CN107085704A - Fast Facial Expression Recognition Method Based on ELM Autoencoding Algorithm - Google Patents

Abstract

The invention discloses a kind of fast face expression recognition method based on ELM own coding algorithms, it is as follows that the present invention includes step：The human face region detection grader of step 1, training based on Adaboost simultaneously carries out Face datection；2nd, the human face region detected is pre-processed, including cutting, size normalization and histogram equalization processing；3rd, feature extraction is carried out to pretreated Facial Expression Image as feature extraction algorithm using the ELM AE algorithms based on self-encoding encoder and the learning machine that transfinites combination；4th, the facial expression classifier based on the learning machine that transfinites is built, the vector of feature extraction is input in expression classifier, output result is the mood of this face.The present invention more can quickly and efficiently extract main information and dimensionality reduction.When Expression Recognition is classified, as long as ELM adjusts a parameter of neuron, identification run time is short, and accuracy rate is high, is a kind of efficient and fast algorithm of pace of learning.

Description

Fast Facial Expression Recognition Method Based on ELM Autoencoding Algorithm

technical field

The invention belongs to the field of image processing, describes the whole process of emotion recognition of human facial expressions, and in particular relates to a fast human facial expression recognition method based on an ELM self-encoding algorithm.

Background technique

Emotion recognition of human facial expressions, that is, recognizing human faces in pictures or videos and performing further emotional analysis, has become a hot spot in the field of biological intelligence feature recognition in the past few decades. Emotion recognition is essentially to give computers the ability to "observe words and colors" and improve the current relatively rigid and immature human-computer interaction environment.

The emotion analysis of facial expressions mainly includes the following aspects: face detection and positioning, image preprocessing, feature extraction of expressions, expression classification and emotion analysis. Face detection often refers to the detection of a face area in an image. If a face is included, the position of the face needs to be located and the size determined. This article uses the face detection method based on Adaboost, which is an iterative algorithm. , for a set of training sets, different training sets Si are obtained by changing the distribution probability of each sample, and each Si is trained to obtain a weak classifier Hi, and then these weak classifiers are classified according to different weights Combined, a strong classifier is obtained. The detected expression images usually have shortcomings such as noise and insufficient contrast, which are often caused by factors such as the intensity of light and the performance of the device. Therefore, preprocessing is a very important link in the process of facial emotion recognition, and effective preprocessing methods can help improve the recognition rate of facial expressions. After preprocessing, the feature extraction algorithm will be applied to the face image to extract the features of different expressions. This patent uses the ELM-AE algorithm combining autoencoder and extreme learning machine as the feature extraction method. This is an efficient and practical autoencoder algorithm. After the sample data is encoded and decoded, if the reconstruction error is small enough, within Within the range, it can be determined that this encoding code is an effective expression of the input sample data, and can be used as a description vector of facial image expression. Finally, realize expression recognition classification and further emotion analysis, build a recognition model training feature library based on the feature vector obtained from facial expression image feature extraction, and give the target category identification, including happiness, sadness, surprise, fear, anger, disgust and neutral.

Expression feature extraction is an important part of the entire facial expression recognition system. The traditional LBP, geometric feature-based and template-based feature extraction methods have certain defects. For example, the LBP feature extraction method is difficult to handle high-dimensional data and runs slowly. ; The adaptability of the method based on geometric features is not strong enough, and some information will be lost at the same time.

The essence of expression recognition is to design an efficient classifier to classify the target expression into one of the six basic expression categories or into neutral expressions according to the feature vector data extracted in the previous stage. It can be seen that the design of the classifier directly affects the final effect of expression recognition and emotion analysis. Due to the large amount of data after feature extraction, the traditional artificial neural network, template matching, support vector machine (SVM) and other methods are not running fast enough, and because of too many training samples and too long training time, they cannot meet real-time sexual needs. Therefore, this paper uses a classifier based on the Extreme Learning Machine (ELM) algorithm for fast expression recognition and classification. ELM is an easy-to-use and effective learning algorithm for single-hidden-layer feed-forward neural networks (SLFNs). It has the following advantages in expression classification: (1) ELM uses random weights between the input layer and the hidden layer. We can train the same dataset multiple times, which gives different output spaces for different classification accuracies. (2) ELM is a simpler learning algorithm for feedforward neural networks. Traditional neural network learning algorithms (such as BP algorithm) need to artificially set a large number of network training parameters, so it is very easy to generate local optimal solutions. In the process of determining network parameters, ELM only needs to set the number of hidden layer nodes of the network, and does not need to adjust the input weights of the network and the bias of hidden elements during the algorithm execution process, and generates a unique optimal solution. Therefore, ELM has faster learning speed and better generalization performance than traditional artificial neural networks, and can realize expression classification and emotion recognition the fastest.

