A facial emotion management method based on cloud human-computer interaction technology
Through cloud human-computer interaction technology, combined with WeChat server and user permissions, an emotion recognition model is built, which solves the problem of insufficient application of facial emotions classification in the existing technology, realizes accurate recognition and tracking of user emotions, and improves user interaction experience and Mental health assistance.
Patent Information
- Application Number
- CN202111637282.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-12-29
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2041-12-29
AI Technical Summary
In the existing technology, commercial face emotion classification application is insufficient, and emotions recognition and tracking cannot be achieved from the user's perspective. The way to obtain face emotions is active acquisition, and the user's permission is not directly collected. The classification accuracy is based on the developer's perspective. Not from the user's perspective.
Adopt face emotion management method based on cloud human-computer interaction technology, obtain user permissions through WeChat servers, record and transmit video streams to cloud servers, establish emotional tags, build user personal machine learning models, realize emotion recognition and tracking, and reduce user Adverse effects of interactive experience.
While ensuring accuracy while reducing the impact of user interaction experience, users can monitor emotional changes anytime and anywhere, relying on historical data to trace emotional information, providing auxiliary basis for users' mental health.
Smart Images

Figure CN114612957B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of communication software interaction technology, and in particular to a facial emotion management method based on cloud-based human-computer interaction technology. Background Art
[0002] Human emotions are divided into 7 basic expressions, such as indifferent, angry, hate, fear, happy, sad, and surprised. At present, the recognition accuracy of these 7 basic emotions is >85%, but it is not widely used in the interactive field. For example, CN202010741026.7 discloses a big data analysis system based on face recognition ReID technology, including a data analysis module, an operator mobile phone data module, and a face recognition module. The data analysis module includes a detection module, a comparison module, a database module, and a tracking module; the operator mobile phone data module includes an association module, an identification module, and a terminal information library module. Compared with the prior art, the advantages of the present invention are: intelligent big data analysis is performed by using big data collection and cleaning, model algorithms, machine learning, etc., to connect customers' accounts at all levels in the store, and provide intelligent data analysis and visualization related to passenger flow and payment for commercial entities, helping shopping centers gradually build complete user portraits, fully recognize customers, continuously optimize operations, improve efficiency, and transform the value of operating data into operating value and revenue. But the problem is that the face acquisition method is active, and the user's permission is not obtained before acquisition. At the same time, its own operating framework is based on big data mining. In fact, whether the customer really likes it is not classified and organized. The accuracy of the classification is based on the developer's perspective rather than the user's perspective, and it is impossible to trace emotional information.
[0003] Therefore, the applicant proposed a facial emotion management method based on cloud-based human-computer interaction technology, which, from the user's perspective, solves the problem of insufficient application of commercial facial emotion classification in the existing technology, introduces emotion recognition technology into communication software, and realizes the recognition and tracking of user emotions. Summary of the invention
[0004] (I) Technical solution
[0005] The present invention is implemented through the following technical solutions: a facial emotion management method based on cloud human-computer interaction technology,
[0006] S1: The user connects to the WeChat server through a terminal device to obtain first interaction information;
[0007] S2: Send a request for obtaining permission information to the terminal device based on the first interaction information; if permitted, proceed to S3; otherwise, return to S1;
[0008] S3: sending at least one second interaction information to the user based on the terminal device, recording the first video stream based on the second interaction information, transmitting the video stream to the cloud server and establishing an emotion label for the video stream through the second interaction information; classifying the video stream with the emotion label to establish a training set, and establishing a user personal machine learning model using the training set;
[0009] S4: Sending an emotion detection request based on the device terminal, and starting to record the second video stream if allowed; importing the personal machine learning model based on the second video stream to generate third interaction information;
[0010] S5: Send a request for obtaining the third interaction information based on the device terminal. If the third interaction information is allowed to be sent and the result is saved to the cloud server, the fourth interaction information is sent at the same time; the third interaction information is bound to the user ID; the user can check the historical third interaction information at any time through the device terminal.
[0011] As a further illustration of the above solution, the first interaction information includes WeChat public account interaction information and WeChat mini program interaction information.
[0012] As a further explanation of the above scheme, the permission information acquisition request includes a WeChat mobile phone number information acquisition request, a public information authorization acquisition request, and a device terminal camera authorization use request.
