3.1 Design of ios painting image style classification model for painting teaching

Beginners often lack a solid foundation in drawing and painting techniques, resulting in their works lacking a distinct style. Challenges arise in objectively assessing students' progress in online modern art teaching. Consequently, they encounter difficulties in receiving specific guidance and training, which hampers effective teaching interaction. Consequently, the development of students' art skills may be impeded ^[20,21,22]. To address this issue, a painting style classification model and a painting description generation model are designed using the portable platform IOS. It is used to evaluate students' works in the process of painting teaching, providing targeted support for teaching interaction during the learning process. The style classification model is used to locate the direction of students' future painting practice. The drawing description generation model can classify students with similar descriptions, making it convenient for students to communicate with each other during the learning process.

The design of the painting image style classification model mainly uses image information entropy technology to classify image art styles. The color entropy and block entropy contained in the participants' works are analyzed for the distribution of painting colors and lines. Based on the image information entropy characteristics, the study classifies the participants' painting images into four basic categories, which are simple line drawing category, complex line drawing category, conservative style category and exaggerated style category. Simple line painting refers to the painting of students who lack a coloring foundation and have simple line composition. Complex line painting refers to student paintings that lack a color foundation but have complex and dense line compositions. The conservative style category refers to students' paintings that use monotonous colors, even with colder tones transitioning between color blocks. Exaggerated style refers to the use of colors in students' paintings that are more exaggerated, without too many color blocks, and the overall color system tends to be warm, as shown in Fig. 1.

It is worth noting that it is often difficult for students to notice the objective characteristics contained in their own works during the drawing process. So it is necessary to classify the students' drawings and to conduct targeted teaching interaction. The study uses the IOS programming language to build a resource library of students' drawings. The specific construction is shown in Fig. 2.

The research focus of the iOS programming section is on the Objective-C programming language. It originated from the C language and is divided into four key features: message mechanism, protocol, category, and memory management method. The message mechanism is a mechanism that determines the method to be called at runtime, while the execution method is determined through function calls at compile time. The protocol is a method declaration that cannot provide the implementation method itself, but can be provided through specific references to the protocol. And using this proxy protocol model is also more convenient for maintaining model code. The category is a way to add methods to the current category directly without the need of inheritance. This coding method allows developers to add custom methods to existing categories. Memory management methods are divided into two main categories, i.e. manual memory reference management and automatic memory reference management. Both mechanisms are implemented through the automatic reference counting mechanism ^[23,24,25]. The specific programming language features are shown in Fig. 3.

The study mainly uses the LAB color model when performing the color entropy calculation of the college works. L denotes the luminance information of the painting. A represents the information interval between the transition of a painting from a green area to a magenta area in color. B represents the information interval between the transition of the painting from the blue area to the yellow area. Color entropy technology more accurately describes the color of an image and can fully capture the color features of the image. Therefore, it has a good classification effect in painting style classification. Since the degree of color saturation is an important criterion in the evaluation of paintings, the study is mainly conducted in HSV color mode. The S channel can be used exclusively for analyzing color saturation. Firstly, RGB images are converted into HSV images, and in this image mode, the range of all channels can be simplified within [0,1], as shown in Equation (1).

In (1), $i$ denotes the hue index and $P$ denotes the overall probability of that hue appearing in the picture of the college painting. A weighted gray scale function is introduced as shown in (2).

In turn, (3) can be obtained.

In (3), $H$ represents the hue channel. If the color entropy calculation only considers the hue element, the participants' painting styles are classified as highly similar. This is mainly due to the lack of saturation determination, where the same hue is interpreted as the same color. To address this limitation, the study analyzed the S channel, which is exclusively dedicated to color saturation. Equation (4) can be used to calculate the average color saturation of the college painting images.

In (4), $x$ and $y$ denote the pixel indexes. $S_{x,y} $ denotes the saturation value at the first $\left(x,y\right)$ pixel, which can be used as a weighting function. The color entropy calculation formulae in HSV mode can be combined in the form of (5).

