Human-computer system design of entrepreneurship education based on artificial intelligence and image feature retrieval

Abstract

The key of deep learning is how to extract abstract, deep and nonlinear target features, in which algorithm plays a crucial role. In this paper, the authors analyze the intelligent system design of entrepreneurship education classroom based on artificial intelligence and image feature retrieval. Pyramid pooling is used to transform any size feature map into fixed size feature vector, which is finally sent to the full connection layer for classification and regression. Experimental results show that the algorithm accelerates the convergence of the whole network and improves the detection speed. The education taught by entrepreneurial class is not only to help college students to seek a stable career, but also to help college students develop their own potential, cultivate entrepreneurial awareness, improve entrepreneurial quality and ability. Entrepreneurship education should not only stay in the design of subject courses, but should integrate entrepreneurship education with internet entrepreneurship practice. On this basis, we provide new countermeasures and suggestions for improving the quality and ability of college students in the process of entrepreneurial activities.

Keywords

Human computer interaction image feature retrieval entrepreneurship education artificial intelligence class

1 Introduction

With the development of entrepreneurship education in Colleges and universities in China, the quality and ability of entrepreneurship of college students are also valued by colleges and universities [1]. From the perspective of entrepreneurs themselves, entrepreneurial behavior is not only an independent employment model, but also an important means to show personal potential and life value [2, 3]. Cultivating the quality and ability of entrepreneurship is the premise for entrepreneurs to optimize their activities and behaviors [4]. It can not only affect the behavior of entrepreneurs from the inside, but also work together with other external factors and run through the whole process of entrepreneurs’ behaviors. Intelligent education is an educational transformation measure under the background of artificial intelligence era [5, 6]. It is not only the upgrading and change of traditional education forms, but also the transformation and improvement of the current education and teaching mechanism. From the traditional teaching emphasis on knowledge transfer and integration to the comprehensive and systematic training of students’ own abilities. In the process of the development of intelligent education, it can show the guiding role of education and teaching in the overall development of students, as well as the full shaping of students’ own wisdom and morality, and lay a good track for students’ growth and life [7]. And in the intelligent society, it can also give education and teaching a new development space, build a more perfect education and teaching environment, update the corresponding infrastructure, design a more flexible and diverse education mechanism, and meet the needs of students’ growth and development to the maximum extent.

One and the only one is that the government has provided many effective policies and venture capital support to support the entrepreneurship of university students. The “Internet plus” entrepreneurship platform has provided a powerful weapon for university students to start their own businesses. It provides a unique opportunity for university students to start their own businesses [8, 9]. Therefore, it is the inevitable result of the development of social economy and technology culture for university students to regard entrepreneurship as the mode of employment and the path of employment. Therefore, by changing the traditional view of employment, establishing a correct view of employment, and shifting the field of employment and vision to the field of entrepreneurship, college students keep up with the trend of the times, and cultivate the awareness of independent entrepreneurship. The necessary abilities for entrepreneurship do not exist originally [10]. The cultivation of these abilities can only be formed by participating in the entrepreneurship education and training carried out by colleges and universities, as well as the accumulation of entrepreneurial activities. Therefore, the best way for college students to cultivate entrepreneurial ability is to participate in entrepreneurial training. During the period of school, college students can participate in entrepreneurial knowledge competition, entrepreneurial lectures, etc. to understand the basic knowledge needed for entrepreneurship and the current economic situation of the market; they can also participate in entrepreneurial skills training activities or entrepreneurial competitions jointly organized by school and enterprise, etc. to accumulate entrepreneurial experience through these practices Improve your ability [11, 12].

No matter in the field of artificial intelligence or in the field of education, in-depth learning pays attention to the extraction of information and knowledge. In the field of artificial intelligence, in-depth learning pays more attention to the application of educational information and better assisted teaching driven by technology and data [13]. In the field of education, in-depth learning pays more attention to students’ deep understanding, internalization and application of knowledge, emphasizes an active inquiry learning method, and cultivates students’ ability of information processing, knowledge transfer and practical application [14, 15]. This research mainly analyzes the classroom entrepreneurship teaching behavior with the help of machine learning and deep learning algorithm in the field of artificial intelligence.

2 Related work

Due to the special technical characteristics of the Internet and the special business management mode of Internet enterprises, Internet entrepreneurship has some differences from traditional industries, and the new wave of entrepreneurship led by the new wave of Internet also has its new characteristics [16, 17]. College students entrepreneurs belong to an independent entrepreneurial group among many entrepreneurs, which not only has the necessary quality of other types of entrepreneurs in society, but also has the special nature that other entrepreneurs do not have, with special and typical characteristics. Students are the main body of entrepreneurial behavior, and also the main object of entrepreneurial education [18]. Consciousness plays a guiding role in behavior. Therefore, college students should start from themselves and have a comprehensive and objective understanding and evaluation of themselves [19]. At the same time, they should also closely follow the social background of Internet, recognize the current employment situation, do a good job in career planning, make clear their own personality characteristics and preferences, not good at high expectations, constantly adjust themselves in the process of employment, and establish a correct concept of employment.