The output of the ELM is Among them: β _i is the weight between the hidden layer node and the output node, G(a _i , _bi , x) is the hidden layer output function. h(x)=[G(a ₁ ,b ₁ ,x),...,G(a _L ,b _L ,x)] ^T is the output vector of the hidden layer relative to the input x. The key to ELM is to minimize the training error and output weight norm. i.e. minimize and ||β||.

The ELM algorithm is summarized as follows: given the training set {( _xi ,t _i ) _xi ∈R ⁿ ,t _i ∈R ^m ,i=1,2,...N}, the hidden layer node output function g(w, b, x) and the number of hidden layer nodes L.

(1) Randomly assign the parameters (w _i , _bi ) of the hidden layer nodes, i=1, 2...,L.

(2) Calculate the hidden layer output matrix H.

(3) Calculate the weight β between the hidden layer node and the output node. β=H ⁺ T.

H ⁺ is the Moore-Penrose generalized inverse matrix of the hidden layer output matrix H, which can be calculated using methods such as orthogonal projection, orthogonalization, and singular value decomposition.

Contents of the invention

The purpose of the invention is to provide a fast human facial expression recognition method based on the ELM self-encoding algorithm for the problems existing in the existing expression recognition algorithm. A faster and more efficient fast expression recognition method.

Technical scheme of the present invention mainly comprises the steps:

Step 1. Train the face region detection classifier

1-1. Given a series of training samples (x ₁ ,y ₁ ),(x ₂ ,y ₂ ),...,( _xi ,y _i ),(x _n ,y _n ), where x _i represents For the i-th sample, when y _i =0, it means it is a negative sample (non-face), when y _i =1, it means it is a positive sample (face), and n is the total number of training samples.

1-2. Initialize the weight and normalize the weight: D _t (i) is the error weight of the i-th sample in the t-th cycle, t=1...T.

1-3. For each feature f, train a weak classifier h(x, f, p, θ); calculate the weighted error rate of the weak classifier corresponding to all features:

1-4. Update the weights of all samples: Where β _t =ξ _t /1−ξ _t , e _i =0 indicates that _xi is correctly classified, and e _i =1 indicates that _xi is incorrectly classified.

1-5. The trained strong classifier can be used for face detection and recognition. If the picture contains a face, it will also locate the position of the face and determine the size. The strong classifier H(x):

where h _t is the weak classifier with the minimum error rate ξ _t during training.

Step 2. Face area preprocessing

2-1. Cut out the ROI of the detected face area, and then normalize the pixel size of the image: reduce/enlarge the image to an appropriate pixel size.

2-2. Perform histogram equalization processing on the normalized image to enhance image contrast. For discrete images, the equalization formula is: Pr(r _k )=r _k /N,0≤r _k <1; k=0,1,...,L-1, where N is the total number of pixels, k is the total number of gray levels, for an 8-bit gray level image k takes 2 ⁸ =256, r _k is the kth gray level value. The equalization transformation function is:

where n _j is the total number of pixels with gray value j.

Step 3, facial expression image feature extraction

3-1. Given training samples: X=[x ₁ , x ₂ , . . . , x _N ], that is, the input and output matrices of the ELM-AE.

3-2. Randomly generate hidden layer input weight matrix a=[a ₁ ,...,a _L ] and orthogonalization bias vector matrix b=[b ₁ ,...,b _L ], input data Mapped to the same or different data dimension space: h=g(ax+b)a ^T a=I,b ^T b=1 where: g() represents the activation function.

3-3. Solve the output weight matrix β of ELM-AE.

Suppose the number of neurons in the input and output layers is d, and the number of neurons in the hidden layer is L.

If d<L or d>L, that is, for sparse and compressed feature expressions,

If d=L, that is, for feature maps of equal dimensions, β=H ^-1 Xβ ^T β=I

Where: H=[h ₁ ,...,h _N ] represents the hidden layer output matrix of ELM-AE.

