embodiment 3 of the present application;
  • FIG. 5 is a schematic structural diagram of a terminal device provided in Embodiment 4 of the present application.
  • FIG. 1 is a schematic diagram of the implementation process of the sleep quality assessment method based on infrared image sequence provided in the first embodiment of the present application.
  • the sleep quality assessment method is applied to a terminal device. As shown in the figure, the sleep quality assessment method may include the following steps:
  • Step S101 Acquire multiple respiratory infrared image sequences to be evaluated, and one respiratory infrared image sequence to be evaluated includes multiple frames of respiratory infrared images to be evaluated.
  • multiple infrared image sequences of breathing to be evaluated can be acquired through an infrared camera device when the user is sleeping.
  • the infrared camera device can be integrated in the terminal device, and It can be independent of the terminal equipment (that is, the infrared camera is not integrated in the terminal equipment).
  • the connection and communication between the infrared camera and the terminal equipment can be established through wireless or wired methods, and the infrared camera can be connected to the terminal equipment.
  • the multiple respiratory infrared image sequences to be evaluated obtained by the camera device are transmitted to the terminal device.
  • the breathing infrared image to be evaluated may refer to the image taken by the infrared camera device of the user (ie, the user whose sleep quality is evaluated) in the nose and mouth area.
  • the infrared camera device does not need to be in contact with the user and can realize non-contact sleep monitoring of the user. , To avoid interference to the user's daily sleep, while reducing the cost of monitoring.
  • the continuous acquisition of multiple frames of respiratory infrared images to be evaluated can capture the temperature changes in the mouth and nose area when the user breathes, so as to extract features such as respiratory frequency and respiratory depth.
  • a breathing infrared image sequence to be evaluated may refer to a breathing infrared image sequence waiting to be evaluated for sleep quality.
  • the infrared camera device When the infrared camera device acquires multiple infrared image sequences of breathing to be evaluated while the user is sleeping, it can use a sliding window method to collect one infrared image sequence of breathing to be evaluated while the user is sleeping with a preset duration as a basic unit, for example, one minute as one Basic unit, continuous collection of five minutes of infrared images to be evaluated, multiple frames of infrared images to be evaluated within one minute compose a sequence of infrared images to be evaluated, five minutes includes five to one minute, that is, five minutes corresponds to five to be evaluated Breathing infrared image sequence.
  • a sliding window method to collect one infrared image sequence of breathing to be evaluated while the user is sleeping with a preset duration as a basic unit, for example, one minute as one Basic unit, continuous collection of five minutes of infrared images to be evaluated, multiple frames of infrared images to be evaluated within one minute compose a sequence of infrared images to be evaluated, five minutes includes five to
  • each respiratory infrared image sequence to be evaluated may also be preprocessed.
  • the preprocessing includes, but is not limited to, adding each respiratory infrared image sequence to be evaluated Adjust the size of the multiple frames of respiratory infrared images to be evaluated in the sequence to be the same (for example, adjust to a preset size) and/or adjust the pixel values of the multiple frames of respiratory infrared images to be evaluated in each respiratory infrared image sequence to be evaluated to the preset size. Set within the range.
  • Step S102 Perform sleep quality assessment on each of the plurality of respiratory infrared image sequences to be evaluated by a classifier to obtain sleep quality assessment results corresponding to each of the respiratory infrared image sequences to be evaluated.
  • the above-mentioned classifier can evaluate a breathing infrared image sequence to be evaluated as a whole for sleep quality evaluation, that is, the above-mentioned classifier can directly evaluate a breathing infrared image sequence to be evaluated for sleep quality, and can effectively retain a breathing infrared image sequence to be evaluated.
  • the spatio-temporal continuity information between multiple frames of respiratory infrared images to be evaluated in the sequence improves the accuracy of the sleep quality evaluation result of a respiratory infrared image sequence to be evaluated.
  • the sleep quality evaluation result is used to indicate whether the sleep quality is good or bad.
  • the sleep quality evaluation result includes but is not limited to the first sleep quality evaluation result and the second sleep quality evaluation result.
  • the first sleep quality evaluation result can mean that the sleep quality is good. 2.
  • the sleep quality assessment result may refer to poor sleep quality, that is, the sleep quality indicated by the first sleep quality assessment result is better than the sleep quality indicated by the second sleep quality assessment result.
  • the content of the sleep quality evaluation results can also be re-divided according to actual needs.
  • the sleep quality evaluation results can be good sleep quality, good sleep quality, poor sleep quality, etc., which are not limited here.
  • the classifier performs sleep quality assessment on a breathing infrared image sequence to be evaluated is to classify a breathing infrared image sequence to be evaluated.
  • the classification category is the sleep quality assessment result.
  • the classification categories include good sleep quality and If the sleep quality is bad, a classifier is used to determine whether the category of a breathing infrared image sequence to be assessed belongs to good sleep quality or bad sleep quality.
  • the sleep quality assessment when the sleep quality assessment is performed on each of the multiple breathing infrared image sequences to be evaluated by the classifier, it may be obtained by classifying the multiple breathing infrared image sequences to be evaluated.
  • the device separately evaluates the sleep quality of multiple breathing infrared image sequences to be evaluated; it can also evaluate the sleep quality of the first breathing infrared image sequence to be evaluated through the classifier when the first breathing infrared image sequence to be evaluated is obtained , Obtain the sleep quality assessment result of the first respiratory infrared image sequence to be evaluated (that is, the sleep quality assessment result corresponding to the first respiratory infrared image sequence to be assessed), and when the second respiratory infrared image sequence to be evaluated is obtained, pass The classifier evaluates the sleep quality of the second breathing infrared image sequence to be evaluated, and obtains the sleep quality evaluation result of the second breathing infrared image sequence to be evaluated, and so on, until the last one of the multiple breathing infrared image sequences to be evaluated is obtained.
  • the sleep quality assessment is performed on each of the multiple breathing infrared image sequences to be evaluated by the classifier to obtain the sleep quality assessment corresponding to each breathing infrared image sequence to be evaluated
  • the results include:
  • the target feature map and the fully connected layer based on tensor decomposition in the classifier perform sleep quality assessment on each of the breathing infrared image sequences to be evaluated, and obtain the corresponding to each breathing infrared image sequence to be evaluated Results of sleep quality assessment.
  • the classifier when it receives a respiratory infrared image sequence to be evaluated, it can use the second-order pooling block to use the second-order information of the respiratory infrared image sequence to be evaluated to automatically extract and Sleep-related breathing characteristics improve the accuracy of the assessment of the sleep quality of the breathing infrared image sequence to be assessed.
  • the network layer based on tensor decomposition includes a convolutional layer based on tensor decomposition and two densely connected blocks based on tensor decomposition.
  • the classifier uses the densely connected mechanism of densely connected blocks to effectively solve the problem of gradient disappearance.
  • the densely connected blocks in the classifier may refer to a dense convolutional neural network, such as a residual network.
  • a densely connected block usually includes multiple convolutional layers.
  • the convolution layer based on tensor decomposition refers to the tensor decomposition of the convolution kernel of the convolution layer.
  • the convolution layer based on tensor decomposition is a 3D convolution layer, and the convolution kernel of the 3D convolution layer is tensor decomposed.
  • the convolution kernel of the 3D convolution layer can be decomposed into the product of two matrices and a three-dimensional tensor.
  • the first and third orders are matrices, and the second order is a three-dimensional tensor.
  • the fully connected layer based on tensor decomposition refers to the tensor decomposition of the weight of the fully connected layer.
  • the weight of the fully connected layer can be decomposed into the product of two matrices and a three-dimensional tensor.
  • the first and third orders are matrices.
  • the second order is a three-dimensional tensor.
  • the breathing infrared image sequence to be evaluated is first taken as a tensor as a whole, and convolved with the convolution kernel after tensor decomposition in the convolutional layer Calculate, and then perform convolution calculation with the convolution kernel after tensor decomposition in a densely connected block, and then pass the second-order pooling block, and then perform convolution calculation with the convolution sum after tensor decomposition in a densely connected block, Finally, the weighted fully connected layer with tensor decomposition is used to obtain the sleep quality evaluation result of the breathing infrared image sequence to be evaluated.
  • the target feature map is the feature map output by the last densely connected block in the classifier. It is a high-order feature map generated after multi-layer convolution, because the number of feature maps output by the last densely connected block in the classifier is multiple. Therefore, it can be called a high-order feature map.
  • the network layer and the fully connected layer in the classifier are represented by tensor decomposition, which can reduce the number of parameters in the classifier, and solve the loss of internal structure information of the tensor data and the number of parameters in the classifier caused by the vectorization calculation of tensor form data.
  • the aforementioned tensor decomposition may refer to Tensor-Train tensor decomposition.
  • Step S103 Count the number of different sleep quality evaluation results according to the sleep quality evaluation results corresponding to the plurality of breathing infrared image sequences to be evaluated, and determine the sleep quality evaluation result with the largest number as the sleep quality evaluation result of the user.
  • one breathing infrared image sequence to be evaluated corresponds to one sleep quality assessment result
  • multiple breathing infrared image sequences to be evaluated correspond to multiple sleep quality assessment results
  • the same sleep quality assessment results may exist in the multiple sleep quality assessment results.
  • the sleep quality of five respiratory infrared image sequences to be evaluated, the first respiratory infrared image sequence to be evaluated, the second respiratory infrared image sequence to be evaluated, and the fifth respiratory infrared image sequence to be evaluated are exemplified by the classifier.
  • the evaluation results are all good sleep quality.
  • the sleep quality evaluation results of the third breathing infrared image sequence to be evaluated and the fourth breathing infrared image sequence to be evaluated are bad sleep quality, and three of the five sleep quality evaluation results are counted. Good sleep quality and two bad sleep quality, the number of good sleep quality is the largest, then it can be determined that the user’s sleep quality evaluation result is good sleep quality.
  • the infrared camera device obtains the user’s breathing infrared image sequence to be evaluated during sleep, and uses a tensor decomposition-based classifier to evaluate the sleep quality of each breathing infrared image sequence to be evaluated as a tensor as a whole.
  • a contactless sleep quality evaluation can be realized, and the accuracy of the evaluation of the user's sleep quality can be improved.
  • FIG. 2 is a schematic diagram of the implementation process of the sleep quality assessment method based on infrared image sequence provided in the second embodiment of the present application.
  • the sleep quality assessment method is applied to a terminal device. As shown in the figure, the sleep quality assessment method may include the following steps:
  • step S201 the classifier is trained through Zhang's quantified ternary generative confrontation network.
  • the quantified ternary generative confrontation network includes a generator, a classifier, and a discriminator.
  • the generator, classifier, and discriminator all use tensor decomposition, which effectively reduces the number of generators, classifiers, and discriminators.
  • the number of parameters, and the overall processing of the respiratory infrared image sequence can be realized.
  • Figure 3a shows an example of the structure of Zhang Quan’s ternary generative confrontation network.
  • G represents the generator
  • C represents the classifier
  • D represents the discrimination.
  • the unmarked respiratory infrared image sequence X c is the respiratory infrared image sequence without the tag
  • the marked infrared image sequence (X l , Y l ) is the respiratory infrared image sequence with the tag
  • Y c is the tag of the unmarked infrared image sequence
  • X g is the respiratory infrared image sequence generated by the generator.
  • the tensor decomposition algorithm Tensor-Train decomposes a d-order tensor, which can be expressed as the product of two matrices and d-2 three-dimensional tensors.
  • the first and dth are matrices, and the remaining d-2 are Three-dimensional tensor, d is an integer greater than 2.
  • the training of the classifier through the quantified ternary generative confrontation network includes:
  • the second respiratory infrared image sequence is input to the classifier, and the third respiratory infrared image sequence is obtained through the second-order pooling block, the network layer based on tensor decomposition, and the fully connected layer in the classifier.
  • the third breath infrared image sequence refers to the second breath infrared image sequence carrying the tag;
  • the discriminator is trained, and the discriminator's response to the third respiratory infrared image sequence is obtained. Judgment result;
  • the generator adopts the idea of conditional generation against the network, taking one-dimensional random noise that obeys a normal distribution as the input of the generator, and taking sleep quality as the target label as the conditional input, and the intermediate network layer adopts 3D reaction Convolutional layer, and then use Leaky ReLU as the activation function, and Batch Norm for batch regularization.
  • the last 3D deconvolution layer of the generator is followed by a tanh activation layer, and the Tensor-Train tensor decomposition auxiliary generator is used to generate and carry
  • the breathing infrared image sequence of the sleep quality label reduces the need for a real breathing infrared image sequence carrying the label.
  • Figure 3b shows an example of the structure of the generator.
  • the one-dimensional random noise carrying the target label is sequentially deconvolved through the 3D deconvolution layer, the Leaky ReLU activation function and batch regularization to perform the gradual deconvolution of the feature map, which can generate the first breath infrared image that is close to the real one carrying the target label. sequence.
  • D(x g ,y g ) represents the discrimination result of the discriminator, if the discrimination result is true, then D(x g ,y g ) is 1, if the discrimination result is false, then D(x g ,y g ) Is 0;
  • ⁇ is the weight parameter (users can set it according to actual needs);
  • x label is the real respiratory infrared image sequence, and
  • x g is the generated respiratory infrared image sequence (ie, the first respiratory infrared image sequence), It represents the L 1 loss between the real breathing infrared image sequence and the generated breathing infrared image sequence, so that the generated breathing infrared image sequence is more close to the
  • the unlabeled respiratory infrared image sequence collected by the infrared camera into two parts: one part is used as the second respiratory infrared image sequence, the sleep quality is evaluated by the classifier, and the sleep quality evaluation result output by the classifier is used as the label , So as to obtain the third breath infrared image sequence; the other part uses sleep experts to evaluate the sleep quality, and use the sleep quality assessment result as a label to obtain the fourth breath infrared image sequence carrying the label; the first breath infrared image sequence, the first breath The three-breath infrared image sequence and the fourth breath infrared image sequence are respectively input to the discriminator, and the discriminator’s discrimination results of the first, third and fourth breath infrared image sequences are obtained.
  • the discriminant results of a sequence of breathing infrared images obtain the loss function of the discriminator, and train the discriminator according to the loss function.
  • the loss function Loss l when the breathing infrared image sequence trains the classifier, and the Loss g and Loss l are integrated into Loss supervised Loss g + ⁇ Loss l , and the classification is obtained according to the discriminator’s discriminating result of the third breathing infrared image sequence
  • Zhang Quan’s ternary generation confrontation network uses a large number of second respiratory infrared image sequences (ie, respiratory infrared image sequences that do not carry tags) and a small number of fourth respiratory infrared image sequences (ie, tag-bearing infrared image sequences). Breathing infrared image sequence) training the ternary generation of Zhang quantified against the network can solve the problem of less data in the respiratory infrared image sequence with tags, and at the same time make full use of the respiratory infrared image sequence without tags, which is conducive to improving the quantification of Zhang. The robustness of the ternary generation against the network.
  • the obtaining the first respiratory infrared image sequence carrying the target label through the deconvolution layer based on tensor decomposition in the generator includes:
  • the deconvolution kernel of the 3D deconvolution layer in the generator after the deconvolution kernel of the 3D deconvolution layer in the generator is decomposed into a tensor, it can be decomposed into the product of two matrices and a three-dimensional tensor (where the first order and the third order Is a matrix, the second order is a three-dimensional tensor), and then the input one-dimensional random noise and the first tensor (that is, the product of two matrices and a three-dimensional tensor) are subjected to multi-layer deconvolution calculations, and then through the graph
  • the activation function Leaky ReLU, batch regularization, and tanh activation layer in the structural example shown in 3b generate a sequence of first breath infrared images that are close to real and carry target tags.
  • the network layer based on tensor decomposition in the discriminator includes a convolutional layer based on tensor decomposition, a first dense connection block, and a second dense connection block;
  • obtaining the discrimination result of the first respiratory infrared image sequence by the discriminator includes:
  • Tensor decomposition is performed on the convolution kernel of the first densely connected block in the discriminator to obtain a third tensor
  • Tensor decomposition is performed on the convolution kernel of the second densely connected block in the discriminator to obtain a fourth tensor
  • the first feature map, the second feature map, and the third feature map are all feature maps of the first respiratory infrared image sequence.
  • the tensor decomposition of the convolution kernel of the 3D convolution layer in the discriminator after the tensor decomposition of the convolution kernel of the 3D convolution layer in the discriminator, it can be decomposed into the product of two matrices and a three-dimensional tensor (where the first order and the third order are matrix , The second order is a three-dimensional tensor), the product of the two matrices and a three-dimensional tensor obtained after decomposition is the second tensor, and the first breathing infrared image sequence is calculated by multi-layer convolution through the second tensor