With the arrival of the era of artificial intelligence, the combination of artificial intelligence and learning science has formed a new field of educational artificial intelligence [20]. A large number of educational artificial intelligence systems integrating artificial intelligence technology (such as data mining, etc.) have been applied in schools, so as to make better use of the big data generated in the campus, analyze and predict the students’ learning performance in school, and provide students with more Personalized and diverse learning styles [21, 22]. At present, the AI technology widely used in the field of education includes image recognition, speech recognition, semantic recognition, machine translation and so on.

Behind the human learning behavior lies the complex processing mechanism of brain science, which is machine learning theory. It mainly uses the computer to simulate and realize the process of human acquiring knowledge. It uses a large number of existing data and various algorithms to train a model that can solve specific problems. This model has a certain generalization ability [23]. When inputting new data, it can get effective results. The ultimate purpose is to acquire knowledge from the data. Machine learning can be divided into two stages: shallow learning and deep learning. Shallow learning mainly relies on artificial experience. Through BP algorithm, sample features are extracted from a large number of samples and predicted. In contrast to shallow learning, deep learning mainly relies on the basic logic framework of machine learning and massive data, and improves the accuracy of classification and prediction through continuous training and learning. Chinese style.

The core goal of speech recognition technology is to convert human speech into words, slice sound waves of speech into frames, combine frames into phonemes, synthesize factors into words, and then speech becomes readable words. Image recognition technology is based on image processing algorithm to analyze and process the image to achieve the purpose of recognizing and understanding the image. The specific application includes recognizing the specific target in the image, recognizing the image scene and recognizing the action of the image subject. Typical applications include face recognition, text recognition, etc.

Under the impact of the new generation of information technology development tide represented by artificial intelligence, the research enthusiasm of technology participation in classroom teaching will continue to increase. However, at present, the practical research on artificial intelligence in classroom teaching field mainly focuses on the development of teaching aids and the analysis of speech behavior, and the theoretical research focuses on the exploration of teaching mode and auxiliary teaching evaluation. In the aspect of classroom teaching behavior analysis, there is a lack of practical research [24]. At present, the analysis of classroom teaching behavior is mainly based on manual annotation and record analysis. If it is a large-scale analysis of classroom teaching behavior, the workload of manual analysis is large, which is easy to cause one-sided and lag of information analysis. Moreover, the analysis of classroom teaching behavior is mostly based on observation, and the timeliness and reliability of the analysis are low.

With the penetration of new technology in the field of education, data-driven intelligent analysis and research has gradually formed. The quality of classroom teaching has a direct impact on talent training. It is of great significance to explore the most direct level of AI in education, classroom teaching, with the aid of AI technology to drive the optimization of teaching [25]. With the help of technology, it is the strategic need of the national medium and long-term education reform and development and the development of education informatization to see through the occurrence and development process of teaching behavior in the classroom and pay attention to the influencing factors of classroom teaching quality. Promote the research of artificial intelligence in classroom teaching behavior, promote the deep integration of information technology and teaching, improve the theoretical system of classroom teaching behavior analysis, increase the scientificity and operability of classroom teaching observation and analysis, at the same time, guide teaching more accurately, and provide new ideas and ideas for the improvement of classroom teaching quality.

3 Theoretical analysis

3.1 Approximate calculation of multi-scale features

The key of deep learning is how to extract abstract, deep and nonlinear target features, in which algorithm plays a crucial role. The excellent traditional target detection algorithm consists of two steps: extracting image features by feature description operator. In the next few years, the open data set including PASCAL VOC has become more and more perfect, the performance of GPU and other computing hardware has been improved, and various computer vision algorithms based on deep learning have been put forward. At the same time, the performance has been rising in various image competitions. Thus, the golden age of deep learning officially comes. Figure feature pyramid as show in Fig. 1.

Fig.1

Figure feature pyramid.

The richer the feature extracted by the algorithm, the more abstract the image can be represented. Therefore, the most important step in target detection is feature extraction. The scanning strategy based on sliding window is the main method of traditional target detection, which consists of three parts: selecting candidate region, extracting target features and designing a reasonable classifier. The most important task of traditional target detection is how to extract efficient target features. The feature extraction stage has experienced a transformation process from global image pixels to local feature descriptors.