[0013] As a further illustration of the above solution, the second interactive information includes pattern information and sound information; each pattern information and sound information is attached with an emotion label.
[0014] As a further explanation of the above solution, S3 is specifically:
[0015] S31: The user determines whether to receive the second interaction information through the device terminal, and if yes, a 3s countdown is generated;
[0016] S32: after the 3s countdown ends, at least one second interaction information is displayed based on the device terminal, and at the same time, a first video stream of 10s recorded based on the device terminal is uploaded to the cloud server;
[0017] S33: using a video segmentation technology to segment the first video stream to generate at least one sub-video stream, and establishing an emotion label for the sub-video stream based on the second interactive information, and classifying the sub-video stream by the emotion label to establish a training set;
[0018] S34: Establish a user's personal machine learning model based on the training set.
[0019] As a further explanation of the above scheme, the S34 is specifically:
[0020] S341: dividing the sub video stream into a plurality of first frame images by frame, and classifying each frame image based on the emotion label;
[0021] S342: Preprocessing the first frame images by grayscale processing and size normalization processing, and performing data enhancement processing by rotation, cropping, and flipping to obtain a data set, and dividing the data set into a training set and a test set according to a ratio of 3:7;
[0022] S343: Extract the HOG features of the facial expressions in the frame image and input them into the SVM classifier for training to obtain a machine learning model.
[0023] As a further illustration of the above scheme, S4 is specifically:
[0024] S41: The user allows the emotion detection request, and the device terminal starts recording a second video stream for 10 seconds; the second video stream is divided into a plurality of second frame images;
[0025] S42: Preprocessing the plurality of second frame images by grayscale processing and size normalization processing,
[0026] S43: extract HOG features from the preprocessed image;
[0027] S44: Import the trained personal machine learning model and output the third interaction information.
[0028] As a further illustration of the above solution, the third interactive information includes classification probability and result report.
[0029] As a further explanation of the above scheme, S5 is specifically:
[0030] S51: The user sends a request for obtaining third interaction information based on a device terminal;
[0031] S52: Feedback third interaction information and fourth interaction information; the fourth interaction information includes: sharing incentive request and sharing assistance request;
[0032] S53: If the user accepts the fourth interaction information, the cloud feeds back the incentive result and the assistance result based on the fourth interaction information.
[0033] As a further illustration of the above scheme, the emotion labels include: a neutral label, an angry label, a hate label, a fear label, a happy label, a sad label, and a surprised label.
[0034] (III) Beneficial effects
[0035] Compared with the prior art, the present invention has the following beneficial effects: an emotion recognition method is established through a WeChat server, and facial expression analysis technology is used to recognize the facial expression of a user; a user facial expression recognition data set is constructed by collecting the user's expression data at a device terminal and emotion tags and voice tags; while ensuring the accuracy, the adverse impact on the user's interactive experience is reduced; through the method established by the present invention, the user can monitor the change of emotion anytime and anywhere, which is convenient for the user to adjust his or her own emotion in time, and at the same time, an emotion information data tracing mechanism is established based on historical data, and through the user's expression, online real-time detection and analysis of the emotion change are carried out, providing an auxiliary basis for the user's mental health. BRIEF DESCRIPTION OF THE DRAWINGS
[0036] Other features, objects and advantages of the novel invention will become more apparent from the detailed description of non-limiting embodiments with reference to the following drawings:
[0037] Figure 1 It is a schematic diagram of the flow chart of the invention method;
[0038] Figure 2 This is a schematic diagram of the interactive display of a mobile phone in step S1 of Example;
[0039] Figure 3 This is a schematic diagram of the interactive display of a mobile phone in step S1 of Example;
[0040] Figure 4 This is a schematic diagram of the interactive display of a mobile phone in step S1 of Example;
[0041] Figure 5 This is a schematic diagram of the interactive display of a mobile phone in step S2 of Example;
[0042] Figure 6 This is a schematic diagram of the interactive display of a mobile phone in step S2 of Example;
[0043] Figure 7 This is a schematic diagram of the interactive display of a mobile phone in step S2 of Example;
