Face micro-expression-based abnormal individual recognition analysis method

Abstract

The invention discloses an abnormal individual recognition and analysis method based on face micro-expressions, which comprises the following steps: acquiring image samples of various types of micro-expressions, including basic micro-expressions of tags of happiness, surprise, sadness, anger, aversion and fear, and performing preprocessing; extracting textural features of a face micro-expression sequence from the preprocessed image sample; constructing a CNN network model based on an attention mechanism by taking the texture features as input and the types of the micro expressions as expected output, and performing training; and optimizing image data needing micro-expression recognition, and inputting the optimized image data into the trained CNN network model based on the attention mechanism to obtain a corresponding abnormal individual recognition result. According to the method, on the basis of the CNN network model of the attention mechanism, the hidden layer of the CNN is optimized by using the self-attention mechanism, so that the precision and generalization ability of the model can be improved, and the hidden micro-expressions in the human face can be better recognized.

Description

Technical Field

[0001] The present invention relates to the technical field of facial micro-expression recognition, and in particular to an abnormal individual recognition and analysis method based on facial micro-expression. Background Art

[0002] Microexpressions are a term in psychology, and are short-lived facial expressions that humans make unconsciously when trying to hide an emotion. They correspond to seven universal emotions: disgust, anger, fear, sadness, happiness, surprise, and contempt. Microexpressions last only 1 / 25 to 1 / 5 of a second, and express the true emotion that a person is trying to suppress and hide. Although a subconscious expression may only last for a moment, it sometimes expresses the opposite emotion.

[0003] Research on micro-expressions has been applied to national security, the judicial system, and medical clinics. In the field of national security, some well-trained terrorists and other dangerous people may easily pass the polygraph test, but through micro-expressions, their true expressions under the false surface can generally be discovered, and because of this feature of micro-expressions, it also has good applications in the judicial system and medical clinics. Film producers, directors, or advertising producers can also predict the revenue of promotional films or advertisements by sampling the micro-expressions of people when they watch promotional films or advertisements.

[0004] In short, with the advancement of science and technology and the continuous development of psychology, the research on facial expressions will become more and more in-depth, the content will become richer and richer, and the application will become more and more extensive. At present, there are also related technologies for the research and recognition of micro-expressions, but the overall accuracy is not very high, and there is still a lot of room for improvement.

[0005] However, since there are few micro-expression databases available for research and the resolution is not high, the development of micro-expression recognition technology is restricted. Therefore, how to improve the accuracy of automatically identifying micro-expressions hidden in the face based on the current small number of samples has become a problem that needs to be solved urgently. Summary of the invention

[0006] The purpose of the present invention is to provide a method for identifying and analyzing abnormal individuals based on facial micro-expressions, which solves the problem of improving the accuracy of identifying micro-expressions hidden in the face even when there are few samples.

[0007] To achieve the above object, the technical solution adopted by the present invention is:

[0008] The present invention provides a method for abnormal individual recognition and analysis based on facial micro-expressions, comprising the following steps:

[0009] S10, obtaining image samples of various types of micro-expressions, including basic micro-expressions with labels of happiness, surprise, sadness, anger, disgust and fear, and performing pre-processing;

[0010] S20, extracting texture features of facial micro-expression sequences from the preprocessed image samples;

[0011] S30, taking the texture feature as input and the type of micro-expression as expected output, constructing a CNN network model based on an attention mechanism, and performing training;

[0012] S40, after optimizing the image data for micro-expression recognition, input it into the trained CNN network model based on the attention mechanism to obtain the corresponding abnormal individual recognition result.

[0013] Furthermore, the preprocessing in step S10 includes:

[0014] 1) Using the Adaboost algorithm to detect the face in the micro-expression image and crop it;

[0015] 2) Using median filtering to denoise the cropped facial micro-expression image area;

[0016] 3) Combine SMOTE algorithm and / or MoEx algorithm to perform data enhancement;

[0017] 4) The image after data enhancement is normalized using a bilinear interpolation algorithm.