  • the resulting feature map is the first feature map; after tensor decomposition of the convolution kernel of the first densely connected block in the discriminator, it can be decomposed into the product of two matrices and a three-dimensional tensor (where the first order and The third order is a matrix, the second order is a three-dimensional tensor), the product of the
  • the discriminator further includes a first transition layer and a second transition layer, the first transition layer and the second transition layer are both 1 ⁇ 1 ⁇ 1 convolution kernels; the first transition layer The layer is located between the first densely connected block and the second densely connected block in the discriminator, and is used to reduce the number of the second feature maps; the second transition layer is located in the second densely connected block in the discriminator Between it and the fully connected layer, it is used to reduce the number of the third feature maps.
  • the number of second feature maps obtained increases, and the convolution calculation is performed through the 1 ⁇ 1 ⁇ 1 3D convolution kernel, which can reduce The number of second feature maps is to reduce the number of channels.
  • the number of third feature maps obtained increases, and the convolution calculation is performed through the 1 ⁇ 1 ⁇ 1 3D convolution kernel, which can reduce the number of third feature maps. That is, reduce the number of channels.
  • the tag-carrying infrared image sequence of breathing input to the discriminator (can be any tag-carrying infrared image sequence input to the discriminator) is convolution based on tensor decomposition Layer (ie 3D quantized convolutional layer), densely connected block based on tensor decomposition (ie 3D quantized densely connected block), transition layer and deep neural network composed of fully connected layer based on tensor decomposition, for breathing infrared images
  • the sequence performs feature extraction to obtain the respiratory feature information that retains the respiratory infrared image sequence in time and space.
  • the extracted respiratory feature information is passed through a fully connected layer based on tensor decomposition to realize the trueness of the respiratory infrared image sequence with tags. False discrimination.
  • the discriminator also includes Leaky ReLU activation function, Batch Norm regularization and sigmoid function.
  • the embodiment of the present application uses densely connected blocks to directly input the feature maps extracted by the network layer before the densely connected blocks to the subsequent network layers for cascading, reducing the feature loss in the gradient transfer process, and solving the deep neural network reverse
  • the problem of the disappearance of gradients in the propagation process stabilizes the training of the confrontation generation network, and improves the discriminator's performance in discriminating between the generated samples and the real samples.
  • the network layer based on tensor decomposition in the classifier includes a convolutional layer based on tensor decomposition, a third densely connected block, and a fourth densely connected block;
  • the second-order pooling in the classifier Block, network layer and fully connected layer based on tensor decomposition, to obtain the third infrared absorption image sequence includes:
  • Tensor decomposition is performed on the convolution kernel of the third densely connected block in the classifier to obtain a seventh tensor
  • Tensor decomposition is performed on the convolution kernel of the fourth densely connected block in the classifier to obtain an eighth tensor
  • the fourth feature map, the fifth feature map, the sixth feature map, and the seventh feature map are all feature maps of the second respiratory infrared image sequence.
  • the convolution kernel of the 3D convolution layer in the classifier can be decomposed into the product of two matrices and a three-dimensional tensor (where the first order and the third order are matrix ,
  • the second order is a three-dimensional tensor
  • the product of the two matrices and a three-dimensional tensor obtained after decomposition is the sixth tensor
  • the second breathing infrared image sequence is calculated by multi-layer convolution through the sixth tensor
  • the resulting feature map is the fourth feature map; after tensor decomposition of the convolution kernel of the third densely connected block in the classifier, it can be decomposed into the product of two matrices and a three-dimensional tensor (where the first order and The third order is a matrix, the second order is a three-dimensional tensor), the product of the two matrices and a three-dimensional tensor obtained after decomposition is the seventh tensor, and the fourth feature
  • the classifier further includes a third transition layer and a fourth transition layer, the third transition layer and the fourth transition layer are both 1 ⁇ 1 ⁇ 1 convolution kernels; the third transition layer The layer is located between the second-order pooling block and the fourth densely connected block in the classifier, and is used to reduce the number of the sixth feature maps; the fourth transition layer is located in the fourth densely connected block in the classifier Between it and the fully connected layer, it is used to reduce the number of the seventh feature maps.
  • the number of fifth feature maps obtained increases, which in turn leads to a larger number of sixth feature maps processed by the second-level pooling block.
  • the number of sixth feature maps can be reduced, that is, the number of channels can be reduced.
  • the number of seventh feature maps obtained increases.
  • the convolution calculation is performed through the 1 ⁇ 1 ⁇ 1 3D convolution kernel, which can reduce the number of seventh feature maps. That is, reduce the number of channels.
  • Figure 3d shows an example of the structure of the classifier.
  • the respiratory infrared image sequence input to the classifier passes through a convolutional layer based on tensor decomposition (i.e. 3D quantized convolutional layer) and densely connected blocks based on tensor decomposition (i.e. 3D quantization dense connection block) performs multi-layer convolution calculation, and reduces the number of feature maps through the transition layer, and then through a fully connected layer based on tensor decomposition, the probability value of the corresponding category can be obtained, thereby obtaining the breathing infrared image Sequence of sleep quality assessment results.
  • the classifier also includes Leaky ReLU activation function, Batch Norm regularization and sigmoid function.
  • the embodiment of the present application uses the self-attention mechanism based on second-order pooling densely connected blocks to weight the feature map in the dimension of the feature channel according to the correlation of different regions, so that the important channel weight is significant and not important.
  • the channel weight is small, which effectively extracts the respiratory features in the infrared image sequence, and improves the accuracy of the assessment of sleep quality.
  • Step S202 Acquire multiple respiratory infrared image sequences to be evaluated, and one respiratory infrared image sequence to be evaluated includes multiple frames of respiratory infrared images to be evaluated.
  • step S101 This step is the same as step S101.
  • step S101 For details, please refer to the related description of step S101, which will not be repeated here.
  • Step S203 Perform sleep quality assessment on each of the plurality of respiratory infrared image sequences to be evaluated by a classifier to obtain sleep quality assessment results corresponding to each of the respiratory infrared image sequences to be evaluated.
  • step S102 This step is the same as step S102.
  • step S102 For details, please refer to the related description of step S102, which will not be repeated here.
  • Step S204 According to the sleep quality evaluation results corresponding to the multiple breathing infrared image sequences to be evaluated, the number of different sleep quality evaluation results is counted, and the sleep quality evaluation result with the largest number is determined as the sleep quality evaluation result of the user.
  • step S103 This step is the same as step S103.
  • step S103 For details, please refer to the related description of step S103, which will not be repeated here.
  • the entire ternary generation confrontation network is quantified, so that the network has a regularization effect, reduces the possibility of network overfitting, and enhances the prediction generalization ability of the network.
  • Zhang Quan reduced the network parameters, accelerated the network training speed, and increased the network operation efficiency.
  • the embodiment of the present application quantizes the respiratory infrared image sequence and replaces the 2D convolution with the 3D convolution, which effectively extracts the timing feature information, removes noise and unnecessary redundant information, and retains The feature relationship between the respiratory infrared image sequences reduces the loss of temporal feature information and improves the classification ability and classification accuracy of the classifier.
  • FIG. 4 is a schematic diagram of a sleep quality assessment system based on infrared image sequences provided in the third embodiment of the present application. For ease of description, only the parts related to the embodiment of the present application are shown.
  • the sleep quality assessment system includes:
  • the image sequence acquisition module 41 is configured to acquire a plurality of respiratory infrared image sequences to be evaluated, and one respiratory infrared image sequence to be evaluated includes multiple frames of respiratory infrared images to be evaluated;
  • the sleep quality assessment module 42 is configured to perform sleep quality assessment on each of the plurality of respiratory infrared image sequences to be evaluated by a classifier, and obtain the sleep corresponding to each of the respiratory infrared image sequences to be evaluated. Quality assessment results;
  • the sleep quality determination module 43 is configured to count the number of different sleep quality evaluation results according to the sleep quality evaluation results corresponding to the plurality of breathing infrared image sequences to be evaluated, and determine that the sleep quality evaluation result with the largest number is the sleep quality of the user evaluation result.
  • the sleep quality assessment module 42 includes:
  • a sequence input unit configured to input each of the breathing infrared image sequences to be evaluated to the classifier
  • the target feature map acquiring unit is configured to acquire the target feature map of each respiratory infrared image sequence to be evaluated according to the second-order pooling block in the classifier and the network layer based on tensor decomposition;
  • the evaluation result acquisition unit is configured to evaluate the sleep quality of each breathing infrared image sequence to be evaluated according to the target feature map and the fully connected layer based on tensor decomposition in the classifier, to obtain each of the to-be-evaluated respiratory infrared image sequences. Evaluate the sleep quality evaluation results corresponding to the breathing infrared image sequence.
  • the sleep quality assessment system further includes:
  • the classifier training module is used to train the classifier through the quantified ternary generative confrontation network.
  • the quantified ternary generative confrontation network includes a generator, the classifier, and a discriminator;
  • the classifier training module includes:
  • the first processing unit is configured to input one-dimensional random noise and a target label to the generator, and obtain a first breathing infrared image carrying the target label through the deconvolution layer based on tensor decomposition in the generator sequence;
  • the second processing unit is configured to input the first respiratory infrared image sequence to the discriminator, and obtain the discriminator’s response to the first tensor decomposition-based network layer and the fully connected layer in the discriminator.
  • the first training unit is used to train the generator according to the discrimination result
  • the first acquiring unit is configured to acquire a second respiratory infrared image sequence without a tag
  • the third processing unit is configured to input the second breathing infrared image sequence to the classifier, and obtain the third order through the second-order pooling block, the network layer based on tensor decomposition, and the fully connected layer in the classifier.
  • a breathing infrared image sequence where the third breathing infrared image sequence refers to a second breathing infrared image sequence carrying a tag;
  • the second acquiring unit is used to acquire the fourth respiratory infrared image sequence carrying the tag
  • the second training unit is used to train the discriminator according to the first respiratory infrared image sequence, the third respiratory infrared image sequence, and the fourth respiratory infrared image sequence, and to obtain the discriminator’s response to the The discrimination result of the third breath infrared image sequence;
  • the third training unit is configured to train the classifier according to the first respiratory infrared image sequence, the discrimination result of the third respiratory infrared image sequence by the discriminator, and the fourth respiratory infrared image sequence.
  • the first processing unit is specifically configured to:
  • the network layer based on tensor decomposition in the discriminator includes a convolutional layer based on tensor decomposition, a first dense connection block, and a second dense connection block; the second processing unit is specifically configured to:
  • Tensor decomposition is performed on the convolution kernel of the first densely connected block in the discriminator to obtain a third tensor
  • Tensor decomposition is performed on the convolution kernel of the second densely connected block in the discriminator to obtain a fourth tensor
  • the first feature map, the second feature map, and the third feature map are all feature maps of the first respiratory infrared image sequence.
  • the discriminator further includes a first transition layer and a second transition layer, the first transition layer and the second transition layer are both 1 ⁇ 1 ⁇ 1 convolution kernels; the first transition layer The layer is located between the first densely connected block and the second densely connected block in the discriminator, and is used to reduce the number of the second feature maps; the second transition layer is located in the second densely connected block in the discriminator Between it and the fully connected layer, it is used to reduce the number of the third feature maps.
  • the network layer based on tensor decomposition in the classifier includes a convolutional layer based on tensor decomposition, a third dense connection block, and a fourth dense connection block; the third processing unit is specifically configured to:
  • Tensor decomposition is performed on the convolution kernel of the third densely connected block in the classifier to obtain a seventh tensor
  • Tensor decomposition is performed on the convolution kernel of the fourth densely connected block in the classifier to obtain an eighth tensor
  • the fourth feature map, the fifth feature map, the sixth feature map, and the seventh feature map are all feature maps of the second respiratory infrared image sequence.
  • the classifier further includes a third transition layer and a fourth transition layer, the third transition layer and the fourth transition layer are both 1 ⁇ 1 ⁇ 1 convolution kernels; the third transition layer The layer is located between the second-order pooling block and the fourth densely connected block in the classifier, and is used to reduce the number of the sixth feature maps; the fourth transition layer is located in the fourth densely connected block in the classifier Between it and the fully connected layer, it is used to reduce the number of the seventh feature maps.
  • the sleep quality evaluation system provided by the embodiment of the present application can be applied in the foregoing method embodiment 1 and embodiment 2.
  • FIG. 5 is a schematic diagram of a terminal device provided in Embodiment 4 of the present application.
  • the terminal device 5 of this embodiment includes a processor 50, a memory 51, and a computer program 52 stored in the memory 51 and running on the processor 50.
  • the processor 50 executes the computer program 52, the steps in the foregoing sleep quality assessment method embodiments are implemented.
  • the processor 50 executes the computer program 52, the functions of the modules/units in the foregoing system embodiments are realized.
  • the terminal device 5 may be a computing device such as a desktop computer, a notebook, a palmtop computer, and a cloud server.
  • the terminal device may include, but is not limited to, a processor 50 and a memory 51.
  • FIG. 5 is only an example of the terminal device 5, and does not constitute a limitation on the terminal device 5. It may include more or less components than those shown in the figure, or a combination of certain components, or different components.
  • the terminal device may also include input and output devices, network access devices, buses, and so on.
  • the so-called processor 50 may be a central processing unit (Central Processing Unit, CPU), other general-purpose processors, digital signal processors (Digital Signal Processor, DSP), application specific integrated circuits (Application Specific Integrated Circuit, ASIC), Ready-made programmable gate array (Field-Programmable Gate Array, FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc.
  • the general-purpose processor may be a microprocessor or the processor may also be any conventional processor or the like.
  • the memory 51 may be an internal storage unit of the terminal device 5, such as a hard disk or a memory of the terminal device 5.
  • the memory 51 may also be an external storage device of the terminal device 5, such as a plug-in hard disk equipped on the terminal device 5, a smart memory card (Smart Media Card, SMC), or a Secure Digital (SD). Card, Flash Card, etc. Further, the memory 51 may also include both an internal storage unit of the terminal device 5 and an external storage device.
  • the memory 51 is used to store the computer program and other programs and data required by the terminal device.
  • the memory 51 can also be used to temporarily store data that has been output or will be output.
  • system/terminal device and method may be implemented in other ways.
  • the system/terminal device embodiments described above are only illustrative.
  • the division of the modules or units is only a logical function division, and there may be other divisions in actual implementation, such as multiple units.
  • components can be combined or integrated into another system, or some features can be omitted or not implemented.
  • the displayed or discussed mutual coupling or direct coupling or communication connection may be indirect coupling or communication connection through some interfaces, devices or units, and may be in electrical, mechanical or other forms.
  • the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the objectives of the solutions of the embodiments.
  • the functional units in the various embodiments of the present application may be integrated into one processing unit, or each unit may exist alone physically, or two or more units may be integrated into one unit.
  • the above-mentioned integrated unit can be implemented in the form of hardware or software functional unit.
  • the integrated module/unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer readable storage medium.
  • the present application implements all or part of the processes in the above-mentioned embodiments and methods, and can also be completed by instructing relevant hardware through a computer program.
  • the computer program can be stored in a computer-readable storage medium. When the program is executed by the processor, it can implement the steps of the foregoing method embodiments.
  • the computer program includes computer program code, and the computer program code may be in the form of source code, object code, executable file, or some intermediate forms.
  • the computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory) , Random Access Memory (RAM, Random Access Memory), electrical carrier signal, telecommunications signal, and software distribution media, etc.
  • ROM Read-Only Memory
  • RAM Random Access Memory
  • electrical carrier signal telecommunications signal
  • software distribution media etc.
  • the content contained in the computer-readable medium can be appropriately added or deleted according to the requirements of the legislation and patent practice in the jurisdiction.
  • the computer-readable medium Does not include electrical carrier signals and telecommunication signals.
  • This application implements all or part of the processes in the above-mentioned embodiments and methods, and can also be completed by a computer program product.
  • the computer program product When the computer program product is run on a terminal device, the terminal device can realize all of the foregoing when the terminal device is executed. Steps in the method embodiment.