$\frac{\partial I_{k}}{\partial x} (i, j) = \frac{1}{k} \frac{\partial I}{\partial x} (i / - k, j / - k), \frac{\partial I_{k}}{\partial x}$ (1)

Firstly, we use Gauss differential function to detect the extremum in each scale space, and then locate, assign and describe the extremum. Sift is invariant to rotation, scale scaling, brightness change, etc. at the same time, it has the advantages of good uniqueness, large quantity, real-time and so on. $M_{K} (i, j) \approx \frac{1}{k} M [⌈ i / - k ⌉, ⌈ j / - k ⌉]$ (2)

so $\begin{matrix} \sum_{i = 1}^{kn} \sum_{j = 1}^{kn} M_{k} (i, j) \approx \sum_{i = 1}^{kn} \sum_{j = 1}^{kn} \frac{1}{k} M [⌈ i / - k ⌉, ⌈ j / - k ⌉] \\ = k^{2} \sum_{i = 1}^{n} \sum_{j = 1}^{m} \frac{1}{k} (i, j) = k \sum_{i = 1}^{n} \sum_{j = 1}^{m} (i, j) \end{matrix}$ (3)

Using a star architecture, including root and component filters and related deformation models for target detection, DPM divides the images by moving windows, and extracts the hog features and classifies SVM for each area after segmentation, even if the target is partially occluded, DPM It can also be detected successfully, but the detection time is too long, which seriously affects its wide application. $\begin{matrix} \frac{\partial I_{k}}{\partial x} (i, j) / - \frac{\partial I_{k}}{\partial y} (i, j) \approx \\ \frac{\partial I}{\partial x} (i / - k, j / - k) / - \frac{\partial I}{\partial y} (i / - k, j / - k) \\ h_{q} \approx {kh}_{q} \end{matrix}$ (4)

Fast r-cnn adopts the design idea of pyramid pooling. The algorithm first normalizes the image and then transmits it to convolutional neural network, but unlike sppnet, fast r-cnn The region of interest is added before the full connection layer. By pooling the region, the feature map of fixed size can be obtained. At the same time, the coordinate information of the candidate frame is considered. Finally, the multi task loss function is introduced to enable the detection model to return to the coordinate position of the target in the full connection layer.

3.2 Approximate estimation of features at multiple scales

In order to increase the speed of real-time target detection, fast r-cnn encapsulates the target detection in a unified end-to-end framework. By introducing the region candidate network, the region most likely to contain the target in the original image is mapped on the feature map, and then further pooled and transmitted to the subsequent fast r-cnn detection network. $f (I, s_{1}) / - f (I, s_{2}) = r (s_{1} - s_{2})$ (5)

Context information refers to the local or global information around the region of interest in a picture. In addition, the author introduces the structure of hop layer connection. By fully extracting and fusing the multi-layer region of interest features of the deep convolution network, the author classifies and regresses each position using the features. In this algorithm, a new idea is put forward: the convolution of the region of interest in Faser r-cnn is moved to the convolution layer of the region of interest, and then r-fcn designs the position sensitive score graph, which aims to deal with the translation invariance and shift transformation of the detected image target. Multi-scale feature extraction as show in Fig. 2.

Fig.2

Multi-scale feature extraction.

3.3 Estimation of parameter λ

The multi-scale space of the feature pyramid is adopted by the algorithm framework. At the same time, the semantic features of other different convolution feature graphs are integrated into the feature graphs in the pyramid space to stack the pixels, and then predict. Finally, the method achieves high detection accuracy, we first calculate: $μ_{s} = [r (I, s)] (s = \frac{1}{8}, \frac{2}{8}, \dots, \frac{24}{8})$

The algorithm is based on fast r-cnn, and uses the feature extraction network of FPN, and then adds a sub part of region of interest. This sub part will search the possible candidate regions in the original image, and then use a complete convolution neural network to predict each region of interest, so as to correct these regions. $λ_{s} = In a - In μ_{s}$ (6)

In order to prevent a large number of simple negative samples from affecting the target detector in the process of training, the training process is mainly focused on a sparse difficult sample set to solve the problem of category imbalance at one time. Experimental results show that, especially for the end-to-end target detection algorithm, using focal loss training method can make the detection algorithm get a huge improvement in accuracy and speed. $C_{S} = Ω (R (I, s)), s \in {1, \frac{1}{2}, \frac{1}{4}, \dots}$ (7)

At the same time, a learning rate selection strategy and BN cross GPU method are proposed, which can greatly reduce the training time of the object detector and achieve higher accuracy. $\begin{matrix} C_{S} = R (C_{S}, s / - s^{'}) {(s / - s^{'})}^{- λ_{Ω}} \\ s^{'} \in {1, \frac{1}{2}, \frac{1}{4}, \dots} \end{matrix}$ (8)

It uses the sliding window to generate the target candidate frame, and then uses the multi-scale sliding window to increase the detection results. Its purpose is to solve the problems such as the change of target shape and size in the image. Finally, it uses the regression model and the depth convolution neural network to classify and locate the target. The algorithm creatively combines classification, location and detection together to solve the problem of target detection as a whole. Fast feature pyramid as show in Fig. 3.

Fig.3

Fast feature pyramid process.

