Image Segmentation and Target Extraction of Preschool Educational Activity Space for Improving Children's Concentration

Children's pointing and concentrate on specific things are examples of children's psychological activity, and these actions are referred to as children's concentration [1, 2]. Children's mental activity are said to be direct when they point away from one thing at a specific point and toward other objects, a sign of their selectivity [3-6]. Although it cannot be compared to a child's IQ, a child's intelligence will be somewhat impacted by their level of concentration. It might be stated that children's character development and excellent character formation depend on their ability to concentrate [7-9].

The majority of the time, teachers will carry out preschool education activities based on the user's needs for concentration training, such as clapping games, operational activities, mirror games, and other enjoyable activities, or group dances, aerobics, cheerleading, stacking arhats, rubber band jumping, archery, and other sports activities, or develop independent activity plans in accordance with student individual differences [10-18]. It is necessary to analyze relevant activity video images to support the follow-up teachers in sorting out the principles and strategies for observing and analyzing children's behaviors in activities in order to achieve effective observation of teachers and children in preschool education activities intended to cultivate children's concentration.

Children find it challenging to focus while learning online. Facial expressions are crucial for grabbing pupils' attention in the present, according to Yu and Luo [19]. The researchers build a student attention prediction model using the self-healing network (SCN) method and the random forest method based on C4.5 based on the facial expression data of students and the attention scores provided by teachers in online teaching activities in an elementary school. This model is then used in actual online teaching. Kokkalia et al. [20] investigated play that supported toddlers with cognitive, intellectual, physical, and literacy challenges. It also investigated how instructional video games can benefit brilliant youngsters and children with developmental disabilities like autism. Through the use of a smart watch device, Hosseini et al. [21] investigated the issue of recognizing children's activity. Bidirectional LSTM model and fully connected deep network are the two deep neural network models that are used, and the outcomes are contrasted with other models that are frequently employed in the field. In comparison to the baseline model, experimental results show that the proposed deep model can greatly improve the results. It also highlights the benefit of activity intensity level identification in health monitoring and confirms the suggested model's excellent performance in this task. Particularly for toddlers and young children, the current system does not account for all age ranges. To obtain benchmark results for various unimodal and multimodal settings, Basak et al. [22] compares the performance of currently available tools and algorithms. Children's faces are the simplest modality to record, according to the data, which also demonstrate great verification performance.

Video surveillance, human-computer interaction, television and movies, the military, and sports are just a few of the many industries where segmentation and extraction of moving objects for dynamic scenarios are used. According to a survey of the literature, the majority of studies in the aforementioned applied research topics focus on object tracking, event monitoring, video content retrieval, and object categorization. Comparatively few studies have focused specifically on the examination of photographs of children's activity spaces. For the purpose of cultivating children's concentration, this study performs research on the image segmentation and target extraction of preschool education activity space. The relationship between preschool educational activities and children's concentration is discussed in Section 2, and the frequency-tuned salience algorithm is introduced into the Gaussian mixture model created to realize the spatial images of preschool educational activities for the development of children's concentration, creation of spatial information and active segmentation of spatial images. In Section 3, the ViBe algorithm ViBe+ of complex scenes is chosen to realize the target recognition of multiple children in a complex preschool educational environment. This algorithm is targeted at issues like small target children, large numbers of target children, overlap of target children and background colors, or strong light interference, etc. The efficiency of the algorithm is confirmed by experimental data.

The setting for early childhood education activities targeted at developing children's attentiveness is fairly complicated, and the scene of the activity area image changes quickly. The background changes in the activity space image between the present active video frame and the preceding active video frame can be more accurately detected using the Gaussian mixture model. However, in the collected preschool education activity space image sequences for the development of children's concentration, Gaussian mixture is only appropriate for simple dynamic scenes with few target children, many target children, overlapping colors between the target children and background, or strong light interference. The results of the model's detection will be far less accurate. The speedy detection of many children's targets in the complicated preschool educational environment is a challenge that this research chooses to answer using the ViBe algorithm ViBe+ with complex sceneries (Figure 3).

3.png

Figure 3. Flow of multitarget detection

The proposed method is divided into three steps: foreground segmentation, target detection post-processing and eliminating false alarm targets. Foreground segmentation aims to characterize the dynamic changes of the current background. Thus, the background complexity BF(a) corresponding to the target child pixel a to be detected is calculated as follows. Assuming that the position of pixel a in the neighborhood is represented by m, the pixel value of a in the neighborhood is represented by u(ai), and the mathematical expectation in the neighborhood of a is represented by U^*. Then, we have

$B F(a)=\sqrt{\frac{\sum_{i=1}^m\left(u\left(a_i\right)-U^*\right)^2}{9}}$             (11)

Then, according to the complexity of the background change of the activity space image, the threshold T used to segment the foreground object and the background of the image is adaptively adjusted based on the following formula, and the initial detection result map of the target child is further obtained. Assume that the global segmentation threshold is denoted by T₀, the global threshold to measure background complexity is denoted by o, and the threshold adjustment parameter is denoted by TZ. Then, threshold T can be configured by:

$T=\left\{\begin{array}{l}T_0+T Z *(B F(a)-o), B F(a) \geq o \\ T_0, \text { other }\end{array}\right.$               (12)

The post-processing step of target detection intends to solve the problems of internal holes, noise and edge breaks in the target area. This paper firstly processes the target child's initial detection result map obtained in the foreground segmentation step with mathematical morphology. Let the input preliminary inspection result map be represented by X, and the structural elements be represented by Y. Then, the discrete noise points are eliminated, and the closed operation for filling the internal cavity of the target area can be expressed as:

$X \cdot Y=(X \oplus Y) \Theta Y$               (13)

To achieve better smooth suppression of edge noise and highlight the information features of key areas such as contours and edges in the activity space image, this paper chooses the discrete differential Sobel operator combined with Gaussian smoothing and differential derivation, and completes the activity space image edge detection and edge enhancement. Assuming that the convolution template of the Sobel operator in the horizontal direction is represented by O_a, and the convolution template in the vertical direction is represented by O_b. Then, the convolution algorithm implemented by the Sobel operator is as follows:

$O_b=\left[\begin{array}{ccc}-1 & -2 & -1 \\ 0 & 0 & 0 \\ 1 & 2 & 1\end{array}\right]$             (14)

$O_a=\left[\begin{array}{rrr}-1 & 0 & 1 \\ -2 & 0 & 2 \\ -1 & 0 & 1\end{array}\right]$             (15)

Assume that the horizontal gradient and vertical gradient of the activity space image are represented by h_a(a,b) and h_b(a,b), respectively, the gradient direction is represented by ω(a,b), and the neighborhood activity space image at the pixel point (a,b) is represented by g(a+θ,b+u). The related formula can be expressed as:

$h_a(a, b)=\sum_{q=-1}^{q=1} \sum_{u=-1}^{u=1} O_a(q, u) g(a+q, b+u)$               (16)

$h_b(a, b)=\sum_{q=-1}^{q=1} \sum_{u=-1}^{u=1} O_b(q, u) g(a+q, b+u)$          (17)

$\omega(a, b)=\arctan \left(\frac{h_b(a, b)}{h_a(a, b)}\right)$          (18)

To obtain the shape edge features of the target child in the activity space image, this paper constructs an edge feature model corresponding to the foreground area in the original activity space image in the step of eliminating the false alarm target, so as to detect and eliminate the false alarm area of the image. To ensure that the identified target children's shape edges overlap with the actual image shape edges as much as possible, the histogram statistics and normalization processing are performed on h_a(a,b) and h_b(a,b), using the Canny operator. Assume that the number of foreground regions in the activity space image is represented by m, the feature model of the foreground region is represented by E, the vector storing the normalization parameter a_i^* is represented by A, the feature model of the original motion region is represented by W, and the vector storing the normalization parameter is represented by B. Further, the E and W of the feature model can be constructed by:

$\left\{\begin{array}{l}E_{i(i=1,2, \ldots, m)}=A_i\left(a_{i 1}, a_{i 2}, \ldots, a_{i 36}\right) \\ W_{i(i=1,2, \ldots, m)}=B_i\left(a_{i 1}, a_{i 2}, \ldots, a_{i 36}\right)\end{array}\right.$              (19)

Assuming that the two M-dimensional vectors of the marked edge and the actual edge are represented by A and B respectively, and the value range of a_ij and b_ij is [0,1]. Then, the feature model similarity R_i(A_i,B_i) can be calculated by:

$R_i\left(A_i, B_i\right)=\frac{A \mathrm{I} B}{A \mathrm{U} B}=\frac{A_i \bullet B_i}{\left\|A_i\right\|^2+\left\|B_i\right\|^2-A_i \bullet B_i}=\frac{\sum_{j=1}^M a_{i j} b_{i j}}{\sqrt{\sum_{j=1}^M\left(a_{i j}\right)^2+\sum_{j=1}^M\left(b_{i j}\right)^2 \sum_{j=1}^M a_{i j} b_{i j}}}$              (20)

It can be seen from the above formula that when A and B are non-zero vectors, and A_i=B_i, R_i(A_i,B_i)=1, according to the following formula, R_i(A_i,B_i)<1 when $A_i \neq B_i$. If one of Ai and Bi is a zero vector, then R_i(A_i,B_i)=0.

$R_i\left(A_i, B_i\right) \leq \frac{\sum_{j=1}^M a_{i j} \cdot b_{i j}}{2 \sum_{j=1}^M a_{i j} b_{i j}-\sum_{j=1}^M a_{i j} b_{i j}}=1$             (21)

False alarm target elimination: The degree of similarity between the foreground area characteristics of the extracted activity space image and the equivalent activity space area features of the original image can be determined when the similarity judgment of vectors A and B is complete. The similarity between determinable quantity features increases as R_i(A_i,B_i) approaches 1. It is determined that the foreground area is not a false alarm area when R_i(A_i,B_i) is less than the similarity threshold T₀, and all of the pixels' values in the foreground area are left alone. Otherwise, set all of the area's pixel values to zero. The technique for removing false alarm targets is depicted in Figure 4.

4.png

Figure 4. Flow of false alarm target elimination

In order to cultivate children's attentiveness, this study performs research on image segmentation and target extraction of activity space in preschool education. Preschool educational activity intervention and children's concentration are related, and the frequency-tuned saliency algorithm is added to the created Gaussian mixture model to realize the construction of spatial information of preschool educational activity space images for the development of children's concentration and Activity space image segmentation. The ViBe algorithm with complex scenes ViBe+ is chosen to realize the rapid detection of multiple child targets in complex preschool educational activity environments. This algorithm targets issues like small target children, large numbers of target children, color overlap between the target children and the background, or strong light interference.

The intervention plans for different preschool education activities were formulated through experiments. The image processing accuracies on different children targets were compared. It was found that our algorithm achieved more accurate and robust effect in image segmentation and target detection than frame difference method and background difference method. In addition, this paper performs frequency counting and significance testing of the concentration behaviors of bottom, middle, and top class children in the preschool education activity scenarios, and further discloses the development features of children of different ages in preschool education activities.