Landscapes

  • Engineering & Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Medical Informatics (AREA)
  • Theoretical Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Public Health (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Pathology (AREA)
  • Biomedical Technology (AREA)
  • Evolutionary Computation (AREA)
  • Animal Behavior & Ethology (AREA)
  • Surgery (AREA)
  • Biophysics (AREA)
  • Veterinary Medicine (AREA)
  • Heart & Thoracic Surgery (AREA)
  • Molecular Biology (AREA)
  • General Physics & Mathematics (AREA)
  • Databases & Information Systems (AREA)
  • Multimedia (AREA)
  • Software Systems (AREA)
  • Computing Systems (AREA)
  • Physiology (AREA)
  • Psychiatry (AREA)
  • Radiology & Medical Imaging (AREA)
  • Pulmonology (AREA)
  • Quality & Reliability (AREA)
  • Fuzzy Systems (AREA)
  • Mathematical Physics (AREA)
  • Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
  • Signal Processing (AREA)
  • Social Psychology (AREA)
  • Human Computer Interaction (AREA)
  • Data Mining & Analysis (AREA)
  • Epidemiology (AREA)
  • Primary Health Care (AREA)
  • Measurement Of The Respiration, Hearing Ability, Form, And Blood Characteristics Of Living Organisms (AREA)

Abstract

一种基于红外图像序列的睡眠质量评估系统和方法，包括：获取多个待评估呼吸红外图像序列，一个待评估呼吸红外图像序列包括多帧待评估呼吸红外图像（S101）；通过分类器（C）对多个待评估呼吸红外图像序列中每个待评估呼吸红外图像序列进行睡眠质量评估，获得每个待评估呼吸红外图像序列对应的睡眠质量评估结果（S102）；根据多个待评估呼吸红外图像序列分别对应的睡眠质量评估结果，统计不同睡眠质量评估结果的数量，并确定数量最多的睡眠质量评估结果为用户的睡眠质量评估结果（S103）。对用户进行无接触的睡眠监测，同时降低监测成本，提高睡眠质量的评估准确度。