[0018] Furthermore, data enhancement combined with the SMOTE algorithm includes:

[0019] a. For each micro-expression image sample X i , calculate the Euclidean distance with similar samples and determine its K nearest neighbor samples of the same type;

[0020] b. From this sample X i Randomly select a sample X from the K nearest neighbors ik , generate new samples;

[0021] c. Repeat step b for N iterations to synthesize N new samples; enhancement methods include: scaling, mirror flipping, Gaussian noise, color jittering, and image rotation.

[0022] Furthermore, the MoEx algorithm performs data enhancement, including:

[0023] For each micro-expression image sample X i The samples are input into the CNN neural network, mapped into low-dimensional vectors of the network layer, and combined transformation is performed directly in the learned feature space for data enhancement.

[0024] Furthermore, the step S20 includes:

[0025] The dynamic spatiotemporal texture features of facial micro-expression sequences are extracted using the HLACLF-TOP algorithm;

[0026] Or extract facial micro-expression and micro-expression texture features based on the LBP-TOP method.

[0027] Furthermore, the HLACLF-TOP algorithm is used to extract the dynamic spatiotemporal texture features of facial micro-expression sequences, including:

[0028] The HLACLF-TOP feature is calculated by applying the HLAC mask to the image sequence, and the HLACLF vector of the XY plane, the HLACLF vectors of the XT plane and the YT plane, and the HLACLF-TOP feature vector of the entire facial micro-expression image sequence are calculated in turn.

[0029] Furthermore, based on the LBP-TOP method, facial micro-expression texture features are extracted, including:

[0030] The LBP-TOP algorithm is used to extract the spatiotemporal texture features of the facial micro-expression image sequence with labels. Eight frames of images that effectively describe the facial micro-expression sequence are used to divide them into 16×16 non-overlapping blocks. The LBP-TOP features are extracted from each sub-block, and the histogram statistics of the LBP-TOP features are performed on the sub-block.

[0031] The feature histograms of all sub-blocks are concatenated into the feature histogram of the entire facial micro-expression sequence, thereby extracting the dynamic spatiotemporal texture features of the facial micro-expression sequence, that is, obtaining the deformation information of the facial micro-expression sequence in the XY plane, as well as the motion information in the XT plane and YT plane.

[0032] Furthermore, in step S30, the CNN network model based on the attention mechanism includes: based on a feature learning module and an attention module; using the gradient descent optimization algorithm Adam, selecting MSE as the loss function, setting the CNN-SA hidden layer to 2 layers, the hidden layer neurons to 6, the CNN layer to 1 layer, and the CNN layer is followed by an SA layer and a fully connected layer.

[0033] Furthermore, the image data required for micro-expression recognition is optimized, including:

[0034] 1) Using the Adaboost algorithm to detect the face in the micro-expression image and crop it;

[0035] 2) Using median filtering to denoise the cropped facial micro-expression image area;

[0036] 3) Use bilinear interpolation algorithm to achieve image size normalization.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] The embodiment of the present invention provides an abnormal individual recognition and analysis method based on facial micro-expressions, including: obtaining image samples of various types of micro-expressions, including basic micro-expressions with labels of happiness, surprise, sadness, anger, disgust and fear, and preprocessing; extracting texture features of facial micro-expression sequences from the preprocessed image samples; using the texture features as input and the type of micro-expressions as expected output, constructing a CNN network model based on an attention mechanism, and training it; optimizing the image data that needs to be recognized for micro-expressions, and inputting it into the trained CNN network model based on the attention mechanism to obtain the corresponding abnormal individual recognition results. The data set is enriched in a variety of ways, and is cropped, filtered, data enhanced and normalized to improve the generalization ability and accuracy of the subsequent detection model. In addition, the CNN network model based on the attention mechanism uses a self-attention mechanism to optimize the hidden layer of the CNN, which can also improve the accuracy and generalization ability of the model, and can better recognize the hidden micro-expressions in the face. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 A flowchart of an abnormal individual recognition and analysis method based on facial micro-expressions provided by an embodiment of the present invention;

[0040] Figure 2 Provides an effect diagram of the sample pre-processed by S10 for an embodiment of the present invention;

[0041] Figure 3 Schematic diagram of LBP-TOP feature composition;

[0042] Figure 4 A structural diagram of a CNN network model of an attention mechanism provided in an embodiment of the present invention;

[0043] Figure 5 A structural diagram of an attention module provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0044] In order to make the technical means, creative features, objectives and effects achieved by the present invention easy to understand, the present invention is further explained below in conjunction with specific implementation methods.

