The analysis of art design under improved convolutional neural network based on the Internet of Things technology

0
The analysis of art design under improved convolutional neural network based on the Internet of Things technology

Datasets collection

Data collection is a critical step. First, sensors and cameras built using IoT technology are employed to collect samples of art pieces. These devices are installed in art studios or exhibition spaces to capture and record images and information of the artworks41,42. When collecting samples of art and design works, it is ensured that the samples come from different times, scenes, and artists with diverse styles and backgrounds. This work also compares various art pieces to ensure that the dataset covers a wide range of artistic styles, colors, textures, and other attributes. Additionally, during the preprocessing of input data, it employs various data augmentation techniques such as rotation, flipping, and scaling to further enhance the robustness and generalization ability of the model.

Additionally, creative expression data from students in art design education are collected. Leveraging IoT technology can capture real-time images, videos, or audio recordings of students’ creative artworks and their behaviors and feedback during the creative process. The specific processes for data collection and model establishment are analyzed. Figure 2 illustrates the overall structure of the model construction.

Fig. 2
figure 2

Specific structure diagram for data collection and model construction.

This work employs a meticulous data integration process to ensure the effective fusion of image feature vectors with sensor data, thereby providing comprehensive support for art and design education. First, it ensures that the collected environmental data corresponds to the respective art image samples in time. In order to achieve this, each art piece’s image and its corresponding environmental data are assigned a unique identifier. By matching these identifiers, it ensures that each set of environmental data corresponds to the correct art image for subsequent analysis. Since image feature vectors and sensor data may exist in different magnitudes and scales, direct fusion may lead to the dominance of certain features, affecting the training effectiveness of the model. In order to mitigate this issue, all data undergo min–max normalization to scale the data to a range of 0 to 1. This step helps balance the weights between different features, enhancing the model’s generalization ability and stability. For image data, CNN is utilized to automatically learn and extract key features from the images. CNN, through multiple layers of convolution and pooling operations, captures hierarchical information in images such as edges, textures, and shapes. As for structured sensor data, since it already exists in numerical form, it can be directly used as feature input. Finally, the normalized image feature vectors are concatenated with the environmental data. This step is typically performed before the fully connected layers of the CNN, and the concatenated feature vector serves as the input to the fully connected layers. Through this approach, the model not only learns the visual features of art images but also considers environmental factors during the art creation process, thereby providing a more comprehensive understanding and evaluation of art pieces. The output of the fully connected layers serves as the model’s final decision basis for various applications in art and design education, such as student work assessment and artistic style analysis. Through the aforementioned data integration process, leveraging environmental data collected by IoT technology combined with the powerful image processing capabilities of CNNs, a comprehensive and accurate analytical tool is provided for art and design education. This not only enhances the technological content of art and design education but also provides educators and students with richer teaching and learning resources.

Experimental environment

Environmental data collected from IoT devices is fused with image feature vectors extracted by the CNN model to achieve a comprehensive understanding of the art pieces. The specific fusion process is as follows:

First, \(\textF=[f_1,f_2,\cdots ,f_n]\) is set as the image feature vector extracted by the CNN model, where \(\textn\) represents the dimension of the features. \(\textS=[s_1,s_2,\cdots ,s_m]\) is set as the environmental data vector collected by the IoT devices, where \(\textm\) represents the number of environmental parameters.

This work uses concatenation to fuse these two vectors, resulting in a comprehensive feature vector \(\textV\). The mathematical equation for this operation is:

$$\textV=Concat\left(F,S\right)=[f_1,f_2,\cdots ,f_n,s_1,s_2,\cdots ,s_m]$$

(7)

The comprehensive feature vector \(\textV\) is then input into the fully connected layer for further processing and analysis. This approach allows the model to learn the visual features of art images and to consider specific environmental factors from the art creation process, providing a more thorough understanding and evaluation of the artwork.