Description

一种基于红外图像序列的睡眠质量评估系统和方法 技术领域
本申请属于深度学习技术领域，尤其涉及一种基于红外图像序列的睡眠质量评估系统和方法。
背景技术
睡眠作为一个复杂的生命行为，与人的健康息息相关，睡眠大约占据人类寿命三分之一的时间。然而现代生活节奏普遍加快，生活工作压力不断增加，睡眠缺失愈发常见，严重影响了人们日常生活质量甚至身体健康。因此对睡眠质量进行检测，评估人的睡眠状况，对不良睡眠及时进行引导治疗具有重要意义。
对于睡眠质量的检测，现有方法通常使用多导生理睡眠检测图和穿戴式设备捕获生理学数据等方法评估用户的睡眠质量。然而，多导生理睡眠检测图方法，需要采用大量传感器，且多数传感器位于敏感的头部和面部，容易给被测者带来生理上的不适和心理上的压力，从而影响被测者在测量时的睡眠，导致检测结果与真实情况产生偏差，且多导生理睡眠监测仪设备操作复杂，不方便移动，被测者需要在医院里使用相关仪器监测约8小时，价格昂贵，时间成本和价格成本都比较高；穿戴式设备捕获生理学数据方法在分析睡眠质量时，需要直接与人体接触，给被测者带来行动不便和心理负担，对被测者睡眠过程造成干扰，影响被测者的睡眠习惯，最终影响对被测者睡眠质量的评估准确度。
技术问题
本申请实施例提供了一种基于红外图像序列的睡眠质量评估系统和方法，以对用户进行无接触的睡眠监测，同时降低监测成本，提高睡眠质量的评估准确度。
技术解决方案
本申请实施例的第一方面提供了一种基于红外图像序列的睡眠质量评估方法，所述睡眠质量评估方法包括：
获取多个待评估呼吸红外图像序列，一个待评估呼吸红外图像序列包括多帧待评估呼吸红外图像；
通过分类器对所述多个待评估呼吸红外图像序列中每个待评估呼吸红外图像序列进行睡眠质量评估，获得所述每个待评估呼吸红外图像序列对应的睡眠质量评估结果；
根据所述多个待评估呼吸红外图像序列分别对应的睡眠质量评估结果，统计不同睡眠质量评估结果的数量，并确定数量最多的睡眠质量评估结果为用户的睡眠质量评估结果。
本申请实施例的第二方面提供了一种基于红外图像序列的睡眠质量评估系统，所述睡眠质量评估系统包括：
图像序列获取模块，用于获取多个待评估呼吸红外图像序列，一个待评估呼吸红外图像序列包括多帧待评估呼吸红外图像；
睡眠质量评估模块，用于通过分类器对所述多个待评估呼吸红外图像序列中每个待评估呼吸红外图像序列进行睡眠质量评估，获得所述每个待评估呼吸红外图像序列对应的睡眠质量评估结果；
睡眠质量确定模块，用于根据所述多个待评估呼吸红外图像序列分别对应的睡眠质量评估结果，统计不同睡眠质量评估结果的数量，并确定数量最多的睡眠质量评估结果为用户的睡眠质量评估结果。
本申请实施例的第三方面提供了一种终端设备，包括存储器、处理器以及存储在所述存储器中并可在所述处理器上运行的计算机程序，所述处理器执行所述计算机程序时实现如上述第一方面所述睡眠质量评估方法的步骤。
本申请实施例的第四方面提供了一种计算机可读存储介质，所述计算机可读存储介质存储有计算机程序，所述计算机程序被处理器执行时实现如上述第一方面所述睡眠质量评估方法的步骤。
本申请的第五方面提供了一种计算机程序产品，当所述计算机程序产品在终端设备上运行时，使得所述终端设备执行如上述第一方面所述睡眠质量评估方法的步骤。
有益效果
由上可见，本方案通过红外摄像装置获取用户在睡眠时的多个待评估呼吸红外图像序列，可以实现对用户的无接触睡眠监测，同时降低监测成本，并通过基于张量分解的分类器，将多个待评估呼吸红外图像序列中每个待评估呼吸红外图像序列作为一个整体进行睡眠质量评估，可以有效保留每个待评估呼吸红外图像序列中待评估呼吸红外图像之间在时空上的连续性信息，提高对每个待评估呼吸红外图像序列的睡眠质量评估结果的准确度，并将多个睡眠质量评估结果中，占比最多的睡眠质量评估结果作为用户的睡眠质量评估结果，提高了对用户睡眠质量的评估准确度。
附图说明
为了更清楚地说明本申请实施例中的技术方案，下面将对实施例或现有技术描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本申请的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动性的前提下，还可以根据这些附图获得其他的附图。
图1是本申请实施例一提供的基于红外图像序列的睡眠质量评估方法的实现流程示意图；
图2是本申请实施例二提供的基于红外图像序列的睡眠质量评估方法的实现流程示意图；
图3a是张量化的三元生成对抗网络的结构示例图；图3b是生成器的结构示例图；图3c是判别器的结构示例图；图3d是分类器的结构示例图；
图4是本申请实施例三提供的基于红外图像序列的睡眠质量评估系统的结构示意图；
图5是本申请实施例四提供的终端设备的结构示意图。
具体实施方式
以下描述中，为了说明而不是为了限定，提出了诸如特定系统结构、技术之类的具体细节，以便透彻理解本申请实施例。然而，本领域的技术人员应当清楚，在没有这些具体细节的其它实施例中也可以实现本申请。在其它情况中，省略对众所周知的系统、装置、电路以及方法的详细说明，以免不必要的细节妨碍本申请的描述。
应当理解，当在本说明书和所附权利要求书中使用时，术语“包括”指示所描述特征、整体、步骤、操作、元素和/或组件的存在，但并不排除一个或多个其它特征、整体、步骤、操作、元素、组件和/或其集合的存在或添加。
还应当理解，在此本申请说明书中所使用的术语仅仅是出于描述特定实施例的目的而并不意在限制本申请。如在本申请说明书和所附权利要求书中所使用的那样，除非上下文清楚地指明其它情况，否则单数形式的“一”、“一个”及“该”意在包括复数形式。
应理解，本实施例中各步骤的序号的大小并不意味着执行顺序的先后，各过程的执行顺序应以其功能和内在逻辑确定，而不应对本申请实施例的实施过程构成任何限定。
为了说明本申请所述的技术方案，下面通过具体实施例来进行说明。
参见图1，是本申请实施例一提供的基于红外图像序列的睡眠质量评估方法的实现流程示意图，该睡眠质量评估方法应用于终端设备，如图所示该睡眠质量评估方法可以包括以下步骤：
步骤S101，获取多个待评估呼吸红外图像序列，一个待评估呼吸红外图像序列包括多帧待评估呼吸红外图像。
在本申请实施例中，可以通过红外摄像装置获取用户睡眠时的多个待评估呼吸红外图像序列(即至少两个待评估呼吸红外图像序列)，该红外摄像装置可以集成在终端设备中，也可以独立于终端设备(即红外摄像装置不集成在终端设备中)，在红外摄像装置独立于终端设备时，可以通过无线或有线等方式建立红外摄像装置与终端设备之间的连接通信，将红外摄像装置获取的多个待评估呼吸红外图像序列传输给终端设备。其中，待评估呼吸红外图像可以是指红外摄像装置对用户(即被评估睡眠质量的用户)口鼻区域所拍摄的图像，该红外摄像装置无需与用户接触，可以实现对用户的无接触睡眠监测，避免对用户日常睡眠造成干扰，同时降低了监测成本。通过连续采集的多帧待评估呼吸红外图像能够捕捉到用户呼吸时在口鼻区域的温度变化，从而提取呼吸频率和呼吸深度等特征。一个待评估呼吸红外图像序列可以是指一个等待进行睡眠质量评估的呼吸红外图像序列。
红外摄像装置在获取用户睡眠时的多个待评估呼吸红外图像序列时，可以采用滑动窗口方式以预设时长为一个基本单位采集用户睡眠时一个待评估呼吸红外图像序列，例如以一分钟为一个基本单位，连续采集五分钟的待评估呼吸红外图像，一分钟内的多帧待评估呼吸红外图像组成一个待评估呼吸红外图像序列，五分钟包括五个一分钟，即五分钟对应五个待评估呼吸红外图像序列。
可选的，在将每个待评估呼吸红外图像序列输入至分类器之前，还可以对每个待评估呼吸红外图像序列进行预处理，该预处理包括但不限于将每个待评估呼吸红外图像序列中多帧待评估呼吸红外图像的尺寸调整为相同(例如均调整为预设尺寸)和/或将每个待评估呼吸红外图像序列中多帧待评估呼吸红外图像的像素值均调整为预设范围内。
步骤S102，通过分类器对所述多个待评估呼吸红外图像序列中每个待评估呼吸红外图像序列进行睡眠质量评估，获得所述每个待评估呼吸红外图像序列对应的睡眠质量评估结果。
其中，上述分类器能够将一个待评估呼吸红外图像序列作为一个整体进行睡眠质量评估，即上述分类器能够直接对一个待评估呼吸红外图像序列进行睡眠质量评估，可以有效保留一个待评估呼吸红外图像序列中多帧待评估呼吸红外图像之间在时空上的连续性信息，提高对一个待评估呼吸红外图像序列的睡眠质量评估结果的准确性。睡眠质量评估结果用于指示睡眠质量的好或坏，睡眠质量评估结果包括但不限于第一睡眠质量评估结果和第二睡眠质量评估结果，第一睡眠质量评估结果可以是指睡眠质量好，第二睡眠质量评估结果可以是指睡眠质量坏，即，第一睡眠质量评估结果指示的睡眠质量优于第二睡眠质量评估结果指示的睡眠质量。需要说明的是，也可以根据实际需要将睡眠质量评估结果的内容重新进行划分，例如睡眠质量评估结果可以是睡眠质量优、睡眠质量良、睡眠质量差等，在此不作限定。需要说明的是，分类器对一个待评估呼吸红外图像序列进行睡眠质量评估是对一个待评估呼吸红外图像序列进行分类，分类的类别即为睡眠质量评估结果，例如分类的类别包括睡眠质量好和睡眠质量坏，通过分类器判断一个待评估呼吸红外图像序列所属类别是睡眠质量好，还是睡眠质量坏。
需要说明的是，在通过分类器对多个待评估呼吸红外图像序列中每个待评估呼吸红外图像序列进行睡眠质量评估时，可以是在获取到多个待评估呼吸红外图像序列之后，通过分类器分别对多个待评估呼吸红外图像序列进行睡眠质量评估；也可以是在获取到第一个待评估呼吸红外图像序列时，通过分类器对第一个待评估呼吸红外图像序列进行睡眠质量评估，获取第一个待评估呼吸红外图像序列的睡眠质量评估结果(即第一个待评估呼吸红外图像序列对应的睡眠质量评估结果)，在获取到第二个待评估呼吸红外图像序列时，通过分类器对第二待评估呼吸红外图像序列进行睡眠质量评估，获得第二个待评估呼吸红外图像序列的睡眠质量评估结果，以此类推，直到获得多个 待评估呼吸红外图像序列中最后一个待评估呼吸红外图像序列的睡眠质量评估结果。例如，红外摄像装置逐分钟(即以一分钟为一个基本单位)采集待评估呼吸红外图像序列时，可以逐分钟进行睡眠质量评估，统计每一分钟内待评估呼吸红外图像序列的睡眠质量评估结果。
可选的，所述通过分类器对所述多个待评估呼吸红外图像序列中每个待评估呼吸红外图像序列进行睡眠质量评估，获得所述每个待评估呼吸红外图像序列对应的睡眠质量评估结果包括：
将所述每个待评估呼吸红外图像序列输入至所述分类器；
根据所述分类器中二阶池化块和基于张量分解的网络层，获取所述每个待评估呼吸红外图像序列的目标特征图；
根据所述目标特征图和所述分类器中基于张量分解的全连接层，对所述每个待评估呼吸红外图像序列进行睡眠质量评估，获得所述每个待评估呼吸红外图像序列对应的睡眠质量评估结果。
在本申请实施例中，分类器在接收到一个待评估呼吸红外图像序列时，可以通过二阶池化块利用该待评估呼吸红外图像序列的二阶信息在自注意力机制作用下自动提取与睡眠有关的呼吸特征，提高对该待评估呼吸红外图像序列的睡眠质量的评估准确度。基于张量分解的网络层包括基于张量分解的卷积层、两个基于张量分解的密集连接块，分类器利用密集连接块的稠密连接机制，可以有效解决梯度消失问题。其中，分类器中的密集连接块可以是指密集型卷积神经网络，例如残差网络。一个密集连接块通常包括多个卷积层。基于张量分解的卷积层是指对卷积层的卷积核进行张量分解，基于张量分解的卷积层是3D卷积层，对3D卷积层的卷积核进行张量分解，可以将3D卷积层的卷积核分解为两个矩阵与一个三维张量的乘积，第一阶和第三阶为矩阵，第二阶为三维张量。
基于张量分解的全连接层是指对全连接层的权重进行张量分解，可以将全连接层的权重分解为两个矩阵和一个三维张量的乘积，第一阶和第三阶为矩阵，第二阶为三维张量。
在本申请实施例中，一个待评估呼吸红外图像序列输入至分类器之后，先将待评估呼吸红外图像序列作为一个张量整体，与卷积层中张量分解后的卷积核进行卷积计算，接着与一个密集连接块中张量分解后的卷积核进行卷积计算，然后通过二阶池化块，再与一个密集连接块中张量分解后的卷积和进行卷积计算，最后再通过权重已进行张量分解的全连接层获得对待评估呼吸红外图像序列的睡眠质量评估结果。目标特征图为分类器中最后一个密集连接块输出的特征图，是通过多层卷积后产生的高阶特征图，因为分类器中最后一个密集连接块输出的特征图的数量为多个，故可以称之为高阶特征图。其中，通过张量分解表示分类器中的网络层和全连接层，可以减少分类器中参数数量，解决对张量形式数据进行向量化计算造成张量数据内部结构信息丢失和分类器中参数数量过大消耗存储空间的问题。上述张量分解可以是指Tensor-Train张量分解。
步骤S103，根据所述多个待评估呼吸红外图像序列分别对应的睡眠质量评估结果，统计不同睡眠质量评估结果的数量，并确定数量最多的睡眠质量评估结果为用户的睡眠质量评估结果。
其中，一个待评估呼吸红外图像序列对应一个睡眠质量评估结果，那么多个待评估呼吸红外图像序列就对应多个睡眠质量评估结果，多个睡眠质量评估结果中可能存在相同的睡眠质量评估结果。
示例性的，通过分类器分别对五个待评估呼吸红外图像序列，第一个待评估呼吸红外图像序列、第二个待评估呼吸红外图像序列以及第五个待评估呼吸红外图像序列的睡眠质量评估结果均为睡眠质量好，第三个待评估呼吸红外图像序列和第四个待评估呼吸红外图 像序列的睡眠质量评估结果是睡眠质量坏，统计出五个睡眠质量评估结果中有三个睡眠质量好和两个睡眠质量坏，睡眠质量好的数量最多，那么就可以确定用户的睡眠质量评估结果是睡眠质量好。
本申请实施例通过红外摄像装置获取用户在睡眠时的待评估呼吸红外图像序列，并通过基于张量分解的分类器，将每个待评估呼吸红外图像序列作为一个张量整体进行睡眠质量评估，可以实现无接触的睡眠质量评估，提高对用户睡眠质量的评估准确度。