[0044] Figure 8 This is a schematic diagram of the mobile phone interactive display of the emotion detection start interface in the embodiment;
[0045] Fig. 9 Schematic diagram of human-computer interaction in step S51 of the embodiment;
[0046] Fig.10 Schematic diagram of human-computer interaction in step S52 of the embodiment. DETAILED DESCRIPTION
[0047] Example
[0048] See also Figure 1, a facial emotion management method based on cloud human-computer interaction technology,
[0049] S1: The user connects to the WeChat server through a terminal device to obtain first interaction information; the first interaction information includes WeChat public account interaction information and WeChat mini program interaction information. Figure 2 , Figure 3 , Figure 4 Specific users connect to the WeChat server through the device terminal, search for the "YueFan Artificial Intelligence" public account on WeChat and follow it, click the "Expression Experience" column to enter the mini program; or directly search for "YueFan Expression Recognition" in the WeChat mini program and click to enter directly;
[0050] S2: Send a permission information acquisition request to the terminal device based on the first interaction information; if allowed, enter S3; otherwise return to S1; the permission information acquisition request includes a WeChat mobile phone number information acquisition request, a public information authorization acquisition request, and a device terminal camera authorization use request. For details, please refer to Figure 5 , Figure 6 , Figure 7 , click "Register / Login", "One-click Authorization Login" and a pop-up will appear Figure 7 Authorize and allow the WeChat mobile phone number, public information, and mobile phone camera as shown. Click Allow to enter the start interface. The start interface is as follows Figure 8 shown.
[0051] The above part is the steps of obtaining WeChat public account and authorization information, which is the pre-step for commercial use of this embodiment and is the prior art, and will not be described in detail here. Next, the core part of the facial emotion management method will be explained.
[0052] S3: Send at least one second interactive message to the user based on the terminal device, record the first video stream based on the second interactive message, transmit the video stream to the cloud server and establish an emotional label for the video stream through the second interactive message; classify the video stream with the emotional label to establish a training set, and use the training set to establish a user's personal machine learning model; it should be further explained that the emotional labels include: bland label, angry label, hate label, fear label, happy label, sad label, and surprised label. The specific steps include the following:
[0053] S31: The user determines whether to receive the second interaction information through the device terminal. If yes, a 3s countdown is generated. After completing S2, the user will enter Figure 8interface; if the user "clicks to start", it is determined that the user accepts the second interactive information through the device terminal, and the second interactive information includes pattern information and sound information; each pattern information and sound information is attached with an emotion label. For example: happy, a happy icon is displayed and a happy sound is emitted, and the icon and sound are attached with a happy emotion label. At the same time, the user "clicks to start" and enters a 3s countdown, giving the user time to adjust his or her own state and buffer time for transmitting the second interactive information from the cloud;
[0054] S32: After the 3s countdown ends, at least one second interactive information is displayed based on the device terminal, and at the same time, the first video stream recorded for 10s based on the device terminal is uploaded to the cloud server; the second interactive information includes pattern information and sound information; each pattern information and sound information is attached with an emotion label. For example: happy, a happy icon is displayed and a happy sound is emitted, and the icon and sound are attached with a happy emotion label; sad, a sad icon is displayed and a sad sound is emitted, and the icon and sound are attached with a sad emotion label. It should be further explained that multiple second interaction information need to appear within the range of 10s to facilitate the collection of sufficient expression data, such as the second interaction information with a calm label appears for 1s, the second interaction information with an angry label appears for 1s, the second interaction information with a hate label appears for 1s, the second interaction information with a fear label appears for 1s, the second interaction information with a happy label appears for 2s, the second interaction information with a sad label appears for 2s, and the second interaction information with a surprise label appears for 2s; the video stream has 10 to 24 frames, then each second interaction information can obtain at least 10 to 24 images; a total of at least 100 to 240 images can be obtained; it should be further explained that the step of S32 can be repeated to obtain more data for subsequent model training.