3-4. Input the preprocessed facial expression image to the trained ELM-AE system, and the obtained hidden layer output matrix vector H is the texture feature vector of the entire facial image.

Step 4. Build a facial expression classifier

4-1. Given training samples: {( _xi ,t _i )| _xi ∈R ⁿ ,t _i ∈R ^m ,i=1,2,...N}, the hidden layer output function g(w, b, x), the number of hidden layer nodes L and the test sample y.

4-2. Randomly generate hidden layer node parameters (w _i , bi ), _i =1, 2,...,L.

4-3. Calculate the hidden layer node output matrix H(w ₁ ,…w _L ,x ₁ ,…,x _N ,b ₁ ,…b _L ), and

Make sure that the H column is full rank, where w is the input weight connecting the hidden layer node and the input neuron, x is the training sample input, N is the number of training samples, b _i is the bias of the i-th hidden layer node, g() means activation function.

4-4. Calculate the optimal external weight β: β=H ⁺ T.

4-5. Calculate the output o=H(w ₁ ,...w _L ,x ₁ ,...,x _N ,b ₁ ,...b _L )β corresponding to the test sample y.

4-6. Carry out expression recognition and classification to the test samples, and the category corresponding to the maximum value in the ELM output vector o is the emotion of the face. which is

The beneficial effects of the present invention are as follows:

What the present invention adopts is deep extreme learning machine self-encoder (ELM-AE) algorithm to carry out facial expression feature extraction, and this algorithm is a kind of self-encoding algorithm more efficient than common AE self-encoding algorithm, and it can process high Dimensional input data, extract its main part information, and realize the feature expression of high dimension, equal dimension and low dimension of original data.

The invention has a faster recognition speed, and when using the ELM-AE algorithm to extract expression features, it can more quickly and efficiently extract main information and reduce dimensionality compared with a gradient descent algorithm with a slow learning rate. For facial expression recognition and classification, ELM only needs to adjust one parameter of the neuron, the recognition running time is short, and the accuracy is high. It is an efficient and fast learning algorithm.

The present invention can reduce the dimension of the data and represent the main components of the original information (i.e. the facial expression image). Compared with other feature extraction algorithms, it has the ability to quickly extract the basic components of the image, and can also process very high-dimensional input data . At the same time, the expression classification algorithm based on the extreme learning machine ELM has a faster learning speed and recognition speed. The combination of the two algorithms can greatly improve the speed and accuracy of facial expression recognition.

Description of drawings

Fig. 1 is a schematic flow chart of the present invention;

Fig. 2 is the Japanese JAFFE facial expression image database;

Fig. 3 is the facial expression image after preprocessing;

Figure 4 is a network structure diagram of ELM-AE;

Fig. 5 is a schematic diagram of a single hidden layer feed-forward neural network.

detailed description

As shown in Figure 1, first use the Adaboost algorithm to train the face region detection classifier, and combine several weak classifiers obtained from each training according to certain weights to obtain a strong classifier that can detect face regions. Then input the picture to be detected to the trained face detection classifier, and perform cropping, size pixel normalization and histogram equalization processing on the detected face area. Input the processed face expression picture into the trained ELM-AE feature extraction neural network, and the obtained hidden layer output matrix vector H is the texture feature vector of the whole face image. Finally, this feature vector is used as the input of the trained ELM expression recognition classifier, and the output of the corresponding expression category can be obtained.

The invention provides a fast human facial expression recognition method based on the ELM self-encoding algorithm. The ELM-AE algorithm is used to extract the facial expression features and use them as the input of the ELM expression classifier. The combination of the two not only improves the running speed, but also High accuracy.

The specific implementation method is as follows:

Step 1: Training face region detection classifier: For a set of training samples, by changing the distribution probability of each sample, different training sets Si are obtained, and each Si is trained to obtain a weak classifier Hi, and then A strong classifier is obtained by combining these weak classifiers according to different weights.

(1-1) As shown in Figure 2, use the Japanese JAFFE facial expression database as a training sample, give each sample initialization weight and perform weight normalization processing: D _t (i) is the error weight of the i-th sample in the t-th cycle, t=1...T.