In order to achieve real-time data collection and transmission, this work establishes a comprehensive IoT environment. Required resources type and quantity: cameras: three high-resolution cameras (such as 1920 × 1080 pixels) for capturing images of artworks. Costs vary depending on brand and quality, roughly between 3000 and 6000 yuan each. Sensors include but are not limited to light sensors and temperature sensors for collecting environmental data during art creation. Costs for each sensor range approximately from 100 to 500 yuan. Sufficient storage space for saving images and sensor data is required, which can be a local server or cloud storage service. The cost of cloud services depends on storage capacity and usage. Network devices including routers and switches to ensure wireless or wired data transmission. Costs vary depending on device performance, generally between 500 and 2000 yuan. Recommendations for different skill levels: for beginners, starting with a simple single-camera system and gradually adding sensors and cameras is advisable. Existing educational resources such as school labs or libraries can be adopted to reduce initial investment. For educators with some experience, exploring the use of open-source software and hardware platforms such as Raspberry Pi with Arduino sensors to build experimental environments is recommended. For advanced users, developing custom sensors or using advanced image capture techniques such as 3D scanners to obtain richer data is suggested. Open source tools and platforms: node-RED is used for programming and integrating different IoT devices, with active community support. Google Cloud or Microsoft Azure provides free storage and computing resources suitable for small projects and educational purposes. Platforms like Coursera and edX offer courses on IoT and machine learning to help educators enhance their professional knowledge. During the construction of the IoT experimental environment, significant initial investment can be reduced by choosing cost-effective devices and utilizing open-source resources. Furthermore, by providing detailed guides and tutorials, even non-professional educators can establish and maintain such experimental environments. In the model training phase, cross-entropy loss function and L2 regularization technique are used, and hyperparameters are adjusted to optimize model performance. Table 1 provides specific details of the experimental environment setup:

Table 1 Setting of experimental parameter values for the IoT environment.

The table above provides specific parameter values and their settings for the IoT experimental environment setup. Data collection and image capture can be tailored to meet specific requirements by configuring the sensitivity threshold of the light sensor, the sampling frequency and measurement range of the temperature sensor, and the camera’s resolution and frame rate43,44.

Parameters setting

This work adopts an improved CNN model specifically tailored to the needs of art and design education. The model, based on DCNN, undergoes optimization of its network architecture to enhance the accuracy and efficiency of feature extraction. Architectural details of the model include increasing the number of convolutional layers, introducing batch normalization layers, incorporating dropout layers, applying L2 regularization, and selecting appropriate loss functions and optimization algorithms. These enhancements aim to improve the model’s ability to process artistic images and provide more comprehensive and accurate support for art and design education. Compared to existing CNN models, this customized model offers several advantages in the field of art and design education. First, by increasing the number of convolutional layers and neurons, the proposed model can more accurately capture the features of artistic images, thus enhancing its feature extraction capability. Second, the introduction of batch normalization and dropout layers enhances the model’s robustness and generalization ability, enabling it to better adapt to different artistic styles and data variations. Additionally, methods such as early stopping and regularization are employed to effectively suppress overfitting. Early stopping strategy is a conventional approach to handle overfitting. This work stops training when the performance on the validation set no longer improves, preventing the model from being overly trained on the training set. Furthermore, techniques like weight decay, and L1/L2 regularization are set to limit the weights, reducing model complexity and the risk of overfitting. Moreover, data augmentation techniques such as rotation, scaling, jittering, and flipping, are applied to existing artistic design sample images to increase the diversity of training data. These techniques effectively improve the model’s generalization ability and further suppress overfitting. Finally, optimized network structures and parameter settings enable the model to process image data more quickly, providing real-time feedback for art and design education, thereby enhancing teaching effectiveness. An automated parameter optimization strategy is introduced in the model design. Bayesian optimization method is adopted, which is a probabilistic model-based global optimization technique that effectively searches for the optimal solution of model parameters. Through this approach, the proposed model can automatically adjust parameters such as the number of convolutional layers, the number of neurons, and regularization coefficients during the training process, finding the optimal parameter configuration. Table 2 provides specific model parameter settings.

Table 2 Specific parameter settings for the improved CNN model.

The table above provides specific parameter settings for the improved CNN model used for real-time monitoring of art sample data and student creative expression data. The performance of the proposed improved educational model of convolution network combined with IoT (IoT-CNN) for education is then compared with the performance of traditional education and teaching model based on reinforcement learning (RLET) and education and teaching model based on multi-modal deep neural network (MDNN). RELT is a reinforcement learning-based model that learns the optimal policy through interaction with the environment. In art and design education, the RELT model can be used to simulate students’ decision-making behaviors during the creative process and guide them towards improving their creative skills through reward mechanisms. However, the RELT model may require a large amount of training data and a long training time to achieve the desired performance, which may be restrictive in practical educational scenarios. MDNN is a deep learning model for handling multimodal data, capable of simultaneously processing different types of data such as images, text, and audio. In art and design education, MDNN can analyze students’ multidimensional creative performance, including visual artworks, design explanations, and verbal expressions. Although MDNN performs well in handling multimodal data, it may not be as effective as the IoT-CNN approach in terms of real-time performance and personalized feedback.

