3.1 Datasets and preprocessing

This study employs several datasets to evaluate the performance of the DKSE-Net model in the environmental art image classification task. These datasets, which cover images from different cultural backgrounds and diverse artistic styles, are designed to test the model’s ability to extract and fuse complex visual features and multicultural information.

To ensure the model’s comprehensiveness and generalization ability, four representative environmental art image datasets were selected, including:

1. Dataset A [31]: The environmental image dataset, encompassing streets, buildings, mountains, glaciers, trees, and their corresponding images, involving over 7000 images. Dataset address: https://www.kaggle.com/datasets/theblackmamba31/landscape-image-colorization.
2. Dataset B [32]: The art image dataset. Dataset source: https://rusmuseumvrm.ru/collections/index.php?lang=en, covering approximately 9000 images from different cultures and art genres. 1767 environmental art images are selected and divided into three categories, each containing various styles of environmental art images (sketches and watercolors, paintings, graphic arts).
3. Dataset C [33]: The environmental art image dataset. It sources https://picsum.photos/. This includes 400 images, each with a unique style, aimed at testing the model’s performance in environmental art image classification.
4. Dataset D [34]: The Earth as Art dataset. Dataset source: https://www.kaggle.com/datasets/aymanhassan121/earth-as-art-nasa-dataset?resource=download. The dataset contains images of different landforms and environments of the Earth and shows the artistic beauty of the Earth’s natural landscape. This dataset meets the requirements of the environmental art image classification task, and can better study the fusion of complex cultural backgrounds and the influence of these backgrounds on image classification.

In the data preprocessing stage, a series of steps are adopted to normalize and enhance the dataset, to improve the model’s training effect and generalization ability. To fit all the input images into the DKSE-Net model’s input layer, the images in all datasets are uniformly adjusted to a size of 224x224 pixels. For image , 224 images are employed to round down the length and width of the image, as described in equations (1) to (3):

(1)

(2)

(3)

W and H represent the width and height of the image; D means the image channel. By the above equations, and refer to the number of segments with 224 in the length and width direction. Finally, n images of size are obtained. In addition, the pixel values of all images are normalized to the range of [0, 1] to eliminate dimensional differences in the input data and improve the model’s training stability and convergence speed.

Due to the small number and uneven distribution of classes in the dataset, data enhancement techniques are applied to the training set to expand the data sample size and improve the model’s robustness. The data enhancement method involves the following. To expand the diversity of the training data, various data augmentation techniques are applied during the model training process, including random flipping, random translation, rotation, color jittering, and lighting variation. Specifically, for each input image, random horizontal and vertical flipping of the image is performed, along with a maximum 20% translation and random rotation at an angle of 30 degrees. Random scaling operations on the images (with a scale factor between 0.8 and 1.2) are conducted to enhance the model’s perception of objects of different sizes. These spatial transformation techniques help simulate different camera angles and object layouts in scenes. Additionally, color jittering is introduced by randomly adjusting the brightness, contrast, and saturation of the images to simulate image capture under different lighting conditions. These augmentation techniques effectively expand the diversity of the dataset by generating images with different visual features, improving the model’s robustness and enabling it to better handle environmental art images containing complex backgrounds. The dataset is proportionally divided into training, validation, and testing sets, with proportions of 70%, 15%, and 15%, respectively. During the partitioning, the distribution of samples of each category in the three subsets should be ensured as evenly as possible to prevent overfitting or underfitting of the model in some categories [35]. The specific dataset partitioning results are exhibited in Table 1.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Partitioning results of the dataset.

https://doi.org/10.1371/journal.pone.0313946.t001

3.2 Architecture of the DKSE-Net model

In the design of the DKSE-Net model, various techniques are employed to enhance the model’s performance and stability. Firstly, dilated convolution is introduced in the initial layers of the model to effectively expand the receptive field of the convolution kernel without increasing computational complexity. By inserting spacing (i.e., “dilation”) between standard convolution kernels, dilated convolution enables the model to capture information within a larger context, especially when processing art images with complex cultural backgrounds. It better captures local and global features, thereby enhancing classification performance. Secondly, L2 regularization is incorporated into the pointwise convolution process to improve the stability and accuracy of the convolution operations. By adding a weight decay term to the loss function, L2 regularization can effectively prevent model overfitting. In this study, a balance between data fitting and model complexity can be achieved by adjusting the regularization coefficient λ, thereby enhancing the model’s generalization capability. Additionally, pointwise convolution is used, employing 1x1 convolution kernels to linearly combine and dimensionally reduce features between channels. The design of pointwise convolution aims to reduce computational complexity while maintaining inter-channel relationships, further enhancing feature diversity and model stability. Finally, the Dropout technique is applied before and after the feature fusion and compression stages to randomly drop out the outputs of some neurons, thus reducing model overfitting. Dropout effectively improves the model’s classification performance in complex cultural backgrounds. The device of the DKSE module fully leverages the advantages of SKNet and SENet, combining adaptive adjustments of multi-scale receptive fields and enhancement of channel features. This enables DKSE-Net to more effectively integrate global and local features of images when processing environmental art images with complicated cultural backgrounds, thereby enhancing classification accuracy and stability. This integrated design makes DKSE-Net perform exceptionally well in environmental art image classification tasks, demonstrating the broad potential of DL technology in the integration of cultural backgrounds.