参见图2，是本申请实施例二提供的基于红外图像序列的睡眠质量评估方法的实现流程示意图，该睡眠质量评估方法应用于终端设备，如图所示该睡眠质量评估方法可以包括以下步骤：
步骤S201，通过张量化的三元生成对抗网络对分类器进行训练。
其中，所述张量化的三元生成对抗网络包括生成器、分类器和判别器，生成器、分类器和判别器均使用了张量分解，有效地减少了生成器、分类器和判别器中的参数数量，并可实现对呼吸红外图像序列的整体处理，如图3a所示是张量化的三元生成对抗网络的结构示例图，图中G表示生成器，C表示分类器，D表示判别器，未标记呼吸红外图像序列X c为未携带标签的呼吸红外图像序列，标记红外图像序列(X l,Y l)为携带标签的呼吸红外图像序列，Y c为未标记红外图像序列的标签，X g为生成器生成的呼吸红外图像序列。
张量分解算法Tensor-Train对一个d阶张量进行分解表示，可以表示为两个矩阵和d-2个三维张量的乘积，第1个和第d个为矩阵，其余d-2个为三维张量，d为大于2的整数。例如d阶张量A进行分解后可以表示为A(l 1,l 2,…,l d)＝G(l 1)G(l 2)…G(l d)，其中，G(l 1)和G(l d)为矩阵，G(l 2)为三维张量。
可选的，所述通过张量化的三元生成对抗网络对分类器进行训练包括：
将一维随机噪声和目标标签输入至所述生成器，通过所述生成器中基于张量分解的反卷积层，获得携带所述目标标签的第一呼吸红外图像序列；
将所述第一呼吸红外图像序列输入至所述判别器，通过所述判别器中基于张量分解的网络层和全连接层，获得所述判别器对所述第一呼吸红外图像序列的判别结果；
根据所述判别结果训练所述生成器；
获取未携带标签的第二呼吸红外图像序列；
将所述第二呼吸红外图像序列输入至所述分类器，通过所述分类器中二阶池化块、基于张量分解的网络层和全连接层，获得第三呼吸红外图像序列，所述第三呼吸红外图像序列是指携带标签的第二呼吸红外图像序列；
获取携带标签的第四呼吸红外图像序列；
根据所述第一呼吸红外图像序列、所述第三呼吸红外图像序列以及所述第四呼吸红外图像序列，训练所述判别器，并获取所述判别器对所述第三呼吸红外图像序列的判别结果；
根据所述第一呼吸红外图像序列、所述判别器对所述第三呼吸红外图像序列的判别结果以及所述第四呼吸红外图像序列，训练所述分类器。
在本申请实施例中，生成器采用条件生成对抗网络的思想，以服从正态分布的一维随机噪声作为生成器的输入，同时以睡眠质量为目标标签作为条件输入，中间网络层采用3D反卷积层，接着使用Leaky ReLU作为激活函数，使用Batch Norm进行批正则化，生成器最后一层3D反卷积层后接着一层tanh激活层，使用Tensor-Train张量分解辅助生成器生成携带睡眠质量标签的呼吸红外图像序列，减少了对真实的携带标签的呼吸红外图像序列的需求。如图3b所示是生成器的结构示例图。其中，将携带目标标签的一维随机噪声依次经过3D反卷积层、Leaky ReLU激活函数和批正则化进行特征图的逐步反卷积，可以生成逼近真实的携带目标标签的第一呼吸红外图像序列。
将生成器生成的第一呼吸红外图像序列输入至判别器，获得判别器对第一呼吸红外图像序列的判别结果，根据该判别结果获取生成器的损失函数，根据该损失函数训练生成器，生成器的损失函数可以表示为
Figure PCTCN2019126038-appb-000001
其中，D(x g,y g)表示判别器的判别结果，如果判别结果为真，则D(x g,y g)为1，如果判别结果为假，则D(x g,y g)为0；λ为权重参数(用户可以根据实际需要自行设定)；x label为真实的呼吸红外图像序列，x g为生成的呼吸红外图像序列(即第一呼吸红外图像序列)，
Figure PCTCN2019126038-appb-000002
表示真实的呼吸红外图像序列与生成的呼吸红外图像序列的L 1损失，使得生成的呼吸红外图像序列更加逼近真实的呼吸红外图像序列。
将通过红外摄像装置采集到的未携带标签的呼吸红外图像序列分为两部分：一部分作为第二呼吸红外图像序列，通过分类器进行睡眠质量评估，并将分类器输出的睡眠质量评估结果作为标签，从而获得第三呼吸红外图像序列；另一部分通过睡眠专家进行睡眠质量评估，并将睡眠质量评估结果为标签，从而获得携带标签的第四呼吸红外图像序列；将第一呼吸红外图像序列、第三呼吸红外图像序列以及第四呼吸红外图像序列分别输入至判别器，获取判别器分别对第一呼吸红外图像序列、第三呼吸红外图像序列以及第四呼吸红外图像序列的判别结果，根据上述三个呼吸红外图像序列的判别结果获取判别器的损失函数，根据该损失函数训练判别器。判别器的损失函数可以表示为Loss D＝logD(x l,y l)+αlog(1-D(x c,y c))+(1-α)log(1-D(x g,y g))，其中，D(x l,y l)为第四呼吸红外图像序列的判别结果，D(x c,y c)为第三呼吸红外图像序列的判别结果，D(x g,y g)为第一呼吸红外图像序列的判别结果，α为权重参数(用户可以根据实际需要自行设定，且α大于或等于0且小于或等于1)。
将第一呼吸红外图像序列和第四呼吸红外图像序列分别输入至分类器，对分类器进行分类训练，获得使用第一呼吸红外图像序列对分类器进行训练时的损失函数Loss g以及使用第四呼吸红外图像序列对分类器进行训练时的损失函数Loss l，并将Loss g和Loss l综合成Loss supervised＝Loss g+αLoss l，根据判别器对第三呼吸红外图像序列的判别结果，获得分类器对第二呼吸红外图像序列进行分类的损失函数Loss unsupervised，分类器的损失函数可以表示为Loss c＝Loss supervised+Loss unsupervised。
在本申请实施例中，张量化的三元生成对抗网络，通过大量的第二呼吸红外图像序列(即未携带标签的呼吸红外图像序列)和少量的第四呼吸红外图像序列(即携带标签的呼吸红外图像序列)进行张量化的三元生成对抗网络的训练，可以解决携带标签的呼吸红外图像序列数据较少的问题，同时充分利用未携带标签的呼吸红外图像序列，有利于提高张量化的三元生成对抗网络的稳健性。
可选的，所述通过生成器中基于张量分解的反卷积层，获得携带所述目标标签的第一呼吸红外图像序列包括：
对所述生成器中反卷积层的反卷积核进行张量分解，获得第一张量；
将所述一维随机噪声与所述第一张量进行反卷积计算，获得携带所述目标标签的第一呼吸红外图像序列。
在本申请实施例中，对生成器中3D反卷积层的反卷积核进行张量分解后，可以分解为两个矩阵和一个三维张量的乘积(其中，第一阶和第三阶为矩阵，第二阶为三维张量)，然后将输入的一维随机噪声与第一张量(即两个矩阵和一个三维张量的乘积)进行多层的反卷积计算，再通过图3b所示结构示例图中的激活函数Leaky ReLU、批正则化以及tanh激活层生成逼近真实的携带目标标签的第一呼吸红外图像序列。
可选的，所述判别器中基于张量分解的网络层包括基于张量分解的卷积层、第一密集连接块以及第二密集连接块；所述通过所述判别器中基于张量分解的网络层和全连接层，获得所述判别器对所述第一呼吸红外图像序列的判别结果包括：
对所述判别器中卷积层的卷积核进行张量分解，获得第二张量；
将所述第一呼吸红外图像序列与所述第二张量进行卷积计算，获得第一特征图；
对所述判别器中第一密集连接块的卷积核进行张量分解，获得第三张量；
将所述第一特征图与所述第三张量进行卷积计算，获得第二特征图；
对所述判别器中第二密集连接块的卷积核进行张量分解，获得第四张量；
将所述第二特征图与所述第四张量进行卷积计算，获得第三特征图；
对所述判别器中全连接层的权重进行张量分解，获得第五张量；
根据所述第三特征图和所述第五张量，获得所述判别器对所述第一呼吸红外图像序列的判别结果；
其中，所述第一特征图、所述第二特征图以及所述第三特征图均为所述第一呼吸红外图像序列的特征图。
在本申请实施例中，对判别器中3D卷积层的卷积核进行张量分解后，可以分解为两个矩阵和一个三维张量的乘积(其中，第一阶和第三阶为矩阵，第二阶为三维张量)，分解后得到的两个矩阵和一个三维张量的乘积即为第二张量，将第一呼吸红外图像序列通过第二张量进行多层的卷积计算后所得特征图即为第一特征图；对判别器中第一密集连接块的卷积核进行张量分解后，可以分解为两个矩阵和一个三维张量的乘积(其中，第一阶和第三阶为矩阵，第二阶为三维张量)，分解后得到的两个矩阵和一个三维张量的乘积即为第三张量，将第一特征图通过第三张量进行多层的卷积计算后所得特征图即为第二特征图；对判别器中第二密集连接块的卷积核进行张量分解后，可以分解为两个矩阵和一个三维张量的乘积(其中，第一阶和第三阶为矩阵，第二阶为三维张量)，分解后得到的两个矩阵和一个三维张量的乘积即为第四张量，将第二特征图通过第四张量进行多层的卷积计算后所得特征图即为第三特征图；对判别器中全连接层的权重进行张量分解，可以分解为两个矩阵和一个三维张量的乘积(其中，第一阶和第三阶为矩阵，第二阶为三维张量)，分解后得到的两个矩阵和一个三维张量的乘积即为第五张量，第三特征图通过第五张量可以实现对第一呼吸红外图像序列的真假判别。
可选的，所述判别器还包括第一过渡层和第二过渡层，所述第一过渡层、所述第二过渡层均为1×1×1的卷积核；所述第一过渡层位于所述判别器中第一密集连接块与第二密集连接块之间，用于减少所述第二特征图的数量；所述第二过渡层位于所述判别器中第二密集连接块与全连接层之间，用于减少所述第三特征图的数量。
在本申请实施例中，第一特征图经过第一密集连接块卷积处理后，得到的第二特征图的数量增加，通过1×1×1的3D卷积核进行卷积计算，可以减少第二特征图的数量，即减少通道数量。第二特征图经过第二密集连接块卷积处理后，得到的第三特征图的数量增加，通过1×1×1的3D卷积核进行卷积计算，可以减少第三特征图的数量，即减少通道数量。
如图3c所示是判别器的结构示例图，输入判别器的携带标签的呼吸红外图像序列(可以是输入至判别器的任一携带标签的呼吸红外图像序列)通过基于张量分解的卷积层(即3D张量化卷积层)、基于张量分解的密集连接块(即3D张量化密集连接块)、过渡层和基于张量分解的全连接层组成的深度神经网络，对呼吸红外图像序列进行特征提取，获得保留了呼吸红外图像序列在时空上的呼吸特征信息，最后将提取的呼吸特征信息通过一层基于张量分解的全连接层，实现对携带标签的呼吸红外图像序列的真假判别。判别器还包括Leaky ReLU激活函数、Batch Norm正则化和sigmoid函数。
对于判别器，本申请实施例利用密集连接块将位于密集连接块之前的网络层提取的特征图直接输入到后续网络层进行级联，减少梯度传递过程中的特征损失，解决深度神经网络反向传播过程中梯度消失的问题，稳定对抗生成网络的训练，提升判别器对生成样本和真实样本的判别性能。
可选的，所述分类器中基于张量分解的网络层包括基于张量分解的卷积层、第三密集连接块以及第四密集连接块；所述通过所述分类器中二阶池化块、基于张量分解的网络层和全连接层，获得第三吸红外图像序列包括：
对所述分类器中卷积层的卷积核进行张量分解，获得第六张量；
将所述第二呼吸红外图像序列与所述第六张量进行卷积计算，获得第四特征图；
对所述分类器中第三密集连接块的卷积核进行张量分解，获得第七张量；
将所述第四特征图和所述第七张量进行卷积计算，获得第五特征图；
通过所述分类器中二阶池化块处理所述第五特征图，获得第六特征图；
对所述分类器中第四密集连接块的卷积核进行张量分解，获得第八张量；
将所述第六特征图与所述第八张量进行卷积计算，获得第七特征图；
对所述分类器中全连接层的权重进行张量分解，获得第九张量；
根据所述第七特征图和所述第九张量，获得所述第三呼吸红外图像序列；
其中，所述第四特征图、所述第五特征图、所述第六特征图以及所述第七特征图均为所述第二呼吸红外图像序列的特征图。
在本申请实施例中，对分类器中3D卷积层的卷积核进行张量分解后，可以分解为两个矩阵和一个三维张量的乘积(其中，第一阶和第三阶为矩阵，第二阶为三维张量)，分解后得到的两个矩阵和一个三维张量的乘积即为第六张量，将第二呼吸红外图像序列通过第六张量进行多层的卷积计算后所得特征图即为第四特征图；对分类器中第三密集连接块的卷积核进行张量分解后，可以分解为两个矩阵和一个三维张量的乘积(其中，第一阶和第三阶为矩阵，第二阶为三维张量)，分解后得到的两个矩阵和一个三维张量的乘积即为第七张量，将第四特征图通过第七张量进行多层的卷积计算后所得特征图即为第五特征图；第五特征图经过二阶池化块处理后所得的特征图即为第六特征图；对分类器中第四密集连接块的卷积核进行张量分解后，可以分解为两个矩阵和一个三维张量的乘积(其中，第一阶和第三阶为矩阵，第二阶为三维张量)，分解后得到的两个矩阵和一个三维张量的乘积即为第八张量，将第六特征图通过第八张量进行多层的卷积计算后所得特征图即为第七特征图；对分类器中全连接层的权重进行张量分解，可以分解为两个矩阵和一个三维张量的乘积(其中，第一阶和第三阶为矩阵，第二阶为三维张量)，分解后得到的两个矩阵和一个三维张量的乘积即为第九张量，第七特征图通过第九张量可以得到对应类别的概率值，从而获得第二呼吸红外图像序列的睡眠质量评估结果(即第二呼吸红外图像的标签)。
可选的，所述分类器还包括第三过渡层和第四过渡层，所述第三过渡层和所述第四过渡层均为1×1×1的卷积核；所述第三过渡层位于所述分类器中二阶池化块与第四密集连接块之间，用于减少所述第六特征图的数量；所述第四过渡层位于所述分类器中第四密集连接块与全连接层之间，用于减少所述第七特征图的数量。
在本申请实施例中，第四特征图经过第三密集连接块卷积处理后，得到的第五特 征图的数量增加，进而导致经过二阶池化块处理的第六特征图的数量也较多，通过1×1×1的3D卷积核进行卷积计算，可以减少第六特征图的数量，即减少通道数量。第六特征图经过第四密集连接块卷积处理后，得到的第七特征图的数量增加，通过1×1×1的3D卷积核进行卷积计算，可以减少第七特征图的数量，即减少通道数量。