[0045] In the description of the present invention, it should be noted that the terms "upper", "lower", "inner", "outer", "front end", "rear end", "two ends", "one end", "the other end" and the like indicate positions or positional relationships based on the positions or positional relationships shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first" and "second" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance.

[0046] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "provided with", "connected", etc. should be understood in a broad sense. For example, "connected" can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0047] Reference Figure 1 As shown, the present invention provides a method for identifying and analyzing abnormal individuals based on facial micro-expressions, comprising the following steps:

[0048] S10, obtaining image samples of various types of micro-expressions, including basic micro-expressions with labels of happiness, surprise, sadness, anger, disgust and fear, and performing pre-processing;

[0049] S20, extracting texture features of facial micro-expression sequences from the preprocessed image samples;

[0050] S30, taking the texture feature as input and the type of micro-expression as expected output, constructing a CNN network model based on an attention mechanism, and performing training;

[0051] S40, after optimizing the image data for micro-expression recognition, input it into the trained CNN network model based on the attention mechanism to obtain the corresponding abnormal individual recognition result.

[0052] This method enriches the data set in a variety of ways, and performs cropping, filtering, data enhancement and normalization on it to improve the generalization ability and accuracy of the subsequent detection model. In addition, the CNN network model based on the attention mechanism uses the self-attention mechanism to optimize the hidden layer of CNN, which can also improve the accuracy and generalization ability of the model and better identify the hidden micro-expressions in the face.

[0053] The following is a detailed description of each of the above steps:

[0054] In step S10, when obtaining image samples of various types of micro-expressions, it includes extracting through real-time video capture, and image resources can also be downloaded from the Internet.

[0055] For example, a high-speed camera is pointed at the face of the subject, and the computer is allowed to store the images frame by frame. Before collecting data, micro-expressions need to be induced, so the subject should be shown some stimuli that have been proven by psychologists. These small stimuli last for a very short time, and the subject's attention must be focused. Before they make expressions, the subject needs to maintain a calm mood, which is equivalent to no muscle changes on the face. Once stimulated, there will be some stimulation on the face. The subject is asked to describe what kind of emotion he was when stimulated, and then the picture can be observed frame by frame to form a collected data set.

[0056] After collecting the data, the next step is to label the data set. Because we need to label the emotions, we need to know which AUs have changed so that we can label the changed AUs. Micro-expressions are a continuous concept over time, so we also need to label the start, end, and climax of the micro-expression.

[0057] Of course, you can also use WordNet to download 1,000,000 facial expression images and related emotion keywords from the Internet. Then use the algorithm to automatically annotate these images with AU, AU intensity, and emotion category, and you can get a very useful database.

[0058] Among them, the preprocessing method and execution order are as follows:

[0059] 1) Use the Adaboost algorithm to detect and crop the face in the micro-expression image; the Adaboost algorithm is a classification method that combines some relatively weak classification methods to form a new and strong classification method. In the face detection process, the Adaboost algorithm is used to select some rectangular features (weak classifiers) that best represent the face, and the weak classifier is constructed into a strong classifier by weighted voting. Then, several trained strong classifiers are connected in series to form a cascaded classifier, which effectively improves the detection speed of the classifier.

[0060] 2) Use median filtering to denoise the cropped facial micro-expression image area; after cropping, perform filtering and denoising again. First, determine a neighborhood with a cropped pixel as the center point. It is usually a square neighborhood, but it can also be a circle, a cross, etc. Then sort the grayscale values ​​of each pixel in the neighborhood, and take the middle value as the new value of the grayscale of the center pixel. The area here is called a window. When the window moves, the median filter can be used to smooth the image. The algorithm is simple and the time complexity is low. The processing effects of the above two steps are as follows: Figure 2 shown.

[0061] 3) Combine SMOTE algorithm and / or MoEx algorithm to perform data enhancement;

[0062] Data enhancement combined with the SMOTE algorithm includes:

[0063] a. For each micro-expression image sample X i , calculate the Euclidean distance with similar samples and determine its K nearest neighbor samples of the same type;

[0064] b. From this sample X i Randomly select a sample X from the K nearest neighbors ik , generate new samples;

[0065] c. Repeat step b for N iterations to synthesize N new samples; enhancement methods include: scaling, mirror flipping, Gaussian noise, color jittering, and image rotation.

