A Study on the Innovation of Instructional Management System in the Perspective of Machine Learning Supported Student Engagement Enhancement

In the process of building the expression recognition model in this paper, the relationship between network depth and model accuracy is considered comprehensively, the number of network layers of the AlexNet-based improved convolutional neural network model in this summary is 8 layers in total, including 6 convolutional layers and 2 full-connectivity layers, and the input is a single-channel grayscale image of 48×48 pixels, which is highly compressed by using a small-scale convolution kernel, and the features of the image are highly compressed, so as to obtain a feature vector to retain the main features of the input samples. feature vector that retains the main features of the input samples, remove the redundant features of the samples through pooling operation, retain the sample features that are effective for expression classification, input the output weights of the upper layer into the fully connected layer for learning, and the dropout is set to 50%, i.e., randomly zero out 50% of the nodes, which is set up to make the expression recognition model less prone to overfitting, and to make the model more stable and robust.

Where the size of convolution kernel: F, the number of convolution kernels: K, the step size: S, the boundary padding: P, the size of input image: W₁×H₁×D₁, the size of output image: W₂×H₂×D₂, the computation process of W₂, H₂, D₂ during the convolution process and the computation process of W₂, H₂, D₂ during the pooling process are shown in Eqs. (7) ~ (12): (7) W2=W1−F+2PS+1 \[{{W}_{2}}=\frac{{{W}_{1}}-F+2P}{S}+1\] (8) H2=H1−F+2PS+1 \[{{H}_{2}}=\frac{{{H}_{1}}-F+2P}{S}+1\] (9) D2=K \[{{D}_{2}}=K\] (10) W2=W1−FS+1 \[{{W}_{2}}=\frac{{{W}_{1}}-F}{S}+1\] (11) H2=H1−FS+1 \[{{H}_{2}}=\frac{{{H}_{1}}-F}{S}+1\] (12) D2=D1 \[{{D}_{2}}={{D}_{1}}\]

In the training process of the convolutional neural network, the classroom expression dataset is firstly imported into the expression recognition model, and the expression recognition model is trained using the training set with a total of 10,000 iterations, and in each iteration, the data with the expression discrimination rate of more than 0.6 are retained, and the rest of the data are discarded, and the model is trained and tested using a randomized batch of data as the input data for each time, so that the time for each training and testing is reduced and the accuracy of the model is not affected. This reduces the training and testing time and does not affect the accuracy of expression recognition. The output of the model is the loss value and the accuracy of the model in recognizing expressions, and the accuracy of the expression recognition is constantly updated as the loss value decreases.

Student behavior recognition is trained based on the structure of ResNet50 pre-trained model [23]. This method is also known as transfer learning, which refers to the influence of one kind of learning on another kind of learning, and since learning activities are always based on existing knowledge and experience, this process of continuously acquiring new knowledge by utilizing one’s own knowledge and experience can be referred to as transfer learning in a broad sense [24]. The process of transfer learning based on the structure of ResNet pre-trained model is shown in Figure 1.

Figure 1.

Process of transfer learning algorithm based on ResNet50

The activation function selected in the deep learning based student behavior recognition algorithm is ReLU. The expression of ReLU function and the derivative expression are: (13) f(x)=max(0,x)f′(x)={ 1x>00x≤0 $\begin{align} & f\left( x \right)=\max \left( 0,x \right) \\ & {f}'\left( x \right)=\left\{ \begin{array}{*{35}{l}} 1 & x>0 \\ 0 & x\le 0 \\ \end{array} \right. \\ \end{align}$

Traditional machine learning consists of three elements: model, strategy and algorithm, when an algorithm is selected, we have to consider what kind of strategy or criterion is used to select the optimal model, the loss function is a function used to measure how good or bad the model’s prediction result is at one time in machine learning, and it is a non-negative real value function. Generally speaking, when the model performs a forward operation it will get a prediction value, however, the prediction value is different from the real value. Therefore, we need to define a way to measure the goodness of the model, which is the loss function, the loss function is usually denoted by L(Y,f(X)), in the deep learning-based student classroom behavior recognition, we use the squared loss function, the squared loss function is a common optimization objective function in the linear regression model, and also based on the ResNet50 pre-training model for training the loss function. It is denoted as equation (14): (14) L(Y, f(X))=(Y−f(X))2 \[L(Y,\ f(X))={{(Y-f(X))}^{2}}\]