如图3d所示是分类器的结构示例图，输入分类器的呼吸红外图像序列通过基于张量分解的卷积层(即3D张量化卷积层)和基于张量分解的密集连接块(即3D张量化密集连接块)进行多层的卷积计算，并通过过渡层减少特征图数量，再通过一层基于张量分解的全连接层，可以得到对应类别的概率值，从而获得呼吸红外图像序列的睡眠质量评估结果。分类器还包括Leaky ReLU激活函数、Batch Norm正则化和sigmoid函数。
对于分类器，本申请实施例利用基于二阶池化的密集连接块的自注意力机制根据不同区域的相关性对特征图在特征通道的维度上进行加权，使重要的通道权重大，不重要的通道权重小，有效地提取红外图像序列中的呼吸特征，提高对睡眠质量的评估精度。
步骤S202，获取多个待评估呼吸红外图像序列，一个待评估呼吸红外图像序列包括多帧待评估呼吸红外图像。
该步骤与步骤S101相同，具体可参见步骤S101的相关描述，在此不再赘述。
步骤S203，通过分类器对所述多个待评估呼吸红外图像序列中每个待评估呼吸红外图像序列进行睡眠质量评估，获得所述每个待评估呼吸红外图像序列对应的睡眠质量评估结果。
该步骤与步骤S102相同，具体可参见步骤S102的相关描述，在此不再赘述。
步骤S204，根据所述多个待评估呼吸红外图像序列分别对应的睡眠质量评估结果，统计不同睡眠质量评估结果的数量，并确定数量最多的睡眠质量评估结果为用户的睡眠质量评估结果。
该步骤与步骤S103相同，具体可参见步骤S103的相关描述，在此不再赘述。
本申请实施例通过张量化整个三元生成对抗网络，使得网络具有正则化效果，减少了网络过拟合的可能，增强了网络的预测泛化能力。同时张量化减少了网络参数，加快了网络训练速度，增加了网络运行效率。另外，本申请实施例通过将呼吸红外图像序列张量化，并将2D卷积替换为3D卷积，在有效地提取了时序特征信息的同时，去除了噪声以及非必要的冗余信息，保留了呼吸红外图像序列之间的特征关系，减少了时序特征信息的损失，提高了分类器的分类能力和分类精度。
参见图4，是本申请实施例三提供的基于红外图像序列的睡眠质量评估系统的示意图，为了便于说明，仅示出了与本申请实施例相关的部分。
所述睡眠质量评估系统包括：
图像序列获取模块41，用于获取多个待评估呼吸红外图像序列，一个待评估呼吸红外图像序列包括多帧待评估呼吸红外图像；
睡眠质量评估模块42，用于通过分类器对所述多个待评估呼吸红外图像序列中每个待评估呼吸红外图像序列进行睡眠质量评估，获得所述每个待评估呼吸红外图像序列对应的睡眠质量评估结果；
睡眠质量确定模块43，用于根据所述多个待评估呼吸红外图像序列分别对应的睡眠质量评估结果，统计不同睡眠质量评估结果的数量，并确定数量最多的睡眠质量评估结果为用户的睡眠质量评估结果。
可选的，所述睡眠质量评估模块42包括：
序列输入单元，用于将所述每个待评估呼吸红外图像序列输入至所述分类器；
目标特征图获取单元，用于根据所述分类器中二阶池化块和基于张量分解的网络层，获取所述每个待评估呼吸红外图像序列的目标特征图；
评估结果获取单元，用于根据所述目标特征图和所述分类器中基于张量分解的全连接层，对所述每个待评估呼吸红外图像序列进行睡眠质量评估，获得所述每个待评估呼吸红外图像序列对应的睡眠质量评估结果。
可选的，所述睡眠质量评估系统还包括：
分类器训练模块，用于通过张量化的三元生成对抗网络对所述分类器进行训练。
可选的，所述张量化的三元生成对抗网络包括生成器、所述分类器和判别器；所述分类器训练模块包括：
第一处理单元，用于将一维随机噪声和目标标签输入至所述生成器，通过所述生成器中基于张量分解的反卷积层，获得携带所述目标标签的第一呼吸红外图像序列；
第二处理单元，用于将所述第一呼吸红外图像序列输入至所述判别器，通过所述判别器中基于张量分解的网络层和全连接层，获得所述判别器对所述第一呼吸红外图像序列的判别结果；
第一训练单元，用于根据所述判别结果训练所述生成器；
第一获取单元，用于获取未携带标签的第二呼吸红外图像序列；
第三处理单元，用于将所述第二呼吸红外图像序列输入至所述分类器，通过所述分类器中二阶池化块、基于张量分解的网络层和全连接层，获得第三呼吸红外图像序列，所述第三呼吸红外图像序列是指携带标签的第二呼吸红外图像序列；
第二获取单元，用于获取携带标签的第四呼吸红外图像序列；
第二训练单元，用于根据所述第一呼吸红外图像序列、所述第三呼吸红外图像序列以及所述第四呼吸红外图像序列，训练所述判别器，并获取所述判别器对所述第三呼吸红外图像序列的判别结果；
第三训练单元，用于根据所述第一呼吸红外图像序列、所述判别器对所述第三呼吸红外图像序列的判别结果以及所述第四呼吸红外图像序列，训练所述分类器。
可选的，所述第一处理单元具体用于：
对所述生成器中反卷积层的反卷积核进行张量分解，获得第一张量；
将所述一维随机噪声与所述第一张量进行反卷积计算，获得携带所述目标标签的第一呼吸红外图像序列。
可选的，所述判别器中基于张量分解的网络层包括基于张量分解的卷积层、第一密集连接块以及第二密集连接块；所述第二处理单元具体用于：
对所述判别器中卷积层的卷积核进行张量分解，获得第二张量；
将所述第一呼吸红外图像序列与所述第二张量进行卷积计算，获得第一特征图；
对所述判别器中第一密集连接块的卷积核进行张量分解，获得第三张量；
将所述第一特征图与所述第三张量进行卷积计算，获得第二特征图；
对所述判别器中第二密集连接块的卷积核进行张量分解，获得第四张量；
将所述第二特征图与所述第四张量进行卷积计算，获得第三特征图；
对所述判别器中全连接层的权重进行张量分解，获得第五张量；
根据所述第三特征图和所述第五张量，获得所述判别器对所述第一呼吸红外图像序列的判别结果；
其中，所述第一特征图、所述第二特征图以及所述第三特征图均为所述第一呼吸红外图像序列的特征图。
可选的，所述判别器还包括第一过渡层和第二过渡层，所述第一过渡层、所述第二过渡层均为1×1×1的卷积核；所述第一过渡层位于所述判别器中第一密集连接块与第二密集连接块之间，用于减少所述第二特征图的数量；所述第二过渡层位于所述判别器中第二密集连接块与全连接层之间，用于减少所述第三特征图的数量。
可选的，所述分类器中基于张量分解的网络层包括基于张量分解的卷积层、第三密集连接块以及第四密集连接块；所述第三处理单元具体用于：
对所述分类器中卷积层的卷积核进行张量分解，获得第六张量；
将所述第二呼吸红外图像序列与所述第六张量进行卷积计算，获得第四特征图；
对所述分类器中第三密集连接块的卷积核进行张量分解，获得第七张量；
将所述第四特征图和所述第七张量进行卷积计算，获得第五特征图；
通过所述分类器中二阶池化块处理所述第五特征图，获得第六特征图；
对所述分类器中第四密集连接块的卷积核进行张量分解，获得第八张量；
将所述第六特征图与所述第八张量进行卷积计算，获得第七特征图；
对所述分类器中全连接层的权重进行张量分解，获得第九张量；
根据所述第七特征图和所述第九张量，获得所述第三呼吸红外图像序列；
其中，所述第四特征图、所述第五特征图、所述第六特征图以及所述第七特征图均为所述第二呼吸红外图像序列的特征图。
可选的，所述分类器还包括第三过渡层和第四过渡层，所述第三过渡层和所述第四过渡层均为1×1×1的卷积核；所述第三过渡层位于所述分类器中二阶池化块与第四密集连接块之间，用于减少所述第六特征图的数量；所述第四过渡层位于所述分类器中第四密集连接块与全连接层之间，用于减少所述第七特征图的数量。
本申请实施例提供的睡眠质量评估系统可以应用在前述方法实施例一和实施例二中，详情参见上述方法实施例一和实施例二的描述，在此不再赘述。
图5是本申请实施例四提供的终端设备的示意图。如图5所示，该实施例的终端设备5包括：处理器50、存储器51以及存储在所述存储器51中并可在所述处理器50上运行的计算机程序52。所述处理器50执行所述计算机程序52时实现上述各个睡眠质量评估方法实施例中的步骤。或者，所述处理器50执行所述计算机程序52时实现上述各系统实施例中各模块/单元的功能。
所述终端设备5可以是桌上型计算机、笔记本、掌上电脑及云端服务器等计算设备。所述终端设备可包括，但不仅限于，处理器50、存储器51。本领域技术人员可以理解，图5仅仅是终端设备5的示例，并不构成对终端设备5的限定，可以包括比图示更多或更少的部件，或者组合某些部件，或者不同的部件，例如所述终端设备还可以包括输入输出设备、网络接入设备、总线等。
所称处理器50可以是中央处理单元(Central Processing Unit，CPU)，还可以是其他通用处理器、数字信号处理器(Digital Signal Processor，DSP)、专用集成电路(Application Specific Integrated Circuit，ASIC)、现成可编程门阵列(Field-Programmable Gate Array，FPGA)或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件等。通用处理器可以是微处理器或者该处理器也可以是任何常规的处理器等。
所述存储器51可以是所述终端设备5的内部存储单元，例如终端设备5的硬盘或内存。所述存储器51也可以是所述终端设备5的外部存储设备，例如所述终端设备5上配备的插接式硬盘，智能存储卡(Smart Media Card，SMC)，安全数字(Secure Digital，SD)卡，闪存卡(Flash Card)等。进一步地，所述存储器51还可以既包括所述终端设备5的内部存储单元也包括外部存储设备。所述存储器51用于存储所述计算机程序以及所述终端设备所需的其他程序和数据。所述存储器51还可以用于暂时地存储已经输出或者将要输出的数据。
所属领域的技术人员可以清楚地了解到，为了描述的方便和简洁，仅以上述各功能单元、模块的划分进行举例说明，实际应用中，可以根据需要而将上述功能分配由不同的功能单元、模块完成，即将所述装置的内部结构划分成不同的功能单元或模块，以完成以上描述的全部或者部分功能。实施例中的各功能单元、模块可以集成在一个处理单元中，也可以是各个单元单独物理存在，也可以两个或两个以上单元集成在一个单元中，上述集成的单元既可以采用硬件的形式实现，也可以采用软件功能单元的形式实现。另外，各功能单元、模块的具体名称也只是为了便于相互区分，并不用于限制本申请的保护范围。上述 系统中单元、模块的具体工作过程，可以参考前述方法实施例中的对应过程，在此不再赘述。
在上述实施例中，对各个实施例的描述都各有侧重，某个实施例中没有详述或记载的部分，可以参见其它实施例的相关描述。
本领域普通技术人员可以意识到，结合本文中所公开的实施例描述的各示例的单元及算法步骤，能够以电子硬件、或者计算机软件和电子硬件的结合来实现。这些功能究竟以硬件还是软件方式来执行，取决于技术方案的特定应用和设计约束条件。专业技术人员可以对每个特定的应用来使用不同方法来实现所描述的功能，但是这种实现不应认为超出本申请的范围。
在本申请所提供的实施例中，应该理解到，所揭露的系统/终端设备和方法，可以通过其它的方式实现。例如，以上所描述的系统/终端设备实施例仅仅是示意性的，例如，所述模块或单元的划分，仅仅为一种逻辑功能划分，实际实现时可以有另外的划分方式，例如多个单元或组件可以结合或者可以集成到另一个系统，或一些特征可以忽略，或不执行。另一点，所显示或讨论的相互之间的耦合或直接耦合或通讯连接可以是通过一些接口，装置或单元的间接耦合或通讯连接，可以是电性，机械或其它的形式。
所述作为分离部件说明的单元可以是或者也可以不是物理上分开的，作为单元显示的部件可以是或者也可以不是物理单元，即可以位于一个地方，或者也可以分布到多个网络单元上。可以根据实际的需要选择其中的部分或者全部单元来实现本实施例方案的目的。
另外，在本申请各个实施例中的各功能单元可以集成在一个处理单元中，也可以是各个单元单独物理存在，也可以两个或两个以上单元集成在一个单元中。上述集成的单元既可以采用硬件的形式实现，也可以采用软件功能单元的形式实现。
所述集成的模块/单元如果以软件功能单元的形式实现并作为独立的产品销售或使用时，可以存储在一个计算机可读取存储介质中。基于这样的理解，本申请实现上述实施例方法中的全部或部分流程，也可以通过计算机程序来指令相关的硬件来完成，所述的计算机程序可存储于一计算机可读存储介质中，该计算机程序在被处理器执行时，可实现上述各个方法实施例的步骤。其中，所述计算机程序包括计算机程序代码，所述计算机程序代码可以为源代码形式、对象代码形式、可执行文件或某些中间形式等。所述计算机可读介质可以包括：能够携带所述计算机程序代码的任何实体或装置、记录介质、U盘、移动硬盘、磁碟、光盘、计算机存储器、只读存储器(ROM，Read-Only Memory)、随机存取存储器(RAM，Random Access Memory)、电载波信号、电信信号以及软件分发介质等。需要说明的是，所述计算机可读介质包含的内容可以根据司法管辖区内立法和专利实践的要求进行适当的增减，例如在某些司法管辖区，根据立法和专利实践，计算机可读介质不包括电载波信号和电信信号。
本申请实现上述实施例方法中的全部或部分流程，也可以通过一种计算机程序产品来完成，当所述计算机程序产品在终端设备上运行时，使得所述终端设备执行时实现可实现上述各个方法实施例中的步骤。
以上所述实施例仅用以说明本申请的技术方案，而非对其限制；尽管参照前述实施例对本申请进行了详细的说明，本领域的普通技术人员应当理解：其依然可以对前述各实施例所记载的技术方案进行修改，或者对其中部分技术特征进行等同替换；而这些修改或者替换，并不使相应技术方案的本质脱离本申请各实施例技术方案的精神和范围，均应包含在本申请的保护范围之内。