[0066] MoEx algorithm performs data enhancement, including:

[0067] For each micro-expression image sample X i The samples are input into the CNN neural network, mapped into low-dimensional vectors of the network layer, and combined transformation is performed directly in the learned feature space for data enhancement.

[0068] You can also choose data augmentation for generative models or data augmentation based on neural style transfer just to expand the dataset.

[0069] 4) The image after data enhancement is normalized using a bilinear interpolation algorithm.

[0070] For example, the size of a facial micro-expression image after facial micro-expression image preprocessing is 150×150 pixels.

[0071] The present invention obtains image samples of various types of micro-expressions, including the collection of video frames, collection from the network and combination with the existing micro-expression database to form a new micro-expression sample data set; and performs data enhancement processing on it to further expand the sample set so as to improve the generalization ability and accuracy of the subsequent detection model.

[0072] In step S20, there are two extraction methods:

[0073] 1. Use the HLACLF-TOP algorithm to extract dynamic spatiotemporal texture features of facial micro-expression sequences;

[0074] Step (1). Calculate the HLACLF-TOP feature by applying the HLAC mask to the image sequence. The HLACLF vector of the facial micro-expression image sequence in the XY plane is obtained by applying the HLAC mask to each frame in the facial micro-expression sequence, obtaining an HLACLF vector, connecting these feature vectors in series, obtaining the feature vector of the sequence in the current plane, and calculating the HLACLF vector of the XY plane.

[0075] Step (2). Calculate the HLACLF vectors of the XT plane and the YT plane: The T axis in the XT plane and the YT plane is the time axis, which is perpendicular to the Y axis and the X axis respectively. The HLAC mask is improved into 6 1×3 masks. These 6 masks contain three 0-order masks and three 1-order masks respectively. The 0-order mask has only one position as 1, and the 1-order mask has two positions as 1. These 6 1×3 masks are used for feature extraction on the XT and YT planes.

[0076] Step (3). Calculate the HLACLF-TOP feature vector of the entire facial micro-expression image sequence: concatenate the feature vectors of the XY, XT and YT planes obtained in the above steps (1) and (2) to obtain 25+6+6 vectors of length m×n as the HLACLF-TOP feature vector of each facial micro-expression video sequence, that is, concatenate 37 vectors as the HLACLF-TOP features of each facial micro-expression video sequence, thereby using the HLACLF-TOP algorithm to extract the dynamic spatiotemporal texture features of the facial micro-expression sequence.

[0077] This extraction method extracts the dynamic spatiotemporal texture features of the facial micro-expression sequence from different frequencies and directions. At the same time, according to the differences in the three planes, the number of masks used on the three planes is different, which can more closely decompose the facial micro-expression image.

[0078] 2. Extract facial micro-expression and micro-expression texture features based on LBP-TOP method:

[0079] The LBP-TOP algorithm is used to extract the spatiotemporal texture features of the facial micro-expression image sequence marked in the second step. The 8 frames of images that effectively describe the facial micro-expression sequence obtained in the second step are divided into 16×16 non-overlapping blocks. The LBP-TOP features are extracted on each sub-block, and the histogram statistics of the LBP-TOP features are performed on the sub-block. Finally, the feature histograms of all sub-blocks are connected in series to form the feature histogram of the entire facial micro-expression sequence, thereby extracting the dynamic spatiotemporal texture features of the facial micro-expression sequence, that is, obtaining the deformation information of the facial micro-expression sequence in the XY plane, as well as the motion information in the XT plane and YT plane.

[0080] The LBP-TOP algorithm is used to extract facial micro-expression sequence features, taking into account the role of the central pixel, significantly reducing the dimension of the histogram, and reducing the impact of white noise on the recognition results.

[0081] For dynamic image sequences, three time-space axes T, X, and Y are set, and the XY plane contains the texture information of each frame of the image, and the XT and YT planes contain the changes in the spatial position of the image sequence over time. The LBP-TOP operator calculates the LBP value on the three orthogonal planes XY, XT, and YT, and then connects the LBP histograms of the three planes in series in the order of XY, XT, and YT to form the LBP-TOP feature, such as Figure 3 shown.