The original intention of the algorithm is to solve the problem of low detection efficiency caused by the large number of candidate regions. Previous target detection methods based on candidate region extraction search targets through a large number of candidate boxes. This method first initializes the candidate boxes after image regression, and then uses the regression box as the original window to continue regression adjustment, so as to essentially suppress negative samples.

Dssd improves the performance of small target detection based on SSD, and integrates the rich semantic information of high convolution layer and the high resolution of low convolution layer. The experimental results show that when the target scale is small, the target distance is close and the background before the target is easy to distinguish, the detection accuracy is improved under the condition that the detection speed does not drop too much. $C_{S} \approx R (C, s) \cdot s^{- λ_{Ω}}$ (9)

Due to the defects of SSD network, the detection effect of small target is not good. Des improves the semantic information of low-level feature map and further improves the accuracy of target detection by introducing segmentation module and global activation module. Ron algorithm, by combining the advantages of two kinds of target detection based on deep learning, uses multi-scale representation to significantly improve the detection of various multi-scale objects, and designs the direction connection structure. At the same time, in order to reduce the object search space, the target priority is created on the convolution feature map to guide the object search, and the target detector is optimized in the training process.

$f_{Ω} (I) = \sum_{i, j, k} w_{i, j, k} C (i, j, k)$ (10)

In the formula, the feature f_Ω represents the weighted sum of the channel features C. $\frac{f_{Ω} (I_{s_{1}})}{f_{Ω} (I_{s_{2}})} = {(s_{1} / - s_{2})}^{- λ_{Ω}} + ɛ$

Location refinement module and object detection module. On the one hand, the relatively rough border information is obtained through the region candidate network, and then further regression is obtained through the normal regression branch, so as to obtain more accurate border information; on the other hand, by introducing the feature fusion operation, while ensuring the accuracy, the speed is also improved. ${\begin{matrix} μ_{s} = \frac{1}{N} \sum_{i = 1}^{N} f_{Ω} (I_{s}^{i}) / - f_{Ω} (I_{1}^{i}) \\ μ_{s} = α_{Ω} s^{- λ_{Ω}} \end{matrix}$ (11)

3.4 Time complexity analysis

By mining the structural association between objects of different scales and introducing low-level fine-grained features, the perceptual generative adversary network enhances the feature representation of small objects and improves the detection rate of small objects. It consists of two subnetworks, i.e. generating network and resolving network. As a depth residual feature generation model, the generation network transforms the original low-level features into high-level features by introducing low-level fine-grained features; the resolution network not only improves the detection rate through perceptual loss, but also distinguishes the high-level features generated by small objects from the real object features, as shown in Equation (12). $\begin{matrix} \sum_{k = 0}^{\infty} n^{2} 2^{- 2 k / - m} = n^{2} \sum_{k = 0}^{\infty} {(4^{- 1 / - m})}^{k} \\ = \frac{n^{2}}{1 - 4^{- 1 / - m}} \approx \frac{{mn}^{2}}{In 4} \end{matrix}$ (12)

In fast RCNN and region candidate network, the calculation amount is shared from the input image to the end of the last convolution layer; the difference is that fast RCNN and region candidate network are two independent steps from the last convolution layer. $ɛ_{t} = \sum_{n = 1}^{N} d_{n}^{(t)} I [y_{n} \neq h_{t} (x_{n})]$ (13) $w_{t} = \sum_{n = 1}^{N} d_{n}^{(t)} I [h_{t} (x_{n}) = 1, y_{n} = 1]$ (14)

Regional candidate network is the biggest highlight of faster r-cnn network. Using any size image as input, the region candidate network outputs a set of target candidate boxes of different sizes and scales, each of which is accompanied by a detection score. $α_{t} = \frac{1}{2} In \frac{1 - ɛ_{t}}{ɛ_{t}} + p \cdot e^{w_{t}}$ (15)

Translation invariance in target detection is very important to the detection task, because it can reduce the size of the model. It means that when a target image moves, the anchor will move with it, and the RPN can also recommend the candidate area. $d_{n}^{(t + 1)} = d_{n}^{(t)} exp (- α_{t} y_{n} h_{t} (x_{n})) / - Z_{t}$ (16)

Faster r-cnn’s anchor based method is based on the anchor pyramid, using anchors of different scales and ratios of width to height as reference classification and regression constraint boxes. The design of multi-scale anchor points can share features without additional calculation overhead, which is the key to solve the multi-scale problem. $Z_{t} = \sum_{n - 1}^{N} d_{n}^{(t)} exp (- α_{t} y_{n} h_{t} (x_{n}))$ (17)

SSD generates a series of fixed number of default frames on the specific characteristic graphs of different sizes by convolution kernels of size 3 * 3, then uses these default frames to predict the category and location, and finally filters the predicted frames according to the detection score by non maximum suppression algorithm to get the target detection results. This method completes classification and regression in the framework of a convolutional neural network, so its detection speed is greatly improved, and the real-time effect is achieved. $H_{t} (x) = sign [\sum_{t = 1}^{T} α_{t} h_{t} (x) - ET]$ (18)

4 Empirical analysis

In order to verify the validity of the data mining model in the competency analysis of the innovative entrepreneurial team based on multi-angle mining, the relevant experimental analysis is needed. The data mining model based on the stepwise linear regression method is used to analyze the data mining model of the innovation entrepreneurial team competency analysis.