Claims (20)

  1. 一种基于红外图像序列的睡眠质量评估方法，其特征在于，所述睡眠质量评估方法包括：
    获取多个待评估呼吸红外图像序列，一个待评估呼吸红外图像序列包括多帧待评估呼吸红外图像；
    通过分类器对所述多个待评估呼吸红外图像序列中每个待评估呼吸红外图像序列进行睡眠质量评估，获得所述每个待评估呼吸红外图像序列对应的睡眠质量评估结果；
    根据所述多个待评估呼吸红外图像序列分别对应的睡眠质量评估结果，统计不同睡眠质量评估结果的数量，并确定数量最多的睡眠质量评估结果为用户的睡眠质量评估结果。
  2. 如权利要求1所述的睡眠质量评估方法，其特征在于，所述通过分类器对所述多个待评估呼吸红外图像序列中每个待评估呼吸红外图像序列进行睡眠质量评估，获得所述每个待评估呼吸红外图像序列对应的睡眠质量评估结果包括：
    将所述每个待评估呼吸红外图像序列输入至所述分类器；
    根据所述分类器中二阶池化块和基于张量分解的网络层，获取所述每个待评估呼吸红外图像序列的目标特征图；
    根据所述目标特征图和所述分类器中基于张量分解的全连接层，对所述每个待评估呼吸红外图像序列进行睡眠质量评估，获得所述每个待评估呼吸红外图像序列对应的睡眠质量评估结果。
  3. 如权利要求1所述的睡眠质量评估方法，其特征在于，在通过分类器对所述多个待评估呼吸红外图像序列中每个待评估呼吸红外图像序列进行睡眠质量评估之前，还包括：
    通过张量化的三元生成对抗网络对所述分类器进行训练。
  4. 如权利要求3所述的睡眠质量评估方法，其特征在于，所述张量化的三元生成对抗网络包括生成器、所述分类器和判别器；所述通过张量化的三元生成对抗网络对所述分类器进行训练包括：
    将一维随机噪声和目标标签输入至所述生成器，通过所述生成器中基于张量分解的反卷积层，获得携带所述目标标签的第一呼吸红外图像序列；
    将所述第一呼吸红外图像序列输入至所述判别器，通过所述判别器中基于张量分解的网络层和全连接层，获得所述判别器对所述第一呼吸红外图像序列的判别结果；
    根据所述判别结果训练所述生成器；
    获取未携带标签的第二呼吸红外图像序列；
    将所述第二呼吸红外图像序列输入至所述分类器，通过所述分类器中二阶池化块、基于张量分解的网络层和全连接层，获得第三呼吸红外图像序列，所述第三呼吸红外图像序列是指携带标签的第二呼吸红外图像序列；
    获取携带标签的第四呼吸红外图像序列；
    根据所述第一呼吸红外图像序列、所述第三呼吸红外图像序列以及所述第四呼吸红外图像序列，训练所述判别器，并获取所述判别器对所述第三呼吸红外图像序列的判别结果；
    根据所述第一呼吸红外图像序列、所述判别器对所述第三呼吸红外图像序列的判别结果以及所述第四呼吸红外图像序列，训练所述分类器。
  5. 如权利要求4所述的睡眠质量评估方法，其特征在于，所述通过生成器中基于张量分解的反卷积层，获得携带所述目标标签的第一呼吸红外图像序列包括：
    对所述生成器中反卷积层的反卷积核进行张量分解，获得第一张量；
    将所述一维随机噪声与所述第一张量进行反卷积计算，获得携带所述目标标签的第一呼吸红外图像序列。
  6. 如权利要求4所述的睡眠质量评估方法，其特征在于，所述判别器中基于张量分解的网络层包括基于张量分解的卷积层、第一密集连接块以及第二密集连接块；所述通过所述判别器中基于张量分解的网络层和全连接层，获得所述判别器对所述第一呼吸红外图像序列的判别结果包括：
    对所述判别器中卷积层的卷积核进行张量分解，获得第二张量；
    将所述第一呼吸红外图像序列与所述第二张量进行卷积计算，获得第一特征图；
    对所述判别器中第一密集连接块的卷积核进行张量分解，获得第三张量；
    将所述第一特征图与所述第三张量进行卷积计算，获得第二特征图；
    对所述判别器中第二密集连接块的卷积核进行张量分解，获得第四张量；
    将所述第二特征图与所述第四张量进行卷积计算，获得第三特征图；
    对所述判别器中全连接层的权重进行张量分解，获得第五张量；
    根据所述第三特征图和所述第五张量，获得所述判别器对所述第一呼吸红外图像序列的判别结果；
    其中，所述第一特征图、所述第二特征图以及所述第三特征图均为所述第一呼吸红外图像序列的特征图。
  7. 如权利要求6所述的睡眠质量评估方法，其特征在于，所述判别器还包括第一过渡层和第二过渡层，所述第一过渡层、所述第二过渡层均为1×1×1的卷积核；所述第一过渡层位于所述判别器中第一密集连接块与第二密集连接块之间，用于减少所述第二特征图的数量；所述第二过渡层位于所述判别器中第二密集连接块与全连接层之间，用于减少所述第三特征图的数量。
  8. 如权利要求4所述的睡眠质量评估方法，其特征在于，所述分类器中基于张量分解的网络层包括基于张量分解的卷积层、第三密集连接块以及第四密集连接块；所述通过所述分类器中二阶池化块、基于张量分解的网络层和全连接层，获得第三吸红外图像序列包括：
    对所述分类器中卷积层的卷积核进行张量分解，获得第六张量；
    将所述第二呼吸红外图像序列与所述第六张量进行卷积计算，获得第四特征图；
    对所述分类器中第三密集连接块的卷积核进行张量分解，获得第七张量；
    将所述第四特征图和所述第七张量进行卷积计算，获得第五特征图；
    通过所述分类器中二阶池化块处理所述第五特征图，获得第六特征图；
    对所述分类器中第四密集连接块的卷积核进行张量分解，获得第八张量；
    将所述第六特征图与所述第八张量进行卷积计算，获得第七特征图；
    对所述分类器中全连接层的权重进行张量分解，获得第九张量；
    根据所述第七特征图和所述第九张量，获得所述第三呼吸红外图像序列；
    其中，所述第四特征图、所述第五特征图、所述第六特征图以及所述第七特征图均为所述第二呼吸红外图像序列的特征图。
  9. 如权利要求8所述的睡眠质量评估方法，其特征在于，所述分类器还包括第三过渡层和第四过渡层，所述第三过渡层和所述第四过渡层均为1×1×1的卷积核；所述第三过渡层位于所述分类器中二阶池化块与第四密集连接块之间，用于减少所述第六特征图的数量；所述第四过渡层位于所述分类器中第四密集连接块与全连接层之间，用于减少所述第七特征图的数量。
  10. 一种基于红外图像序列的睡眠质量评估系统，其特征在于，所述睡眠质量评估系统包括：
    图像序列获取模块，用于获取多个待评估呼吸红外图像序列，一个待评估呼吸红外图像序列包括多帧待评估呼吸红外图像；
    睡眠质量评估模块，用于通过分类器对所述多个待评估呼吸红外图像序列中每个待评估呼吸红外图像序列进行睡眠质量评估，获得所述每个待评估呼吸红外图像序列 对应的睡眠质量评估结果；
    睡眠质量确定模块，用于根据所述多个待评估呼吸红外图像序列分别对应的睡眠质量评估结果，统计不同睡眠质量评估结果的数量，并确定数量最多的睡眠质量评估结果为用户的睡眠质量评估结果。
  11. 一种终端设备，包括存储器、处理器以及存储在所述存储器中并可在所述处理器上运行的计算机程序，其特征在于，所述处理器执行所述计算机程序时实现如下步骤：
    获取多个待评估呼吸红外图像序列，一个待评估呼吸红外图像序列包括多帧待评估呼吸红外图像；
    通过分类器对所述多个待评估呼吸红外图像序列中每个待评估呼吸红外图像序列进行睡眠质量评估，获得所述每个待评估呼吸红外图像序列对应的睡眠质量评估结果；
    根据所述多个待评估呼吸红外图像序列分别对应的睡眠质量评估结果，统计不同睡眠质量评估结果的数量，并确定数量最多的睡眠质量评估结果为用户的睡眠质量评估结果。
  12. 如权利要求11所述的终端设备，其特征在于，所述处理器执行所述计算机程序时，所述通过分类器对所述多个待评估呼吸红外图像序列中每个待评估呼吸红外图像序列进行睡眠质量评估，获得所述每个待评估呼吸红外图像序列对应的睡眠质量评估结果包括：
    将所述每个待评估呼吸红外图像序列输入至所述分类器；
    根据所述分类器中二阶池化块和基于张量分解的网络层，获取所述每个待评估呼吸红外图像序列的目标特征图；
    根据所述目标特征图和所述分类器中基于张量分解的全连接层，对所述每个待评估呼吸红外图像序列进行睡眠质量评估，获得所述每个待评估呼吸红外图像序列对应的睡眠质量评估结果。
  13. 如权利要求11所述的终端设备，其特征在于，所述处理器执行所述计算机程序时还实现如下步骤：
    在通过分类器对所述多个待评估呼吸红外图像序列中每个待评估呼吸红外图像序列进行睡眠质量评估之前，通过张量化的三元生成对抗网络对所述分类器进行训练。
  14. 如权利要求13所述的终端设备，其特征在于，所述张量化的三元生成对抗网络包括生成器、所述分类器和判别器；所述处理器执行所述计算机程序时，所述通过张量化的三元生成对抗网络对所述分类器进行训练包括：
    将一维随机噪声和目标标签输入至所述生成器，通过所述生成器中基于张量分解的反卷积层，获得携带所述目标标签的第一呼吸红外图像序列；
    将所述第一呼吸红外图像序列输入至所述判别器，通过所述判别器中基于张量分解的网络层和全连接层，获得所述判别器对所述第一呼吸红外图像序列的判别结果；
    根据所述判别结果训练所述生成器；
    获取未携带标签的第二呼吸红外图像序列；
    将所述第二呼吸红外图像序列输入至所述分类器，通过所述分类器中二阶池化块、基于张量分解的网络层和全连接层，获得第三呼吸红外图像序列，所述第三呼吸红外图像序列是指携带标签的第二呼吸红外图像序列；
    获取携带标签的第四呼吸红外图像序列；
    根据所述第一呼吸红外图像序列、所述第三呼吸红外图像序列以及所述第四呼吸红外图像序列，训练所述判别器，并获取所述判别器对所述第三呼吸红外图像序列的判别结果；
    根据所述的第一呼吸红外图像序列、所述判别器对所述第三呼吸红外图像序列的判别结果以及所述第四呼吸红外图像序列，训练所述分类器。
  15. 如权利要求14所述的终端设备，其特征在于，所述处理器执行所述计算机程序时，所述通过生成器中基于张量分解的反卷积层，获得携带所述目标标签的第一呼吸红外图像序列包括：
    对所述生成器中反卷积层的反卷积核进行张量分解，获得第一张量；
    将所述一维随机噪声与所述第一张量进行反卷积计算，获得携带所述目标标签的第一呼吸红外图像序列。
  16. 如权利要求14所述的终端设备，其特征在于，所述判别器中基于张量分解的网络层包括基于张量分解的卷积层、第一密集连接块以及第二密集连接块；所述处理器执行所述计算机程序时，所述通过所述判别器中基于张量分解的网络层和全连接层，获得所述判别器对所述第一呼吸红外图像序列的判别结果包括：
    对所述判别器中卷积层的卷积核进行张量分解，获得第二张量；
    将所述第一呼吸红外图像序列与所述第二张量进行卷积计算，获得第一特征图；
    对所述判别器中第一密集连接块的卷积核进行张量分解，获得第三张量；
    将所述第一特征图与所述第三张量进行卷积计算，获得第二特征图；
    对所述判别器中第二密集连接块的卷积核进行张量分解，获得第四张量；
    将所述第二特征图与所述第四张量进行卷积计算，获得第三特征图；
    对所述判别器中全连接层的权重进行张量分解，获得第五张量；
    根据所述第三特征图和所述第五张量，获得所述判别器对所述第一呼吸红外图像序列的判别结果；
    其中，所述第一特征图、所述第二特征图以及所述第三特征图均为所述第一呼吸红外图像序列的特征图。
  17. 如权利要求16所述的终端设备，其特征在于，所述判别器还包括第一过渡层和第二过渡层，所述第一过渡层、所述第二过渡层均为1×1×1的卷积核；所述第一过渡层位于所述判别器中第一密集连接块与第二密集连接块之间，用于减少所述第二特征图的数量；所述第二过渡层位于所述判别器中第二密集连接块与全连接层之间，用于减少所述第三特征图的数量。
  18. 如权利要求14所述的终端设备，其特征在于，所述分类器中基于张量分解的网络层包括基于张量分解的卷积层、第三密集连接块以及第四密集连接块；所述处理器执行所述计算机程序时，所述通过所述分类器中二阶池化块、基于张量分解的网络层和全连接层，获得第三吸红外图像序列包括：
    对所述分类器中卷积层的卷积核进行张量分解，获得第六张量；
    将所述第二呼吸红外图像序列与所述第六张量进行卷积计算，获得第四特征图；
    对所述分类器中第三密集连接块的卷积核进行张量分解，获得第七张量；
    将所述第四特征图和所述第七张量进行卷积计算，获得第五特征图；
    通过所述分类器中二阶池化块处理所述第五特征图，获得第六特征图；
    对所述分类器中第四密集连接块的卷积核进行张量分解，获得第八张量；
    将所述第六特征图与所述第八张量进行卷积计算，获得第七特征图；
    对所述分类器中全连接层的权重进行张量分解，获得第九张量；
    根据所述第七特征图和所述第九张量，获得所述第三呼吸红外图像序列；
    其中，所述第四特征图、所述第五特征图、所述第六特征图以及所述第七特征图均为所述第二呼吸红外图像序列的特征图。
  19. 如权利要求18所述的终端设备，其特征在于，所述分类器还包括第三过渡层和第四过渡层，所述第三过渡层和所述第四过渡层均为1×1×1的卷积核；所述第三过渡层位于所述分类器中二阶池化块与第四密集连接块之间，用于减少所述第六特征图的数量；所述第四过渡层位于所述分类器中第四密集连接块与全连接层之间，用于减少所述第七特征图的数量。
  20. 一种计算机可读存储介质，所述计算机可读存储介质存储有计算机程序，其特征在于，所述计算机程序被处理器执行时实现如权利求1至9任一项所述的睡眠质量评估方法。
PCT/CN2019/126038 2019-12-17 2019-12-17 一种基于红外图像序列的睡眠质量评估系统和方法 WO2021120007A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
PCT/CN2019/126038 WO2021120007A1 (zh) 2019-12-17 2019-12-17 一种基于红外图像序列的睡眠质量评估系统和方法
US17/786,840 US20230022206A1 (en) 2019-12-17 2019-12-17 Infrared image sequence-based sleep quality evaluation system and method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2019/126038 WO2021120007A1 (zh) 2019-12-17 2019-12-17 一种基于红外图像序列的睡眠质量评估系统和方法