(1-2) For each feature f, train a weak classifier h(x, f, p, θ); calculate the weighted error rate of the weak classifier corresponding to the + feature: And update the weights of all samples: Where β _t =ξ _t /1−ξ _t , e _i =0 indicates that _xi is correctly classified, and e _i =1 indicates that _xi is incorrectly classified.

(1-3) The strong classifier obtained after training can be used for face detection and recognition. If the picture contains a face, the center position and size of the face can be determined.

Strong classifier H(x):

where h _t is the weak classifier with the minimum error rate ξ _t during training.

Step 2: Face area preprocessing: As shown in Figure 3, the detected face area is cropped by the ROI of the region of interest, and then the pixel size is normalized to reduce/enlarge the image to a suitable pixel size , and perform histogram equalization.

(2-1) After ROI region segmentation is performed on the picture of the detected face area, the pixel size is normalized, and a fixed-size facial expression image is output.

(2-2) Perform histogram equalization processing on the processed image, and the equalization transformation function is:

where n _j is the total number of pixels with gray value j.

Step 3: Feature extraction of facial expression images: train the network structure of ELM-AE (as shown in Figure 4), and calculate different output weight matrix β according to the different feature expression dimensions. The trained ELM-AE network can be used for feature extraction of expression pictures.

(3-1) Given training samples: X=[x ₁ , x ₂ , . . . , x _N ], that is, the input and output matrices of the ELM-AE.

(3-2) Randomly generate the hidden layer input weight matrix a=[a ₁ ,...,a _L ] and the orthogonalized bias vector matrix b=[b ₁ ,...,b _L ].

(3-3) Map the input data to the same or different data dimension spaces:

h=g(ax+b) a ^T a=I,b ^T b=1 where: g() represents an activation function.

(3-4) Solve the output weight matrix β of ELM-AE.

Suppose the number of neurons in the input and output layers is d, and the number of neurons in the hidden layer is L.

If d<L or d>L, that is, for sparse and compressed feature expressions,

If d=L, that is, for feature maps of equal dimensions, β=H ^-1 Xβ ^T β=I

Where: H=[h ₁ ,...,h _N ] represents the hidden layer output matrix of ELM-AE.

(3-5) Input the preprocessed facial expression image to the trained ELM-AE system, and the obtained hidden layer output matrix vector H is the texture feature vector of the entire facial image.

Step 4: Build a facial expression classifier: as shown in Figure 5, build an expression classifier based on an extreme learning machine, randomly generate the parameters of hidden layer nodes, and optimize the only adjustment parameter β for training.

(4-1) Given a training sample: {( _xi , t _i )| _xi ∈ R ⁿ , t _i ∈ R ^m , i=1,2,...N}, the hidden layer output function g(w ,b,x), the number of hidden layer nodes L and the test sample y.

(4-2) Randomly generate hidden layer node parameters (w _i , bi ), _i =1, 2,...,L.

(4-3) Calculate the hidden layer node output matrix H(w ₁ ,…w _L ,x ₁ ,…,x _N ,b ₁ ,…b _L ), and

Make sure that the H column is full rank, where w is the input weight connecting the hidden layer node and the input neuron, x is the training sample input, N is the number of training samples, b _i is the bias of the i-th hidden layer node, g() means activation function.

(4-4) Calculate the optimal outer weight β: β=H ⁺ T.

H ⁺ is the Moore-Penrose generalized inverse matrix of the hidden layer output matrix H, which can be calculated using methods such as orthogonal projection, orthogonalization, and singular value decomposition.

(4-5) Calculate the output corresponding to the test sample y:

o=H(w ₁ ,...w _L ,x ₁ ,...,x _N ,b ₁ ,...b _L )β

The expression recognition classification is carried out to the test sample, and the maximum value in the ELM output vector o corresponds to

The category of is the emotion of the face, namely

Claims (1)

1. the fast face expression recognition method based on ELM own coding algorithms, it is characterised in that comprise the following steps：

Step 1, training face region detection grader

1-1. gives a series of training sample (x₁,y₁),(x₂,y₂),...,(x_i,y_i),(x_n,y_n), wherein x_iRepresent i-th of sample This, y_iIt is negative sample (non-face), y that it is represented when=0_iIt is positive sample (face) that it is represented when=1, and n is training sample altogether Quantity；

1-2. initializes weight and makees weight normalization：D_t(i) it is i-th of sample in the t times circulation This Error weight, t=1...T；