[0055] S33: Use video segmentation technology to segment the first video stream to generate at least one sub-video stream, and establish emotion labels for the sub-video streams based on the second interactive information, and classify the sub-video streams through the emotion labels to establish a training set; it should be further explained that in order to facilitate subsequent data training, the emotion labels need to be divided first, such as the second interactive information with a calm label appears for 1s, the second interactive information with an angry label appears for 1s, the second interactive information with a hate label appears for 1s, the second interactive information with a fear label appears for 1s, the second interactive information with a happy label appears for 2s, the second interactive information with a sad label appears for 2s, and the second interactive information with a surprised label appears for 2s; then the entire 10-second video needs to be segmented into 1s, 1s, 1s, 1s, 2s, 2s, 2s; and at the same time, each video is classified according to the label.
[0056] S34: Establish a user's personal machine learning model based on the training set. Specifically:
[0057] S341: split the sub-video stream into several first frames by frame, and classify each frame based on the emotion label; it should be further explained that each frame in the video is extracted by video segmentation technology, and since the 10s video needs to be classified into 7 emotion label data, each frame image must be included in the data set to ensure sufficient data volume. At the same time, all images are classified according to the 7 emotion labels for subsequent processing.
[0058] S342: Preprocess several first frame images by grayscale processing and size normalization, and obtain data sets by rotating, cropping, and flipping to perform data enhancement processing, and divide the data sets into training sets and test sets according to 3:7; It should be further explained that the purpose of grayscale processing is that it is more complicated to process the data of three channels. The image is first grayscaled, and the grayscale process is to unify the RGB value of each pixel into the same value. The grayscaled image will change from three channels to a single channel, and the single channel data processing will improve the data processing efficiency and ensure that the recognition rate is greater than 85%. As the traffic and speed pursued by commercial interaction, although the three-channel processing can increase the recognition rate to more than 95%, the time consumed will also increase exponentially, which is not conducive to the efficiency of user interaction; the purpose of size normalization is to make the image size meet the input requirements of the classifier; the purpose of data enhancement is to further increase the amount of data, improve the generalization ability, and avoid overfitting.
[0059] S343: Extract the HOG features of the facial expressions in the frame image and input them into the SVM classifier for training to obtain a machine learning model. It should be further explained that the accuracy and time consumption of the HOG features and the SVM classifier reach a relatively balanced state, please refer to the following Table 1:
[0060] Table 1 Comparison of classification effects of feature and classifier combinations
[0061] Features and classification algorithms Test set accuracy time consuming HOG+SVM 85.73% 6.32s LBP+SVM 90.32% 50.89s Gabor+SVM 92.55% 287.33s CNN+CNN 95.67% 397.46s
[0062] From the above table, we can know that the combination of HOG features and SVM classifiers can achieve a relatively balanced state between accuracy and time consumption. As commercial interactions pursue traffic and speed, although other methods can increase the recognition rate to more than 90%, the time consumed also increases exponentially, which is not conducive to the efficiency of user interaction. What users need is to get a result immediately. Generally speaking, the time users are willing to wait for in an interactive interface is roughly in the range of 5-10s. Therefore, although other algorithms have high recognition rates, they actually have an adverse effect on the user interaction experience. At the same time, what this embodiment requires is the output probability of the classifier rather than the output result, so the combination of HOG features and SVM classifiers is the best choice for this embodiment. It should be further explained that the above combination is the preferred embodiment of this embodiment and is not a further limitation of the claims.
[0063] S4: Send an emotion detection request based on the device terminal, and start recording the second video stream if allowed; import the personal machine learning model based on the second video stream to generate third interaction information; S4 specifically:
[0064] S41: The user allows the emotion detection request, and the device terminal starts recording the second video stream for 10 seconds; the second video stream is divided into a number of second frames; it should be further explained that S3 is suitable for collecting user data for training, and it is in S4 that the user's emotion recognition in the current scene is considered. At the same time, when the device terminal records a 10-second video, the second interactive information will not be displayed. The purpose of the second interactive information is to guide the user to be in a certain emotional state in a certain time period, and this step requires the user's current emotional state.