In the previous subsection, we introduced the loss function calculation method is the difference between the predicted value of the student category and the true value of the student category, of course, this difference should be minimized as much as possible, then it means that the model has the best learning effect. In mathematics, in order to calculate the minimum value of the function, that is, you need to find the rate of change of the function image, generally in the function of the largest rate of change of the function will appear in the function of the maximum value, and in the multivariate function, the gradient is a vector, the vector has a direction, the direction of the gradient points to the function at the given point in the direction of the fastest rise, then the gradient of the direction of the opposite direction of the function at the given point in the direction of the fastest decline, so in order to find the minimum of the loss function, we need to find the direction of the gradient of the loss function, and the direction opposite to its gradient is also the direction of the minimum of the loss function. The loss function we used in the previous subsection is the squared loss function, as shown in Equation (15): (15) J(θ)=minθ12∑i=1m(hθ(x(i)−y(i)))2 \[J\left( \theta \right)=\underset{\theta }{\mathop{\min }}\,\frac{1}{2}\sum\limits_{i=1}^{m}{{{\left( {{h}_{\theta }}\left( {{x}^{\left( i \right)}}-{{y}^{\left( i \right)}} \right) \right)}^{2}}}\]

where J(θ) denotes the function that minimizes the variance with respect to θ, which is actually the value of the loss function, representing the fit between the model and the training data, and then look at an equation: (16) θi=θi−α∂∂θiJ(θ) \[{{\theta }_{i}}={{\theta }_{i}}-\alpha \frac{\partial }{\partial {{\theta }_{i}}}J\left( \theta \right)\]

Where θ_i represents the value of the loss function, α represents the value of the learning rate, the right side of the learning rate indicates the direction of descent, so it can be seen that the learning rate determines the rate of gradient descent, if the learning rate is set too large, it may skim over the lowest point of the function, resulting in the model can not be converged, if the learning rate is set too small, resulting in the gradient descent is too slow. It can be seen that the appropriate learning rate is important for the training process, and the learning rate is set to 0.001 in this research program.

The main function of the face-to-face classroom management system is to provide automated and intelligent management support functions for classroom teaching, and to improve the efficiency and quantitative evaluation of classroom teaching. It mainly provides functions including student face check-in and attendance, student abnormal behavior capture and prompt, teacher face attendance, teacher classroom management tools, comprehensive classroom evaluation and other aspects. The participants of traditional face-to-face classroom mainly consist of students, teachers and teaching management personnel. By analyzing the usage scenarios, the system will mainly include the following user roles.

Starting from the use of user roles, this paper designs the system to have three application scenarios: learning scenario, teaching scenario and teaching management scenario, corresponding to the three user roles of students, teachers and teaching management personnel, relying on the system to carry out face-to-face classroom teaching management.

According to the analysis of user scenarios and according to the user operation view, the whole system framework can be composed of three functional components, mainly including the application operation component, face recognition component, and scoring component. Among them, the application operation component is responsible for the actual application and operation part of the system, facing the actual users and providing interaction and operation interface. The face recognition component is responsible for the core functions such as face recognition and facial expression comparison. The scoring component is responsible for establishing a set of student behavior and classroom efficiency scoring models for quantitative classroom assessment. The three components are on the data flow of the core face recognition function. The application operation component is responsible for providing the user with the use interaction and obtaining the operation instructions, inputting the face data into the face recognition component, and after the face recognition component processes the results, entering the results into the scoring component, obtaining the quantitative results, and feeding back to the application operation component. The system framework is shown in Figure 2.

Figure 2.

System framework

Each component is positioned according to different functions and provides different functional services according to the link in the operation process.

Figure 3.

Face recognition component function structure