1-3. trains a Weak Classifier h (x, f, p, θ) to each feature f；Calculating corresponds to adding for the Weak Classifier of all features Weigh error rate:

1-4. is updated to the weight of all samples：Wherein β_t=ξ_t/1-ξ_t, e_i=0 represents x_iQuilt Correctly classify, e_i=1 represents x_iMistakenly classified；

Strong classifier after 1-5. training can be used for Face datection identification, if including face in picture, can also Navigate to the position of face and determine size, strong classifier H (x)：

<mfenced open = "" close = ""> <mtable> <mtr> <mtd> <mrow> <mi>H</mi> <mrow> <mo>(</mo> <mi>x</mi> <mo>)</mo> </mrow> <mo>=</mo> <mfenced open = "{" close = ""> <mtable> <mtr> <mtd> <mn>1</mn> </mtd> <mtd> <mrow> <munderover> <mo>&amp;Sigma;</mo> <mrow> <mi>t</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>T</mi> </munderover> <msub> <mi>&amp;alpha;</mi> <mi>t</mi> </msub> <msub> <mi>h</mi> <mi>t</mi> </msub> <mrow> <mo>(</mo> <mi>x</mi> <mo>)</mo> </mrow> <mo>&amp;GreaterEqual;</mo> <mn>1</mn> <mo>/</mo> <mn>2</mn> <munderover> <mo>&amp;Sigma;</mo> <mrow> <mi>t</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>T</mi> </munderover> <msub> <mi>&amp;alpha;</mi> <mi>t</mi> </msub> </mrow> </mtd> </mtr> <mtr> <mtd> <mn>0</mn> </mtd> <mtd> <mrow> <mi>o</mi> <mi>t</mi> <mi>h</mi> <mi>e</mi> <mi>r</mi> <mi>s</mi> </mrow> </mtd> </mtr> </mtable> </mfenced> </mrow> </mtd> <mtd> <mrow> <msub> <mi>&amp;alpha;</mi> <mi>t</mi> </msub> <mo>=</mo> <mi>l</mi> <mi>o</mi> <mi>g</mi> <mn>1</mn> <mo>/</mo> <msub> <mi>&amp;beta;</mi> <mi>t</mi> </msub> <mo>=</mo> <mi>l</mi> <mi>o</mi> <mi>g</mi> <mfrac> <mrow> <mn>1</mn> <mo>-</mo> <msub> <mi>&amp;xi;</mi> <mi>t</mi> </msub> </mrow> <msub> <mi>&amp;xi;</mi> <mi>t</mi> </msub> </mfrac> </mrow> </mtd> </mtr> </mtable> </mfenced>

Wherein h_tThere is minimal error rate ξ during for training_tWeak Classifier；

Step 2, human face region pretreatment；

The human face region detected is carried out region of interest ROI and cuts out by 2-1., then makees Pixel Dimensions normalizing to image Change is handled：Picture is reduced/enlarged into a certain suitable Pixel Dimensions；

2-2. strengthens picture contrast to the image after normalized as histogram equalization processing；For discrete picture, Equalizing formula is：Pr(r_k)=r_k/N,0≤r_k＜ 1；K=0, wherein 1 ..., L-1, N are the total numbers of pixel, and k is gray scale Level sum, 2 are taken for the gray level image k of 8⁸=256, r_kFor k-th of gray-scale value；Equalizing transforming function transformation function is：

Wherein n_jIt is the total pixel number that gray value is j；

Step 3, Facial Expression Image feature extraction；

3-1. gives training sample：X=[x₁,x₂,...,x_N], i.e. ELM-AE input and output matrix；

3-2. generates hidden layer input weight matrix a=[a at random₁,...,a_L] and orthogonalization bias vector matrix b=[b₁,..., b_L], input data is mapped to identical or different data dimension space：H=g (ax+b) a^TA=I, b^TB=1 is wherein：g () represents activation primitive；

3-3. solves ELM-AE output weight matrix β；

Assuming that input and output layer neuronal quantity is d, hidden layer neuron quantity is L；

If d ＜ L or d ＞ L, i.e., for sparse and compression feature representation,

If d=L, i.e., for etc. dimension Feature Mapping, β=H^-1Xβ^Tβ=I

Wherein：H=[h₁,...,h_N] represent ELM-AE hidden layer output matrix；

3-4. inputs pretreated Facial Expression Image to the ELM-AE systems that train, obtained hidden layer output matrix to Amount H is the texture feature vector of view picture facial image；