The “safety detection rate” is an important metric for assessing the model’s ability to identify and respond to potential safety threats in the art design education environment. This metric is defined as the ratio of the number of safety threats correctly identified by the model to the total number of safety threats. Its mathematical expression reads:

$$\textSecurity detection rate=\fracNumber\; of\; correctly\; identified\; security\; threatsTotal\; number\; of\; security\; threats\times 100\%$$

(8)

In the experiment, various data from the art creation environment are first collected using IoT devices, including but not limited to images, temperature, and humidity. The CNN model is then used to analyze these data to identify potential safety threats, such as abnormal behavior or environmental changes. For each identification cycle, the number of safety threats detected by the model is recorded and compared with the actual number of safety threats that have occurred to calculate the safety detection rate.

Performance evaluation

Figure 3 displays the data variation curves for different models concerning the accuracy of recommendations in art design education. Figure 4 shows the data variation curves for different models regarding the safety detection rate in art design education.

Fig. 3
figure 3

Data variation curves depicting the recommendation accuracy for different models in art design education.

Fig. 4
figure 4

Data variation curves depicting the safety detection rates for different models in art design education.

Figure 3 illustrates that for the RLET model, the recommendation accuracy starts relatively low in the initial phase (first 100 iterations). However, it gradually improves with an increase in the number of iterations, eventually stabilizing in later iterations. This suggests that the model may require more training data and iterations to achieve higher recommendation accuracy. The MDNN model exhibits relatively high recommendation accuracy in the initial phase (around 100–200 iterations) but experiences a slight decline in later iterations, maintaining a relatively stable level.

Figure 4 reveals that the IoT-CNN model achieves a relatively high safety detection rate of 53.45% after 100 iterations. However, as the number of iterations continues to increase, its safety detection rate starts to decline, reaching 77.87% after 600 iterations. Table 3 presents the real-time performance of various models in art design education.

Table 3 Real-time performance comparison of various models in art design education.

Table 3 illustrates the real-time performance of different models in art design education. The IoT-CNN model excels in image processing speed, sensor data processing speed, and response time. Specifically, the IoT-CNN model achieves the highest image processing speed at 25 frames per second, showcasing remarkable swiftness compared to other models. Simultaneously, it demonstrates outstanding capabilities in sensor data processing speed, reaching 1200 data points per second. Ultimately, the IoT-CNN model boasts the shortest response time, clocking in at 40 ms, underscoring its exceptional real-time performance.

Additionally, Fig. 5 displays the data trend for image compression quality scores for different models in art design education. Figure 6 presents the data trend for student satisfaction with different models in art design education. Finally, Fig. 7 illustrates the data trend for student creative expression ability scores with different art design education models. The quality score of artistic design image compression is calculated by considering the compression ratio, image clarity, and preservation of key details. First, the compressed images are visually assessed, including their clarity and recognizability. Then, the compression ratio is calculated to determine the degree of image compression. Finally, a comprehensive score is given after considering these factors together. Student satisfaction score is calculated through feedback surveys and assessments targeting students participating in art and design education activities. Student creative performance score is evaluated by assessing students’ creative abilities and performance levels demonstrated during art and design education activities. The assessment includes aesthetic evaluation of student works, creativity, and technical application, along with students’ presentations and verbal statements, to comprehensively evaluate their creative performance.

Fig. 5
figure 5

The data trend for image compression quality scores for different models in art design education.

Fig. 6
figure 6

The data trend for student satisfaction with different models in art design education.

Fig. 7
figure 7

The data trend for student creative expression ability scores with different art design education models.

Figure 5 reveals that the scores for all three models gradually increase as the number of iterations increases. This indicates a significant improvement in the quality of compressing art design images as the models continue to learn and optimize. Specifically, the RLET model’s score increases from an initial value of 2.1 to a final value of 4.0. Although the rate of increase is relatively slow, it exhibits an overall upward trend. The MDNN model’s score increases from an initial value of 4.8 to a final value of 6.5.

Figure 6 illustrates that all three models have made some progress in increasing student satisfaction. The IoT-CNN model, in particular, demonstrates higher performance and stability in art design education, consistently receiving relatively high satisfaction scores. However, further research and practical application are needed to validate the applicability and effectiveness of these models in different teaching scenarios.