[0082] In step S30, a CNN network model based on the attention mechanism is used, such as Figure 4 As shown, it includes: feature learning module and attention module; based on the feature learning module, it automatically pays attention to the tiny muscle changes in different local areas of the face when people make expressions. At the same time, an associated attention module is used to pay attention to the connection between the changes in different local face areas, and the salient feature map containing the overall face information is passed to the output. In addition, the gradient descent optimization algorithm Adam is used, the loss function is selected as MSE, the hidden layer is set to 2 layers, the hidden layer neurons are 6, the CNN layer is 1 layer, and the CNN layer is followed by the SA layer and the fully connected layer.

[0083] The above basic feature learning module first performs convolution operation on the input features:

[0084] X 0 =(W c *X)+b (1)

[0085] Among them, X is the input of the convolutional layer, W c is the convolution kernel, * represents the convolution operation, and b is the bias term of the convolution operation. 0 Use a nonlinear activation function to get the output of this layer. You can use a rectified linear unit as the activation function as follows:

[0086]

[0087] Among them, x 0 Yes X 0 The convolutional neural network repeats this operation to obtain the final basic convolution graph. ResNet adds skip layer connections on this basis.

[0088] Attention module, the structure is as follows Figure 5 As shown:

[0089] When an expression is produced on a human face, the facial muscle movements in different local areas of the face usually have different patterns and features. For subtle facial expressions, this muscle movement will be relatively weak, so its corresponding pattern is not obvious. For such a weak pattern, if the convolutional network is used directly for feature extraction, it may not be able to capture useful semantic information well, and the attention mechanism can amplify the significant useful information and filter out the non-significant irrelevant information. Therefore, the present invention considers using the attention mechanism in the spatial domain to assign different weights to different facial areas, so that subsequent operations can focus on weak local patterns. However, the inconspicuous features in different local areas may be difficult to learn through a single attention module. Since the patterns in a local area of ​​the face are basically the same, while the patterns in different local areas are different, based on this observation, multiple local attention modules can be used, so that each local attention module can only focus on a single pattern in a local part of a human face. This operation can improve the network's ability to extract discriminative information and reduce the difficulty of network learning.

[0090] Therefore, the local feature extraction module of the embodiment of the present invention includes M attention masks, each of which focuses on a specific significant face region. For example, a certain attention mask may be more sensitive to frowned eyebrows, while another attention mask may focus more on squinting eyes.

[0091] The structure of the attention module is as follows Figure 3 As shown, the basic feature map F∈R in the basic feature learning module H×W×C As input, H, W, C represent the height, width and number of channels of the feature map respectively. This module obtains a set of attention masks by acting on the basic feature map with a 1×1 convolution kernel. Each mask focuses on a specific salient face area, and the corresponding local attention map F is obtained by multiplying the attention mask with the basic feature map point by point. A ∈R H×W×MThat is to say, each attention mask is a filter in the spatial domain, which assigns a larger weight to the corresponding salient local area in the basic feature map, and assigns a smaller weight or no weight to other parts. Through this operation, the local attention feature map finally obtained by each attention submodule only contains the corresponding local saliency discrimination information.

[0092] The texture features extracted in step S20 are used as input, and the type of micro-expression is used as the expected output. A CNN network model based on the attention mechanism is constructed and trained.

[0093] In step S40, the image data to be used for micro-expression recognition is optimized, including:

[0094] 1) Using the Adaboost algorithm to detect the face in the micro-expression image and crop it;

[0095] 2) Using median filtering to denoise the cropped facial micro-expression image area;

[0096] 3) Use bilinear interpolation algorithm to achieve image size normalization.

[0097] Finally, the data is input into the trained CNN network model based on the attention mechanism to obtain the corresponding abnormal individual recognition results.