The DKSE module is a novel module proposed in this study, which aims to effectively integrate the images’ global and local features, to improve the accuracy and stability of environmental art image classification. This module combines the adaptive adjustment characteristics of SKNet’s multi-scale receptive field [36, 37] and SENet’s channel feature enhancement characteristics [38, 39]. The adaptability and flexibility of feature extraction are achieved by introducing a dual kernel strategy. The structure of SKNet and SENet modules is displayed in Fig 1.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. The structure of SKNet and SENet modules.

https://doi.org/10.1371/journal.pone.0313946.g001

In Fig 1, SENet mainly inserts a Sequeeze and Excitation structure into the CNN and changes the input X into output U through the convolution operation. Moreover, the Squeeze operation is a global average pooling [40, 41], which is calculated in equation (4):

(4)

The calculation of the Excitation operation is:

(5)

In equations (4) and (5), the result obtained by Squeeze is z; and z perform a fully connected (FC) layer operation; The dimension of is r represents a scaling parameter, which is to reduce the number of channels and thus reduce the amount of computation. The result of is A ReLU layer is added, then passed through a FC layer with whose dimension is the same as and finally passed through the sigmoid function to obtain S. The weight s obtained earlier and the U obtained by the previous convolution are multiplied to obtain the output which is expressed in equation (6):

(6)

In this study, DKSE-Net consists of four parts: split, squeeze, excitation, and scale, and the expression is provided in equation (7).

(7)

U represents the feature map fused by DKSE; means the global mean pooling operation; refers to the channel compression processing; denotes the channel feature activation operation; N indicates the number of branches in the DKSE module [42–45]. The DKSE-Net model’s structure is revealed in Fig 2.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. The structure of the DKSE-Net model.

https://doi.org/10.1371/journal.pone.0313946.g002

The convolution mapping process of the convolution kernel in the Split module is shown in equations (8) and (9):

(8)

(9)

and denote the mapping processes that go through convolutional mapping, batch standardization, and ReLU excitation function respectively. H, W, and C represent the height, width, and number of channels of the and operation of feature maps [46]. The convolution mapping of the intermediate feature map by each filter on the different branches is given in equations (10) and (11):

(10)

(11)

w means the convolutional filter, which can be expressed as represents the parameter of the c th filter; c^′ refers to the number of channels for the filter and feature map; is the bias; denotes the batch normalization. The ReLU operation is as follows:

(12)

(13)

The weight module Z obtained by the Squeeze and Excitation operations of the Scale module is the image feature information that has been filtered and suppressed by the main features; Z is weighted and mapped with feature maps and and the corresponding channel feature operations are performed, and the expression is:

(14)

refers to the feature information of the c th channel of the feature map [47, 48].

3.3 Improvement and optimization of the DKSE module

In the design of the DKSE module, dilated convolution is used to expand the receptive field of the convolution kernel without increasing computational complexity or losing spatial resolution. The introduction of dilated convolution plays a crucial role in the image classification model presented in this study, especially for environmental art images with complex cultural backgrounds, which can capture meaningful contextual information on a larger scale.

The core idea of dilated convolution is to insert intervals (i.e., “dilation”) into standard convolution kernels to expand the receptive field. Let the size of the standard convolution kernel and dilated convolution kernel be , but with dilations inserted between adjacent kernel elements, where r is the dilation rate. The relationship between the output of dilated convolution and the input X and weight W is defined in equation (15):

(15)

refers to the position of the output feature map. When , the dilated convolution degenerates into a standard convolution. Dilated convolution allows the receptive field of the convolutional kernel to be expanded with fewer computational resources while keeping the feature map size constant. Especially when dealing with artistic images, which often contain local and global information with a high degree of complexity and diversity, dilated convolution helps to better capture this information [49]. Fig 3 shows dilated convolution.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Dilated convolution.