[0065] S42: Pre-process several second frame images using grayscale processing and size normalization processing; it should be further explained that the purpose of grayscale processing is to process three-channel data, which is relatively complicated. The image is first grayscaled. The grayscale process is to unify the RGB value of each pixel into the same value. The grayscaled image will change from three channels to a single channel. Single-channel data processing improves data processing efficiency while ensuring a recognition rate greater than 85%. As commercial interactions pursue traffic and speed, although three-channel processing can increase the recognition rate to more than 95%, the time consumed also doubles, which is not conducive to the efficiency of user interaction; the purpose of size normalization is to make the image size meet the input requirements of the classifier;
[0066] S43: extract HOG features from the preprocessed image;
[0067] S44: Import the trained personal machine learning model and output the third interactive information. The third interactive information includes classification probability and result report. It should be further explained that this embodiment outputs the classification probability instead of the specific classification result. When creating the SVM classifier in step S3, probability=True is set in Scikit-Learn. After the training is completed, the algorithm will use logistic regression to calibrate the SVM score, and perform an additional 5-fold cross-validation training on the training data, so that the classification probability value can be output in step S4. At the same time, this embodiment can also introduce a kernel function in the SVM model so that the SVM model considers nonlinear spatial feature conversion, and finally outputs the classification probability in step S4. Or directly introduce a kernel function in the logistic regression so that the SVM model can consider the nonlinear spatial feature conversion while outputting the classification probability. The specific description is not expanded. The classification result generated at this time is the classification probability, such as happy 80%, bland 10%, surprised 10%, and the result report is generated through the result; such as happy 80%, bland 10%, surprised 10%, the result report displays the classification probability and displays "mainly happy in the current state". It should be further explained that the video recorded in step S4 will be returned to the data set for storage, so as to further improve the accuracy of the classifier.
[0068] S5: Send a request for obtaining the third interaction information based on the device terminal. If the third interaction information is allowed to be sent and the result is saved to the cloud server, the fourth interaction information is sent at the same time; the third interaction information is bound to the user ID; the user can check the historical third interaction information at any time through the device terminal.
[0069] The S5 is specific as follows:
[0070] S51: The user sends a request for obtaining the third interaction information based on the device terminal. Fig. 9 ; At this time, the user interface will display "Click to generate report". If clicked, it is considered that the user sends a request for acquisition, and the process goes to step S52;
[0071] S52: Feedback the third interaction information and the fourth interaction information; see Fig.10 The fourth interactive information includes: sharing incentive request, sharing assistance request; the third interactive information includes classification probability and result report. In this embodiment, the fourth interactive information is used to share the results with friends using the sharing function of WeChat.
[0072] S53: If the user accepts the fourth interactive information, the cloud will feedback the incentive result and the assistance result based on the fourth interactive information. The purpose of this step is to encourage the user to share the report with friends, so that friends can assist the user. If the assistance is successful, the user can receive a red envelope; the user assistance can be registration or testing. The purpose of this step is to establish a traffic channel, promote based on the user, and broaden the traffic channel.
[0073] It should be further explained that after users connect to the WeChat server through the user interface, they can check their mood changes and historical records at any time in the personal records of the user interface, analyze future trends, and adjust their mood and health status in time. Please refer to Figure 5 , making emotional information data traceable.
[0074] The above shows and describes the basic principles and main features of the present invention and the advantages of the present invention. It is obvious to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and the present invention can be implemented in other specific forms without departing from the spirit or basic features of the present invention. Therefore, no matter from which point of view, the embodiments should be regarded as exemplary and non-restrictive. The scope of the present invention is defined by the attached claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims are included in the present invention. Any figure mark in the claims should not be regarded as limiting the claims involved.
[0075] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.