4.1 Comparison of the analysis accuracy

The data mining model is used to analyze the risk situation in the analysis of competency analysis of different innovation and entrepreneurial teams, and the results are compared with the actual pressure results. The results are compared with the results obtained by using this model and the stepwise linear regression method. The results is described in Table 1.

Table 1
The accuracy comparison of two models

Experiment Actual Model in Stepwise linear

times result this paper regression model

1 0.987 0.989 0.912

2 0.983 0.988 0.829

3 0.982 0.976 0.884

4 0.913 0.904 0.865

5 0.939 0.942 0.911

6 0.925 0.929 0.892

7 0.928 0.926 0.884

8 0.966 0.971 0.907

9 0.972 0.975 0.912

10 0.963 0.969 0.913

Experiment	Actual	Model in	Stepwise linear
1	0.987	0.989	0.912
2	0.983	0.988	0.829
3	0.982	0.976	0.884
4	0.913	0.904	0.865
5	0.939	0.942	0.911
6	0.925	0.929	0.892
7	0.928	0.926	0.884
8	0.966	0.971	0.907
9	0.972	0.975	0.912
10	0.963	0.969	0.913

From Table 1, we can be seen that using this model in the paper to carry out analysis on risk situation for innovation and entrepreneurship team competency, the obtained results and the actual results are not much difference, while based on stepwise linear regression, compared with innovation entrepreneurial team competency analysis and the model in the paper, the actual analysis of the results vary widely, indicating that the analysis of this model are more accurate.

In order to further verify the accuracy of the model, a total of 60 experiments were conducted, and the average value of every 10 experiments was obtained. The results were compared with the accuracy of the model and the model based on the stepwise linear regression method. It can be seen that compared with the accuracy of innovation and entrepreneurial team competency analysis based on stepwise linear regression model, that of this model in this paper is significantly higher, and it is always higher than the data mining model based on the stepwise linear regression method. And the accuracy curve of this model is more stable, indicating that the model not only has high accuracy, but also has high stability, validating the effective of the model.

4.2 Comparison of the analysis accuracy

In order to further verify the validity of the model, the time required to complete the same data mining analysis based on the stepwise linear regression method and method in this paper are described. The results are described in Table 2.

Table 2
Comparison result of two model efficiency

Experimental Model in this Stepwise linear regression

times paper/s model/s

1 1.4 3.6

2 1.6 4.2

3 1.7 3.9

4 1.5 4.6

5 1.8 4.8

6 2.1 4.1

7 2.3 5.2

8 1.9 5.6

Experimental	Model in this	Stepwise linear regression
1	1.4	3.6
2	1.6	4.2
3	1.7	3.9
4	1.5	4.6
5	1.8	4.8
6	2.1	4.1
7	2.3	5.2
8	1.9	5.6

As shown in Table 2, it can be seen that the time required for this model is significantly lower than that based on the stepwise linear regression model, and it is always lower than the stepwise linear regression model, which shows that the model has high efficiency.

According to the feedback evaluation of the medical student ability evaluation model, reference to relevant literature, we construct the entrepreneurial innovation capability system as shown in Table 3.

Table 3

Entrepreneurial innovation capability system

Level 1 indicators	Level 2 indicators	Level 3 indicators
Entrepre-neurial innovation capability system	Self ability	Innovation ability, knowledge quality, psychological quality.
	Working philosophy	Collective consciousness, service consciousness, democratic consciousness, sense of responsibility, innovation and sense of competition.
	Ability to work	Planning organizational capacity, ability to perform, teamwork, flexibility and innovation, Self - control ability, language ability, observation and analysis of problem - solving ability.
	Work performance	Work in a timely manner, accurate work, outstanding achievements.

Table 4

Average satisfaction value

Evaluation index	Satisfaction mean	Evaluation index	Satisfaction mean
Innovation ability	3.82	Innovative sense of competition	3.33
Knowledge quality	3.66	Responsibility	3.97
Psychological quality	3.77	Planning organizational skills	3.67
Collective consciousness	3.50	Execution ability	3.76
Service awareness	3.86	Teamwork	3.77
Democratic consciousness	3.44	Flexible and innovative ability	3.32
Self - control	3.54	Work in a timely manner	3.74
Language expression skills	3.74	Work accurate	3.24
Observe and analyze problems	3.27	Outstanding results	3.47