Since the participant's drawing image comprises not only the color component but also the composition element, evaluating and categorizing the participant's drawing based solely on color entropy would not be accurate. Therefore, additional analysis of drawing lines and composition information is required. In this part of the study, the block entropy technique is used for model construction. The two-dimensional entropy model mainly distinguishes images that obtain basic information entropy through a blocking strategy, and then constructs two-dimensional image information entropy. It achieves the effect of two-dimensional student drawing images by minimizing the interference of color information in the images. Supposing the potential segmentation boundary point of the image is $S$, and the probability condition of the image gray level is represented by $pL$, which can be divided into two probability distributions. The first probability distribution is shown in (6).

The second probability distribution is shown in (7).

The $l$ in (7) indicates the number of gray levels. The entropy value after image segmentation is shown in (8).

$E_{A} $ in (8) can be calculated in the form of (9).

In (10), $E_{B} $ can be calculated as the form of (10).

Then the optimal threshold of the segmented image can be expressed in the form of (11).

In addition, in the block entropy section, research mainly uses variance form to characterize the spatial distribution of images and distinguishes students' painting images based on specific rules. The image is divided into small blocks and the information entropy of each individual block is calculated. Finally, the information entropy variance and block grayscale distribution of all blocks are calculated to analyze the connections between the blocks. The overall architecture of the image entropy style classification model is shown in Fig. 4.

3.2 Painting description generation model design

The painting description generation model utilized in this study primarily utilizes a multi-attention mechanism to generate image descriptions via a triple-attention mechanism. The specific structure is shown in Fig. 5.

As shown in Fig. 5, ATT1 represents the attention mechanism for the spatial information of the image. V represents the attention context vector of the previous moment. ATT2 represents the attention mechanism for the hidden unit of the previous moment. ATT3 represents the attention mechanism of the image information with the hidden unit of the LSTM model at the current moment. The study uses the LSTM model to generate description statements for the participant's painting. In the input stage, the weight matrix can be described as (12).

In (12), $Z_{it} $ denotes the weight matrix. $W_{iv} $, $w_{ih} $, and $W_{th} $ denote the parameters obtained by the training method, respectively. $V$ denotes the feature elements of the image space. $h_{t-1} $ denotes the hidden state of the LSTM model at the previous moment. Then the mapping of the attention model in the image mapping space dimension can be expressed in the form of (13).

The attention mechanism can be expressed in the form of (14).

The design of the model first extracts data from words that can describe college painting, and then summarizes it into a database of painting description words. Next, by indexing, the word vector of each word can be set to correspond to its position in the word matrix. The attention mechanism design part adopts the method of feature vector matching to match the text feature vectors of the subject's painting with the visual features of the painting image, forming a vector matrix with correlation. The elements in this matrix are strongly correlated with each other. The study utilizes the text feature and visual feature vectors as input for the attention module, respectively. The final attention vector is then formed through a weighted sum calculation, as demonstrated in (15).

In (15), $w$ represents the output formed after the parameter matrix is input to the SoftMax layer. $v$ represents the weights of the visual feature vector.

4.1 Analysis of the effect of painting image style classification

When testing the effectiveness of painting image style classification in this study, it is first necessary to determine the sample size, type labels for testing applications, and testing indicators for experimental applications, as shown in Table 1.

Table 1 outlines the experiment designed for this study, which comprises 800 participant drawings categorized into four types, each containing 200 images. The labels used in the experiment are classified according to four basic categories commonly found among participants: simple drawing, complex drawing, conservative style, and exaggerated style. Ultimately, four types of labels are formed: conservative simplicity, exaggerated simplicity, conservative complexity, and exaggerated complexity. The study utilizes two primary indexes, accuracy and recall, to evaluate the efficacy of the application. For the model training phase, 100 images from each category in the painting image collection are selected as the training set, with the remaining 100 used as the test set. The accuracy results of the model are shown in Fig. 6.