Claims
1. A facial emotion management method based on cloud human-computer interaction technology, characterized in that: S1: A user connects to a WeChat server through a terminal device to obtain first interaction information; the first interaction information includes WeChat public account interaction information and WeChat mini-program interaction information; S2: Sending a request for obtaining permission information to the terminal device based on the first interaction information; If allowed, enter S3; Otherwise return to S1; S3: Send at least one second interaction information to the user based on the terminal device, record the first video stream based on the second interaction information, transmit the video stream to the cloud server and establish an emotion label for the video stream through the second interaction information; classify the video stream with the emotion label to establish a training set, and use the training set to establish a user's personal machine learning model; specifically, the following steps are included: S31: The user determines whether to receive the second interaction information through the device terminal, and if yes, a 3s countdown is generated; if yes, it is determined that the user accepts the second interaction information through the device terminal, and the second interaction information includes pattern information and sound information; Each pattern information and sound information is attached with an emotion label; the emotion labels include: bland label, angry label, hate label, fear label, happy label, sad label, and surprised label; S32: after the 3s countdown ends, at least one second interaction information is displayed based on the device terminal, and at the same time, a first video stream of 10s is recorded based on the device terminal and uploaded to the cloud server; the second interaction information includes pattern information and sound information; each pattern information and sound information is attached with an emotion label; multiple second interaction information needs to appear within the range of 10s, and the step of S32 can be repeated to obtain more data for subsequent model training; S33: using a video segmentation technology to segment the first video stream to generate at least one sub-video stream, and establishing an emotion label for the sub-video stream based on the second interactive information, and classifying the sub-video stream by the emotion label to establish a training set; S34: Establish a user's personal machine learning model based on the training set; specifically: S341: split the sub video stream into a number of first frame images by frame, and classify each frame image based on the emotion label; it should be further explained that each frame in the video is extracted by the video segmentation technology, and since the 10s video needs to be classified into data of 7 emotion labels, each frame image must be included in the data set to ensure sufficient data volume; at the same time, all images are classified according to the 7 emotion labels; S342: Preprocessing the first frame images by grayscale processing and size normalization processing, and performing data enhancement processing by rotation, cropping, and flipping to obtain a data set, and dividing the data set into a training set and a test set according to a ratio of 3:7; S343: extracting HOG features of facial expressions in the frame image and inputting them into the SVM classifier for training to obtain a machine learning model; S4: Sending an emotion detection request based on the device terminal, and starting to record the second video stream if allowed; importing the personal machine learning model based on the second video stream to generate third interaction information; S5: Send a request for obtaining the third interaction information based on the device terminal. If the third interaction information is allowed to be sent and the result is saved to the cloud server, the fourth interaction information is sent at the same time; the third interaction information is bound to the user ID; the user can check the historical third interaction information at any time through the device terminal.
2. According to claim 1, a method for managing facial emotions based on cloud-based human-computer interaction technology is characterized in that: The first interaction information includes WeChat public account interaction information and WeChat mini-program interaction information.
3. The method for managing facial emotions based on cloud-based human-computer interaction technology according to claim 1, characterized in that: The permission information acquisition request includes a WeChat mobile phone number information acquisition request, a public information authorization acquisition request, and a device terminal camera authorization use request.
4. The method for managing facial emotions based on cloud-based human-computer interaction technology according to claim 1, characterized in that: The second interactive information includes pattern information and sound information; each pattern information and sound information is attached with an emotion label.
5. The method for managing facial emotions based on cloud-based human-computer interaction technology according to claim 1, characterized in that: The S4 is specific as follows: S41: The user allows the emotion detection request, and the device terminal starts recording a second video stream for 10 seconds; the second video stream is divided into a plurality of second frame images; S42: pre-processing the plurality of second frame images by grayscale processing and size normalization processing; S43: extract HOG features from the preprocessed image; S44: Import the trained personal machine learning model and output the third interaction information.
6. The method for managing facial emotions based on cloud-based human-computer interaction technology according to claim 5, characterized in that: The third interactive information includes classification probability and result report.
7. The method for managing facial emotions based on cloud-based human-computer interaction technology according to claim 1, characterized in that: The S5 is specific as follows: S51: The user sends a request for obtaining third interaction information based on a device terminal; S52: Feedback third interaction information and fourth interaction information; the fourth interaction information includes: sharing incentive request and sharing assistance request; S53: If the user accepts the fourth interaction information, the cloud feeds back the incentive result and the assistance result based on the fourth interaction information.
8. A method for managing facial emotions based on cloud-based human-computer interaction technology according to claim 1, 4, 5 or 6, characterized in that: The emotion labels include: a neutral label, an angry label, a hate label, a fear label, a happy label, a sad label, and a surprised label.
Citation Information
Patent Citations
Big data analysis system based on face recognition Reid technology
CN111898522A
Man-machine interaction method and device, storage medium and terminal
CN108491519A
Facial expression recognition-based emotion recognition method and device
CN108564007A
A man-machine interaction method and system based on emotion recognition
CN109683709A
Facial expression recognition method and system based on deep privileged network
CN112906500A