Step 4, structure facial expression classifier；

4-1. gives training sample：{(x_i,t_i)|x_i∈Rⁿ,t_i∈R^m, i=1,2 ... N }, hidden layer output function g (w, b, x), The number of hidden nodes L and test sample y；

4-2. generates hidden node parameter (w at random_i,b_i), i=1,2 ..., L；

4-3. calculates hidden node output matrix H (w₁,…w_L,x₁,…,x_N,b₁,…b_L), and

<mrow> <mi>H</mi> <mrow> <mo>(</mo> <msub> <mi>w</mi> <mn>1</mn> </msub> <mo>,</mo> <mo>...</mo> <msub> <mi>w</mi> <mi>L</mi> </msub> <mo>,</mo> <msub> <mi>x</mi> <mn>1</mn> </msub> <mo>,</mo> <mo>...</mo> <mo>,</mo> <msub> <mi>x</mi> <mi>N</mi> </msub> <mo>,</mo> <msub> <mi>b</mi> <mn>1</mn> </msub> <mo>,</mo> <mo>...</mo> <msub> <mi>b</mi> <mi>L</mi> </msub> <mo>)</mo> </mrow> <mo>=</mo> <mfenced open = "[" close = "]"> <mtable> <mtr> <mtd> <mrow> <mi>g</mi> <mrow> <mo>(</mo> <msub> <mi>w</mi> <mn>1</mn> </msub> <mo>&amp;CenterDot;</mo> <msub> <mi>x</mi> <mn>1</mn> </msub> <mo>+</mo> <msub> <mi>b</mi> <mn>1</mn> </msub> <mo>)</mo> </mrow> </mrow> </mtd> <mtd> <mn>...</mn> </mtd> <mtd> <mrow> <mi>g</mi> <mrow> <mo>(</mo> <msub> <mi>w</mi> <mi>L</mi> </msub> <mo>&amp;CenterDot;</mo> <msub> <mi>x</mi> <mn>1</mn> </msub> <mo>+</mo> <msub> <mi>b</mi> <mi>L</mi> </msub> <mo>)</mo> </mrow> </mrow> </mtd> </mtr> <mtr> <mtd> <mo>.</mo> </mtd> <mtd> <mrow></mrow> </mtd> <mtd> <mo>.</mo> </mtd> </mtr> <mtr> <mtd> <mo>.</mo> </mtd> <mtd> <mn>...</mn> </mtd> <mtd> <mo>.</mo> </mtd> </mtr> <mtr> <mtd> <mo>.</mo> </mtd> <mtd> <mrow></mrow> </mtd> <mtd> <mo>.</mo> </mtd> </mtr> <mtr> <mtd> <mrow> <mi>g</mi> <mrow> <mo>(</mo> <msub> <mi>w</mi> <mn>1</mn> </msub> <mo>&amp;CenterDot;</mo> <msub> <mi>x</mi> <mi>n</mi> </msub> <mo>+</mo> <msub> <mi>b</mi> <mn>1</mn> </msub> <mo>)</mo> </mrow> </mrow> </mtd> <mtd> <mn>...</mn> </mtd> <mtd> <mrow> <mi>g</mi> <mrow> <mo>(</mo> <msub> <mi>w</mi> <mi>L</mi> </msub> <mo>&amp;CenterDot;</mo> <msub> <mi>x</mi> <mi>N</mi> </msub> <mo>+</mo> <msub> <mi>b</mi> <mi>L</mi> </msub> <mo>)</mo> </mrow> </mrow> </mtd> </mtr> </mtable> </mfenced> </mrow>

Ensure H sequency spectrums, wherein w is the input weight for connecting hidden node and input neuron, and x is training sample input, and N is Training sample number, b_iIt is the deviation of i-th of hidden node, g () represents activation primitive；

4-4. calculates optimal outer power β：β=H⁺T；

4-5. calculates the corresponding output o=H (w of test sample y₁,…w_L,x₁,…,x_N,b₁,…b_L)β；

4-6. carries out Expression Recognition classification to test sample, is the face to the corresponding classification of maximum in ELM output vectors ο Mood；I.e.