In Fig. 6, the number of style classification results of the designed painting image style classification model is 58 for conservative simple types, 31 for exaggerated types, 46 for conservative types, and 55 for exaggerated complex types. The conservative simple type has the highest number of classifications, while the exaggerated simple type has the lowest number of classifications. In terms of the correct classification number of military academy students' paintings, the conservative simple type is 51, the exaggerated simple type is 25, the conservative complex type is 37, and the exaggerated complex type is 44. Combining the number of classification results and the number of correct classifications for each type of academy drawing, the classification accuracy of each type can be obtained. Regarding the classification accuracy of military school student paintings, the conservative simple type is 96.55%, the exaggerated simple type is 80.64%, the conservative complex type is 80.43%, and the exaggerated complex type is 80.00%. The classification model used in this study has the highest classification accuracy for participating paintings, with conservative simple models having the highest accuracy and exaggerated complex models having the lowest accuracy. This is because the composition and color of the latter are the most complex. The recall test results of the painting image style classification model are shown in Fig. 7.

In terms of the number of correctly classified pictures in Fig. 7, there are 51 for the conservative simple type, 25 for the exaggerated simple type, 37 for the conservative complex type, and 44 for the exaggerated complex type. The number of correct classifications is the highest for the conservative simple type and the lowest for the exaggerated simple type. In terms of the number of misclassifications, the conservative simple type is 11 times, the exaggerated simple type is 9 times, the conservative complex type is 6 times, and the exaggerated complex type is 9 times. Conservative and simple participants have the highest number of misclassifications in their paintings, while conservative and complex participants have the lowest number. The recall status of the model can be determined by combining the number of correct classifications and the number of incorrect classifications for each category of the model. In terms of the recall rate of the participant chart, the conservative simple type was 85.35%, the exaggerated simple type was 64.00%, the conservative complex type was 83.78%, and the exaggerated complex type was 79.54%. The painting image style classification model in this study has the highest recall rate for conservative and simple types of paintings, and the lowest recall rate for exaggerated and simple types. This is mainly because the composition of exaggerated and minimalist participants' paintings is relatively simple, and the composition information features are not obvious, which is easy to be misclassified.

4.2 Analysis of drawing description generation effect

In the analysis of the effect of drawing description generation, this study primarily compares the same type of models to test the effect. This study compared single-layer attention mechanism, double-layer attention mechanism, and simple LSTM model with the proposed three-layer attention mechanism model. The purpose is to test the accuracy and recall changes of the model under changes in sample size, that is, to determine whether the description generated by the model can accurately describe the painting. The model accuracy results are shown in Fig. 8.

In Fig. 8, the accuracy line of the single-layer attention mechanism model shows a significant decreasing trend from 85.17% to 34.23% as the sample size varies from 1 to 100. The accuracy line of the pure LSTM model also shows a decreasing trend with the increasing sample size from 90.23% to 66.18%. Compared with the single-layer attention mechanism model and the pure LSTM model, the accuracy fold of the two-layer attention mechanism model does not show a significant decreasing trend, but still decreases from 93.24 to 78.63%. The accuracy line of the three-layer attention mechanism painting description generation model also shows a decreasing trend, but only from 97.74% to 84.45%. The highest accuracy point of the drawing description generation model of the three-layer attention mechanism model is the peak in the comparison model. The lowest point is still the peak in the comparative model. The overall folding line accuracy of newspapers is above 80%. The description generated by the three-layer attention mechanism model painting description generation model is more accurate. The variation of the recall rate is shown in Fig. 9.

In Fig. 9, the recall folds of all four models show a decreasing trend as the sample size varies from 1 to 100. The downward trend of recall folding in the single-layer attention mechanism model is most significant compared to the simple LSTM model and the double-layer attention mechanism model. But as the sample size increases, when recall folding drops below 80%, only the three-layer attention mechanism model remains above 80%. The recall rate of the three-layer attention mechanism model is the best among the comparison models, fluctuating between 83.71% and 98.27%. This shows that the three-layer attention mechanism model drawing description generation model has more performance advantages in drawing description generation.