Table 5

Evaluation index importance degree

Evaluation index	Importance	Evaluation index	Impor-tance
Innovation ability	0.19	Innovative sense of competition	0.14
Knowledge quality	0.13	Responsibility	0.21
Psychological quality	0.19	Planning organizational skills	0.10
Collective consciousness	0.11	Execution ability	0.06
Service awareness	0.23	Teamwork	0.21
Democratic consciousness	0.20	Flexible and innovative ability	0.07
Self - control	0.20	Analyze problems	0.19
Language expression skills	0.05	Work accurate	0.17
Observe and analyze problems	0.14	Outstanding results	0.14

The questionnaire was distributed to a class teacher, counselor, classmate, teacher and college leader in a university. a total of 300 questionnaires were distributed and 286 valid questionnaires were collected. the survey of the survey indicators are designed with reference to the literature has been the scale, after pre-trial and small-scale interviews, the questionnaire has considerable content validity. in this study, “Cronbach’s alpha” coefficients were used to determine the internal consistency of the questionnaire, according to Nunnally’s suggestion, “Cronbach’s alpha” coefficient is acceptable as long as it is greater than 0. 7. from the reliability point of view, the consistency coefficient of each variable is greater than 0.7, the questionnaire has a considerable content reliability.

The purpose of this paper is to evaluate the satisfaction of students’, to determine the urgent need to improve the factors, so we research on the quality of entrepreneurial ability evaluation of medical college students for the importance of analysis. In this paper, the principal component analysis method is a statistical method to convert multiple indicators into a few comprehensive indexes and maintain a large amount of information on the original index. Through the analysis of the many factors (indicators) that affect the overall goal, the sample variance matrix of the original index and the eigenvalue and eigenvector of the matrix are solved.

5 Model performance testing

The goal of innovative entrepreneurship education is to train high-quality compound talents. As one of the modes of innovative entrepreneurship education, the project participatory Innovation education has been recognized and practiced by colleges and universities in recent years. College students can participate in the project team through research, science and technology competitions, practical training and other ways, from the traditional knowledge teaching passive learning to Team independent interactive learning, highlighting the practical characteristics of innovative entrepreneurship education, through the “dry middle school” to enable students to learn knowledge and skills. This article takes the innovative Entrepreneurship Training Program project of university students as the angle of view to study the innovative entrepreneurship education model.

In order to verify the feasibility and effectiveness of classroom teaching behavior analysis framework based on artificial intelligence technology, the black box of classroom teaching process is opened by using the emerging artificial intelligence technology to realize the process of behavior collection, marking and analysis. Visualization of the real behavior of teachers and students in the classroom, reduce the workload of subsequent manual analysis, realize technology driven analysis teaching, and better improve the quality of education and teaching. Based on the analysis of technology driven classroom teaching behavior, this study proposes an analysis system of classroom teaching behavior based on artificial intelligence technology. class face detection procees as show in Fig. 4.

Fig.4

Class face detection procees.

Teaching video is the carrier of recording classroom teaching behavior. In order to facilitate intelligent analysis, it is usually necessary to collect real video of classroom teaching scene. It mainly relies on the infrastructure layer of the analysis system of classroom teaching behavior based on artificial intelligence technology to collect the teaching video obtained by the equipment, and then processes and analyzes the collected data through the data layer, and finally forms the analysis result visualization to the managers, teachers, students, parents and other interested groups. Based on the data required by the observation dimension and evaluation dimension of the analysis framework, preprocess, determine the specific analysis dimension, select the corresponding artificial intelligence technology including voice recognition, face detection, etc. for analysis, get the frequency, time and other statistical data of various aspects of behavior, speech, emotion, etc. based on these statistical data, form a visual analysis of the whole classroom teaching behavior result. Classroom intelligence analysis as show in Fig. 5.

Fig.5

Classroom intelligence analysis.

After getting the text and image information, the convolution neural network is used to analyze and process the image information, and the cyclic neural network is used to process the voice information, so as to recognize and analyze the specific activities and speech behaviors. Because emotional behavior is mainly reflected in facial expression and speech behavior, we will first use face recognition technology to process, get the face image analysis of students and teachers, and then extract the speech features and text features in the speech flow, synthesize the analysis results of both, get the analysis of teachers and students’ emotional behavior. Image information analysis as show in Fig. 6.

Fig.6

Image information analysis.

Through the data processing and analysis process, the corresponding processing of data, combined with the evaluation dimension of the analysis framework of classroom teaching behavior based on artificial intelligence technology, obtain the evaluation dimension data of each index item, combined with the data analysis of the activity rate of teachers and students, the language rate of teachers and students in classroom teaching, and then analyze the entire classroom. The final generated data can be visualized to the application level managers, teachers, students and parents and other education related groups to help form a health education ecosystem.

The analysis framework of classroom teaching behavior based on artificial intelligence technology involves speech recognition, image recognition and machine learning technology of artificial intelligence to recognize and analyze classroom teaching behavior in specific education scenarios. At present, the behavior identification and analysis in other fields, such as shopping mall security system and public security system, has been very mature. Due to the particularity of education scene, there are not a large number of data samples for free use in the classroom voice or behavior data. If we label the samples and train the database by ourselves, the analysis effect will be greatly reduced due to the limitation of research period.