[0098] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A method for abnormal individual recognition and analysis based on facial micro-expressions, characterized in that: The following steps are involved: S10, obtaining image samples of various types of micro-expressions, including basic micro-expressions with labels of happiness, surprise, sadness, anger, disgust and fear, and performing pre-processing; S20, extracting texture features of facial micro-expression sequences from the preprocessed image samples; S30, taking the texture feature as input and the type of micro-expression as expected output, constructing a CNN network model based on an attention mechanism, and performing training; S40, after optimizing the image data for micro-expression recognition, input it into the trained CNN network model based on the attention mechanism to obtain the corresponding abnormal individual recognition result.

2. The abnormal individual recognition and analysis method based on facial micro-expressions according to claim 1 is characterized in that: The pre-processing in step S10 includes: 1) Using the Adaboost algorithm to detect the face in the micro-expression image and crop it; 2) Using median filtering to denoise the cropped facial micro-expression image area; 3) Combine SMOTE algorithm and / or MoEx algorithm to perform data enhancement; 4) The image after data enhancement is normalized using a bilinear interpolation algorithm.

3. The abnormal individual recognition and analysis method based on facial micro-expressions according to claim 2 is characterized in that: Data enhancement combined with the SMOTE algorithm includes: a. For each micro-expression image sample X i , calculate the Euclidean distance with similar samples and determine its K nearest neighbor samples of the same type; b. From this sample X i Randomly select a sample X from the K nearest neighbors ik , generate new samples; c. Repeat step b for N iterations to synthesize N new samples; enhancement methods include: scaling, mirror flipping, Gaussian noise, color jittering, and image rotation.

4. The abnormal individual recognition and analysis method based on facial micro-expressions according to claim 2 is characterized in that: The MoEx algorithm performs data enhancement, including: For each micro-expression image sample X i The samples are input into the CNN neural network, mapped into low-dimensional vectors of the network layer, and combined transformation is performed directly in the learned feature space for data enhancement.

5. The abnormal individual recognition and analysis method based on facial micro-expressions according to claim 1 is characterized in that: The step S20 comprises: The dynamic spatiotemporal texture features of facial micro-expression sequences are extracted using the HLACLF-TOP algorithm; Or extract facial micro-expression and micro-expression texture features based on the LBP-TOP method.

6. The abnormal individual recognition and analysis method based on facial micro-expressions according to claim 5 is characterized in that: The HLACLF-TOP algorithm is used to extract dynamic spatiotemporal texture features of facial micro-expression sequences, including: The HLACLF-TOP feature is calculated by applying the HLAC mask to the image sequence, and the HLACLF vector of the XY plane, the HLACLF vectors of the XT plane and the YT plane, and the HLACLF-TOP feature vector of the entire facial micro-expression image sequence are calculated in turn.

7. The abnormal individual recognition and analysis method based on facial micro-expressions according to claim 5 is characterized in that: Extract facial micro-expression texture features based on LBP-TOP method, including: The LBP-TOP algorithm is used to extract the spatiotemporal texture features of the facial micro-expression image sequence with labels. Eight frames of images that effectively describe the facial micro-expression sequence are used to divide them into 16×16 non-overlapping blocks. The LBP-TOP features are extracted from each sub-block, and the histogram statistics of the LBP-TOP features are performed on the sub-block. The feature histograms of all sub-blocks are concatenated into the feature histogram of the entire facial micro-expression sequence, thereby extracting the dynamic spatiotemporal texture features of the facial micro-expression sequence, that is, obtaining the deformation information of the facial micro-expression sequence in the XY plane, as well as the motion information in the XT plane and YT plane.

8. The abnormal individual recognition and analysis method based on facial micro-expressions according to claim 1 is characterized in that: In step S30, the CNN network model based on the attention mechanism includes: based on a feature learning module and an attention module; using the gradient descent optimization algorithm Adam, selecting MSE as the loss function, setting the CNN-SA hidden layer to 2 layers, the hidden layer neurons to 6, the CNN layer to 1 layer, and the CNN layer is followed by an SA layer and a fully connected layer.

9. The abnormal individual recognition and analysis method based on facial micro-expressions according to claim 2 is characterized in that: Optimize the image data required for micro-expression recognition, including: 1) Using the Adaboost algorithm to detect the face in the micro-expression image and crop it; 2) Using median filtering to denoise the cropped facial micro-expression image area; 3) Use bilinear interpolation algorithm to achieve image size normalization.

Images