Figure 7 suggests that the RLET model has relatively low student creative expression ability scores in the initial phase (first 100 iterations). However, these scores gradually improve with an increase in the number of iterations, eventually stabilizing in later iterations. This suggests that the model can enhance students’ creative expression ability through iterative learning and continuously improve their scores during training. Next, the MDNN model exhibits relatively high student creative expression ability scores in the initial phase (around 100–300 iterations), but they experience a slight decline in later iterations, maintaining a relatively stable level.

Compared to the RELT and MDNN methods, the IoT-CNN method demonstrates advantages in several aspects: by increasing the number of convolutional layers and neurons and introducing batch normalization and dropout layers, IoT-CNN improves the accuracy of image processing and analysis. The IoT-CNN model outperforms RELT and MDNN models in terms of image processing speed, sensor data processing speed, and response time, which is particularly important in educational scenarios requiring real-time feedback. Additionally, the IoT-CNN model performs better in terms of user input response time and user experience ratings, providing a smoother and more satisfactory interactive experience. In summary, the application of the IoT-CNN method in art and design education provides more accurate image analysis and real-time feedback. Moreover, it offers personalized guidance based on students’ creative performance, thereby effectively enhancing teaching quality and students’ creative abilities.

Table 4 presents the interactive performance of different models in art design education.

Table 4 Interactive performance comparison of different models in art design education.

Table 4 illustrates the interactive performance of various models in art design education. The IoT-CNN model has demonstrated outstanding results, achieving a user input response time of 40 ms faster than the other two models. Furthermore, the IoT-CNN model attains the highest user experience score of 9.2, indicating superior user satisfaction with its interactive performance. This achievement can be attributed to incorporating IoT technology and designing an enhanced CNN model, making the model more responsive and efficient in interactions with users.

Figure 8 illustrates the feature extraction for different art styles. It shows the high-level features extracted by the model when processing works of impressionism, cubism, and abstract expressionism. It is evident that artworks of different styles exhibit significant differences in color, shape, and texture, which are effectively captured by the CNN.

Fig. 8
figure 8

Feature extraction for different art styles.

Figure 9 shows the CNN visualization results of student-created works. The analysis highlights the student’s innovations in color usage and composition, providing teachers with targeted feedback basis.

Fig. 9
figure 9

CNN visualization results for student-created works.

Comparison and analysis with existing educational models

In the field of art and design education, common educational models include traditional lecture-based models, problem-based learning (PBL) models, and technology-based interactive learning models. These models each have their strengths. The traditional lecture-based model emphasizes the systematic delivery of fundamental knowledge, the PBL model focuses on practical skills and problem-solving abilities, and the technology-based model utilizes modern information technology to enhance teaching interactivity and learning efficiency. Compared to the aforementioned models, the IoT-CNN model combines IoT technology and deep learning to provide real-time monitoring of art sample data and students’ creative expressions, offering precise feedback and guidance. The model outperforms traditional models in terms of image processing speed, sensor data processing speed, and response time, demonstrating outstanding real-time performance. Through deep learning algorithms, the IoT-CNN model can provide personalized guidance based on students’ creative characteristics, which is challenging for traditional models to achieve. Table 5 displays the differences in key performance indicators between the IoT-CNN model and traditional lecture-based, PBL, and technology-based interactive models. It can be observed that the IoT-CNN model exhibits relative advantages in multiple key performance indicators, particularly in real-time feedback and learning efficiency. Besides, the traditional lecture-based model shows relatively lower levels of student engagement and creative expression. The PBL model demonstrates high levels of learning outcomes and student engagement, but slightly lacks in real-time feedback. The technology-based interactive model performs well in terms of learning outcomes and student engagement but slightly lags behind the PBL model in creative expression.

Table 5 Comparison of IoT-CNN model with traditional lecture model, PBL model, and technology-interactive model on key performance indicators.

Empirical validation

In order to further validate and optimize the IoT-based CNN model proposed, collaborative agreements have been reached with schools and art education training institutions to apply and test the model in their courses. The model runs in real teaching environments, collecting various types of data generated during the teaching process, including but not limited to samples of art design images, environmental data, and student work. Additionally, feedback is collected from both teachers and students to understand the performance and acceptance of the model in practical applications. Empirical research involves periodic satisfaction surveys with teachers and students using the model to understand their acceptance and the actual effects of the model in teaching. Regular communication is maintained with teachers using the model, meticulously recording their experiences, and any issues or suggestions they may have. The learning outcomes of students are tracked to evaluate the effectiveness of the model in enhancing students’ artistic skills. Table 6 presents the empirical validation results. It reveals that the IoT-CNN model has received positive evaluations in art design education, particularly in enhancing students’ innovativeness and expressiveness.