Publications (1)

Publication Number Publication Date
WO2021120007A1 true WO2021120007A1 (zh) 2021-06-24

Family

ID=76476983

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/126038 WO2021120007A1 (zh) 2019-12-17 2019-12-17 一种基于红外图像序列的睡眠质量评估系统和方法

Country Status (2)

Country Link
US (1) US20230022206A1 (zh)
WO (1) WO2021120007A1 (zh)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113610153A (zh) * 2021-08-06 2021-11-05 长沙理工大学 人体红外图像识别方法、装置、计算机设备及存储介质
CN114699040A (zh) * 2022-02-21 2022-07-05 华南师范大学 基于生理信号的醒睡检测方法、装置、设备以及存储介质
CN117351049A (zh) * 2023-12-04 2024-01-05 四川金信石信息技术有限公司 热成像与可见光融合的测点配准引导方法、设备和介质

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US12073538B2 (en) * 2021-04-08 2024-08-27 Canon Medical Systems Corporation Neural network for improved performance of medical imaging systems
CN116386120B (zh) * 2023-05-24 2023-08-18 杭州企智互联科技有限公司 一种用于智慧校园宿舍的无感监控管理系统

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5704367A (en) * 1995-03-28 1998-01-06 Nihon Kohden Corporation Respiration monitor for monitoring respiration based upon an image signal of a facial region
CN101489478A (zh) * 2006-06-01 2009-07-22 必安康医疗有限公司 用于监视生理症状的装置、系统和方法
CN102973273A (zh) * 2012-11-29 2013-03-20 中国人民解放军第四军医大学 一种基于红外辐射检测的睡眠呼吸功能监测系统
CN104834946A (zh) * 2015-04-09 2015-08-12 清华大学 一种非接触式睡眠监测方法及系统
CN106236013A (zh) * 2016-06-22 2016-12-21 京东方科技集团股份有限公司 一种睡眠监测方法及装置
CN107280673A (zh) * 2017-06-02 2017-10-24 南京理工大学 一种基于关键帧提取技术的红外成像呼吸信号检测方法

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US11734955B2 (en) * 2017-09-18 2023-08-22 Board Of Trustees Of Michigan State University Disentangled representation learning generative adversarial network for pose-invariant face recognition
CN110647973A (zh) * 2018-06-27 2020-01-03 北京中科寒武纪科技有限公司 运算方法及相关方法和产品
WO2021112335A1 (ko) * 2019-12-06 2021-06-10 주식회사 애자일소다 생성적 적대 신경망 기반의 분류 시스템 및 방법

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5704367A (en) * 1995-03-28 1998-01-06 Nihon Kohden Corporation Respiration monitor for monitoring respiration based upon an image signal of a facial region
CN101489478A (zh) * 2006-06-01 2009-07-22 必安康医疗有限公司 用于监视生理症状的装置、系统和方法
CN102973273A (zh) * 2012-11-29 2013-03-20 中国人民解放军第四军医大学 一种基于红外辐射检测的睡眠呼吸功能监测系统
CN104834946A (zh) * 2015-04-09 2015-08-12 清华大学 一种非接触式睡眠监测方法及系统
CN106236013A (zh) * 2016-06-22 2016-12-21 京东方科技集团股份有限公司 一种睡眠监测方法及装置
CN107280673A (zh) * 2017-06-02 2017-10-24 南京理工大学 一种基于关键帧提取技术的红外成像呼吸信号检测方法

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113610153A (zh) * 2021-08-06 2021-11-05 长沙理工大学 人体红外图像识别方法、装置、计算机设备及存储介质
CN114699040A (zh) * 2022-02-21 2022-07-05 华南师范大学 基于生理信号的醒睡检测方法、装置、设备以及存储介质
CN117351049A (zh) * 2023-12-04 2024-01-05 四川金信石信息技术有限公司 热成像与可见光融合的测点配准引导方法、设备和介质
CN117351049B (zh) * 2023-12-04 2024-02-13 四川金信石信息技术有限公司 热成像与可见光融合的测点配准引导方法、设备和介质

Also Published As

Publication number Publication date
US20230022206A1 (en) 2023-01-26

Similar Documents

Publication Publication Date Title
WO2021120007A1 (zh) 一种基于红外图像序列的睡眠质量评估系统和方法
CN107392109A (zh) 一种基于深度神经网络的新生儿疼痛表情识别方法
Hasan et al. SmartHeLP: Smartphone-based hemoglobin level prediction using an artificial neural network
Muhaba et al. Automatic skin disease diagnosis using deep learning from clinical image and patient information
WO2021068781A1 (zh) 一种疲劳状态识别方法、装置和设备
Yang Medical multimedia big data analysis modeling based on DBN algorithm
CN104636580A (zh) 一种基于人脸的健康监控手机
CN117542474A (zh) 基于大数据的远程护理监测系统及方法
CN112509696A (zh) 基于卷积自编码器高斯混合模型的健康数据检测方法
CN111954250A (zh) 一种轻量级Wi-Fi行为感知方法和系统
US11244456B2 (en) System and method for image segmentation and digital analysis for clinical trial scoring in skin disease
Joshi et al. Deep learning based person authentication using hand radiographs: A forensic approach
Hajabdollahi et al. Segmentation of bleeding regions in wireless capsule endoscopy images an approach for inside capsule video summarization
CN111063438B (zh) 一种基于红外图像序列的睡眠质量评估系统和方法
Elshwemy et al. A New Approach for Thermal Vision based Fall Detection Using Residual Autoencoder.
CN116469148A (zh) 一种基于面部结构识别的概率预测系统及预测方法
Bharathi et al. Diabetes diagnostic method based on tongue image classification using machine learning algorithms
Soundrapandiyan et al. AI-based wavelet and stacked deep learning architecture for detecting coronavirus (COVID-19) from chest X-ray images
Ramkumar et al. Attention induced multi-head convolutional neural network organization with MobileNetv1 transfer learning and COVID-19 diagnosis using jellyfish search optimization process on chest X-ray images
Arai et al. Human gait gender classification using 2D discrete wavelet transforms energy
CN111700592A (zh) 一种癫痫脑电自动分类模型的获取方法、系统及分类系统
CN114694234B (zh) 情绪识别方法、系统、电子设备和存储介质
Oz et al. Efficacy of biophysiological measurements at FTFPs for facial expression classification: A validation
CN111798455A (zh) 一种基于全卷积密集空洞网络的甲状腺结节实时分割方法
Yudhana et al. Glucose Content Analysis using Image Processing and Machine Learning Techniques

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 19956493

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 19956493

Country of ref document: EP

Kind code of ref document: A1

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 16.01.2023)

122 Ep: pct application non-entry in european phase

Ref document number: 19956493

Country of ref document: EP

Kind code of ref document: A1