Teaching video data usually contains image data and voice data, only image analysis can not fully show the whole picture of teachers and students’ emotional behavior. Therefore, this study also aims at emotional analysis of voice data, which is used to assist emotional analysis of video image data of example teaching. The feedback scores of different time stages are shown in Figs. 7–8.

Fig.7

First stage feedback score.

Fig.8

Second stage feedback score.

From the overall analysis results, this study can get the change trend of middle school students and teachers, emotional evaluation scores and a series of classroom teaching evaluation indicators, such as positive emotion frequency, positive emotion rate, etc. At the same time, it monitors the changes of teachers’ and students’ emotions in classroom teaching. From the perspective of the comprehensive evaluation of classroom emotions, this is a more active classroom, but the changes of emotional scores are relatively large at some times in the classroom, which needs to be analyzed in combination with specific teaching subjects. For example, when a certain period of classroom time, students’ emotions are more negative, teachers can consider optimizing this part of teaching.

Comprehensive classroom teaching analysis requires comprehensive data collection and data cross analysis with the help of voice recognition, image recognition, human detection and action recognition. Limited by the existing data collection means and data analysis technology, at present, teachers and students at specific time can not participate in the analysis of the situation, so it is impossible to conduct a complete analysis and evaluation of the sample video. However, from the perspective of current emotional analysis, the current classroom evaluation can be assisted by intelligent technology and means to improve the evaluation efficiency of classroom teaching.

6 Analysis of results

This paper combines theoretical research and practical research methods to provide a new perspective for the existing classroom teaching behavior analysis and evaluation system. From the perspective of AI technology driven teaching, this paper aims to build a scientific framework for classroom teaching analysis, reduce the repetitive and inefficient work of human analysis by using AI technology, effectively improve the efficiency and quality of classroom teaching behavior analysis, and promote the diversified development of classroom evaluation methods, the development of teachers’ Professional skills and the improvement of teaching quality.

Teaching is a professional activity, which needs to be evaluated according to scientific evaluation methods. If it is only a traditional one-sided observation, it will only lead to one-sided evaluation, and it can not be objectively understood. At the same time, manual classification through teaching videos will be too immersed in people’s subjective evaluation, resulting in evaluation bias. At this time, the value of technology participation in classroom observation is immediately highlighted. Although the current technology can not completely replace manual observation and classification, it can save human and material resources to a certain extent and improve the evaluation efficiency.

The development of innovation and entrepreneurship education in Colleges and universities is the inevitable result of the development mode of global colleges and universities. In the whole process of development, colleges and universities should closely combine the cultivation of scientific research ability and their own quality, so as to cultivate comprehensive talents for the society and provide assistance for promoting the innovation and entrepreneurship ability of college students. Colleges and universities must change the traditional education mode and concept in view of the increasingly severe employment situation. Therefore, colleges and universities need to change their roles, incorporate the innovation and entrepreneurship education into their education concepts, and carry out the creative employment and innovative employment post education for college students, so that the graduates have the choice and consciousness of innovation and entrepreneurship. At this time, the graduates have the identity of both post creators and job seekers, and improve the graduates’ own Value, to find their own foothold.

In the Internet plus background, we need more innovative and technical personnel. Therefore, our university’s goal of cultivating talents should be defined as cultivating entrepreneurs with creativity, self innovation and even job creation. Our entrepreneurship education should not only focus on the design of subject courses, but also integrate the entrepreneurship education with the Internet entrepreneurship University. The education taught by the entrepreneurship university is not only to help college students to seek a stable career, but also to help college students develop their own potential, cultivate entrepreneurship awareness, improve entrepreneurship quality and ability. Compared with the established traditional university, when the university changes to the entrepreneurial university, it should not only have the traditional core operation mode such as classroom teaching, but also have the characteristics that the traditional university does not have, that is, directly let students participate in entrepreneurial activities and business operations.

7 Conclusion

Internet plus promotes the transformation of entrepreneurship education for college students. The arrival of the era of big data has played an important role in the entrepreneurship education of college students. In the past, due to the lack of data support, the traditional entrepreneurship education curriculum can not achieve the desired results. The big data analysis brought by Internet plus has effectively compensated for this deficiency, making the entrepreneurship education curriculum in the new era more targeted and more satisfying the individual requirements of students. Therefore, under the background of Internet plus, entrepreneurship education for college students needs to integrate text image, video multimedia and classic teaching cases. Teachers also need to learn how to use Internet technologies such as microblog and wechat to improve their own education level, which not only puts forward higher requirements for college students’ Entrepreneurship Education workers, but also faces the entrepreneurship education for college students Important topics.

References

Leon

M.C.

and Jukka

, Searching for big data: How incumbents explore a possible adoption of big data technologies, Scandinavian Journal of Management 34(2) (2018), 129–140.