Table 6 Empirical validation results.

This work also aims to understand the direct impact of the proposed model on students’ learning outcomes in art and design education. In order to achieve this, a series of field teaching experiments and qualitative surveys are conducted to collect and analyze data on students’ creative performance and engagement. The IoT-CNN model is deployed in a real teaching environment, and comparisons are made with traditional teaching methods. The experimental design is as follows: the experimental subjects are students majoring in art and design, who are divided into two groups. One group uses traditional teaching methods, while the other group uses the IoT-CNN model as an auxiliary teaching tool. The teaching observations last for eight consecutive weeks. Table 7 presents the comparative results of the model in the field teaching environment.

Table 7 Comparative results of the model in field teaching environment.

In order to gain deeper insights into students’ perceptions and user experiences with the IoT-CNN model, this work conducts the following survey. The survey participants are students using the IoT-CNN model, and Table 8 summarizes the survey results. The survey findings indicate that the majority of students hold a positive attitude towards the IoT-CNN model, believing that it provides a more flexible and personalized learning approach.

Table 8 Survey results summary.

In summary, the consistent results from field teaching experiments and qualitative research indicate that the IoT-CNN model effectively promotes students’ creativity and engagement in art and design education. These findings provide strong evidence for the application of the model in art and design education and point the way for future developments in educational technology.

In order to gain deeper insights into user experience, feedback from art and design teachers and students is collected. Based on the data, an analysis is conducted on users’ perceptions during the model’s usage and their system usability scale (SUS) scores. Table 9 presents the results of the user experience survey. The survey results indicate that users are generally satisfied with the model’s interface, believing that it provides timely and accurate feedback on art and design. Particularly in the extraction of artistic features and analysis of environmental factors, users’ feedback suggests that the model demonstrates good performance. However, there are also areas for improvement. Some users mention that it takes them a relatively long time to familiarize themselves with the operation process when using the model for the first time. This indicates a need for enhanced user-friendliness in the design of subsequent developments. Additionally, users also note that there is further room for optimization in capturing complex textures and subtle color variations when dealing with intricate art images.

Table 9 Survey results of user experience.

Based on user feedback, the following optimizations are made to the model proposed:

(1) Interface design: addressing the steep initial learning curve raised by users, the interface is optimized to simplify the operation steps. More intuitive tutorials and explanations are introduced to help users grasp the model’s usage faster.

(2) Detail capturing: improvements are made to enhance the model’s learning ability for texture and color variations in images. By introducing finer convolutional layers and utilizing advanced activation functions in the CNN, the model’s accuracy in capturing subtle features in artworks is enhanced.

(3) Improvement in interactive performance: optimization is conducted on the system’s response speed, and the backend data processing flow is reconstructed to reduce user waiting time, providing a smoother interactive experience. More efficient algorithms are employed to accelerate data transmission and processing, ensuring timely real-time feedback.

Feedback data from real teaching environments where the model is used are collected through teacher and student interviews and observational studies. The analysis focuses on teaching methods, selection of teaching content, learning motivation, and satisfaction. Table 10 presents the qualitative analysis results of teaching effectiveness:

Table 10 Qualitative analysis results of teaching effectiveness.

This work designs and implements a before-and-after comparative experiment to monitor the changes in students’ creativity and technical proficiency in their works before and after using the model. Table 11 presents the assessment results of creative expression abilities:

Table 11 Assessment results of creative expression abilities.

The qualitative analysis and before-and-after comparative experiment results indicate that the improved CNN model combined with IoT technology has played a positive role in art and design education. Teachers can utilize the instant feedback provided by the model to adjust their teaching strategies, making the teaching content more tailored to students’ individualized needs, thereby enhancing the attractiveness and interactivity of teaching. Meanwhile, through interaction with the model, students can intuitively grasp artistic design concepts, thereby stimulating their creativity and critical thinking abilities. Additionally, there has been a significant improvement in students’ creative expression abilities, particularly in the originality and complexity of their works. These findings demonstrate that the model proposed enhances the information processing capabilities in art and design education and promotes the enhancement of students’ creativity, offering a new teaching aid for the field of art and design education.

link

Leave a Reply

Your email address will not be published. Required fields are marked *