Samiya

, Xiufeng

, Kashish

A.S.

and Mansaf

, A survey on scholarly data: From big data perspective, Information Processing & Management 53(4) (2017), 923–944.

Puth

M.T.

, Neuhäuser

and Ruxton

G.D.

, Effective use of Pearson’s product– moment correlation coefficient, Animal Behaviour 93 (2014), 183–189.

Tabachnick

B.G.

and Fidell

L.S.

, Using multivariate analysis, California State University Northridge: Harper Collins College Publishers 43(12) (2001), 55–60.

, Hu

and Wang

, Research on building development and comprehensive evaluation operation mechanism, Building Economy 3 (2007), 36–38 green.

Pereira

and Ferreira

, Identification of IT Value Management Practices and Resources in COBIT 5, RISTI-Revista Ib érica de Sistemas e Tecnologias de Informação (15) (2015), 17–33.

Subbu

K.P.

and Athanasios

, Big Data for Context Aware Computing-Perspectives and Challenges, Big Data Research 10(7) (2017), 33–43.

Shaoyan

and Yunjian

, Research on Application of Multimedia Technology in College Physical Education, Procedia Engineering 29(2) (2012), 4213–4217.

Lee

, Dong-hyuk

, Canh

, Ui

and Dat

, Pre-liminary design and fabrication of smart handheld surgical tool with tactile feedback, In2013 IEEE RO-MAN (2013), 76–80.

10.

, Tiago

and Milene

, User experience design and agile development: From theory to practice, Journal of Software Engineering and Applications 5(10) (2012), 743.

11.

Moreira

, Ana

and Atif

, A pattern-based approach for GUI modeling and testing, In2013 IEEE 24th International Symposium on Software Reliability Engineering (ISSRE) (2013), 288–297.

12.

Derboven

, Dries

and Mathijs

, Semiotic analysis of multi-touch interface design: The Mu Table case study, International Journal of Human-Computer Studies 70(10) (2012), 714–728.

13.

Jiang

, Wu

T.Y.

, Zheng

, et al., Learning Summary Statistic for Approximate Bayesian Computation via Deep Neural Network, Statistica Sinica 27(4) (2015), 11–13.

14.

Siyuan

, Wei

, Yuanxin

, et al., A Deep Network Architecture for Super-Resolution-Aided Hyperspectral Image Classification With Classwise Loss, IEEE Transactions on Geoscience and Remote Sensing 14(3) (2018), 1–14.

15.

Kang

, Kim

J.M.

, et al., Reliable Fault Diagnosis for Low-Speed Bearings Using Individually Trained Support Vector Machines With Kernel Discriminative Feature Analysis, IEEE Transactions on Power Electronics 30(5) (2015), 2786–2797.

16.

Atienza

F.A.

, Morgado

and Fernandezmartinez

, et al., Detection of Life-Threatening Arrhythmias Using Feature Selection and Support Vector Machines, IEEE Trans Biomed Eng 61(3) (2014), 832–840.

17.

Jegadeeshwaran

and Sugumaran

, Fault diagnosis of automobile hydraulic brake system using statistical features and support vector machines, Mechanical Systems and Signal Processing 52-53 (2015), 436–446.

18.

De Yong

, Bhowmik,s. and Magnago,F., An effective Power Quality classifier using Wavelet Transform and Support Vector Machines Expert Systems with Applications 42(15-16) (2015), 6075–6081.

19.

Halder

, Bensch

, Mellinger

, et al., Online artifact removal for brain-computer interfaces using support vector machines and blind source separation, (Research Article) Computational Intelligence & Neuroscience 2007(10) (2014), 8–19.

20.

Claesen

, De Smet

, Suykens,J.A.K., et al., Ensemble SVM: A Library for Ensemble Learning Using Support Vector Machines, Journal of Machine Learning Research 15(1) (2014), 141–145.

21.

Gao

, Li

, Khodadadzadeh

, et al., Subspace-Based Support Vector Machines for Hyperspectral Image Classification, IEEE Geoscience and Remote Sensing Letters 12(2) (2015), 349–353.

22.

Ifenthaler

and Widanapathirana

, Development and Validation of a Learning Analytics Framework: Two Case Studies Using Support Vector Machines Technology, Knowledge and Learning 19(1-2) (2014), 221–240.

23.

Rienow

and Goetzke

, Supporting SLEUTH – Enhancing a cellular automaton with support vector machines for urban growth modeling Computers, Environment and Urban Systems 49 (2015), 66–81.

24.

Kercheval

A.N.

and Zhang

, Modelling high-frequency limit order book dynamics with support vector machines, Quantitative Finance 15(8) (2015), 1–15.

25.

Majid

, Ali

, Iqbal

, et al., Prediction of human breast and colon cancers from imbalanced data using nearest neighbor and support vector machines, Computer Methods & Programs in Biomedicine 113(3) (2014), 792–808.