https://doi.org/10.1371/journal.pone.0313946.g003

The introduction of dilated convolution provides the model with a larger receptive field, allowing it to capture important local and global information in images while maintaining spatial resolution. Especially when processing environmental art images with complex cultural backgrounds, dilated convolution effectively captures texture details and cultural features by expanding the receptive range of the convolution kernel. This operation helps the model extract multi-scale contextual information during the classification process, enhancing its ability to recognize complex images. Furthermore, by combining other techniques such as L2 regularization and Dropout, the model further strengthens its generalization capability during the feature fusion and compression stages, thus performing well on diverse datasets. The core role of dilated convolution lies in expanding the receptive field of the convolution kernel by inserting dilations, allowing it to perceive a larger range of contextual information in images without increasing computational costs. For environmental art images, cultural backgrounds often contain rich visual features, such as architectural styles, texture details in artistic styles, color distributions, etc., all of which exhibit multi-scale characteristics. Dilated convolution is particularly suitable for capturing this multi-scale cultural background information. At the pixel level, cultural backgrounds can manifest as local textures (such as painting brushstrokes, and sculptural textures), specific color gradients, or repeating geometric patterns, which usually maintain consistency or exhibit gradual characteristics over a larger range. Dilated convolution, through the expansion of the receptive field, can capture both local and global information. For example, in an art image of traditional architecture, dilated convolution can capture the geometric shapes and pattern arrangements of architectural elements while perceiving the overall hue and texture gradients over a larger range. More specifically, when processing these art images, dilated convolution can effectively combine global features with local features in the cultural background. For instance, in carpets or murals with complex patterns, dilated convolution can not only capture local texture details but also the overall structure of the patterns, which is crucial for enhancing the accuracy of classification models.

In the DKSE module, L2 regularization and pointwise convolution are combined to improve the accuracy and stability of convolution operations and reduce the risk of overfitting. L2 regularization prevents model overfitting by adding a weight decay term to the loss function. The loss function for L2 regularization reads:

(16)

represents the original loss function; λ denotes the regularization coefficient; means the parameter of the model. By controlling the size of λ, the data fitting is balanced with the complexity of the model to prevent overfitting [50].

Pointwise convolution refers to the use of 1x1 convolution kernels to perform pointwise convolution operations on feature maps. The main role of pointwise convolution in the DKSE module is to perform linear combination and dimensionality reduction of features between channels. The output of the pointwise convolution is given in equation (17):

(17)

is the 1x1 convolution kernel. Pointwise convolution positively reduces computational complexity while maintaining the interrelationships between channels. In the DKSE module, pointwise convolution is used to constrain feature combinations between channels, while L2 regularization is used to prevent model overfitting. Through this combination, the model’s stability and generalization ability can be improved without losing feature diversity [51].

Dropout is an effective regularization technique that randomly discards some neurons during training to prevent model overfitting and improve generalization ability. In the DKSE module, Dropout is used before and after the feature fusion and compression phases. The basic idea of dropout is to randomly set the output of a subset of neurons to zero in each forward propagation, thus reducing the interdependencies between nodes. Assuming that the activation value of the output neuron is , the dropout operation can be written as equation (18):

(18)

represents a random variable that follows a Bernoulli distribution; where p is the retention rate. In the DKSE module, Dropout techniques are used in both the feature fusion (before and after global mean pooling) and feature reweighting stages. By randomly discarding some features, the model can prevent overfitting to the training data, thereby improving its classification performance in complex cultural backgrounds.

To address the potential performance degradation of the DKSE-Net model in complex scenes, this study further introduces adversarial training, optimized attention mechanisms, and data augmentation strategies to enhance the model’s robustness in noisy and complex scenes. Firstly, adversarial training is introduced by generating adversarial samples, and exposing the model to image data with noise and interference during the training, thereby learning more robust feature representations. Adversarial sample generation employs the Fast Gradient Sign Method (FGSM) technique, which adds minor perturbations to the input images to create samples with similar visual features but more challenging, as calculated in equation 19:

(19)

represents the adversarial sample; ϵ refers to the amplitude of the perturbation; L means the loss function; denotes the output of the model. In addition, the self-attention is introduced, and the attention layer is added to the DKSE module to dynamically adjust the model’s attention to the local area of the image. By learning the dependencies between different features, the weights of features are adaptively adjusted, to enhance the model’s ability to identify complex and noisy images.

3.4 Model lightweight strategies in resource-constrained environments

To address deployment issues on resource-constrained devices, various model lightweight strategies have been adopted in the DKSE-Net model, including:

1. (1) Model Pruning: By analyzing the weight distribution of each convolutional layer, convolution kernels that contribute less to the model’s performance are removed, thereby reducing computational complexity.
2. (2) Quantization: Floating-point weights are converted to l low-bit integers (8-bit) to reduce storage requirements and accelerate the inference process.
3. (3) Knowledge Distillation: A small student model is trained to approach the performance of a large teacher model, significantly reducing the number of parameters.

Specifically, in terms of convolution kernel adjustments, lightweight adjustments are achieved as follows:

1. (1) Adjustment of convolution kernel size: In some layers, the original 3x3 convolution kernels are replaced with 1x1 convolution kernels to reduce the number of parameters, especially in subsequent convolutional layers where feature mapping dimensionality reduction is performed.
2. (2) Reduction of convolutional layers: Combining empirical rules, redundant convolutional layers are removed to simplify the network structure and reduce computational load.

3.5 Training process and parameter setting

During the training of the DKSE-Net model, the Adam optimizer and cross-entropy loss function are used. The Adam optimizer combines momentum and adaptive learning rates, allowing the learning rate for each parameter to be dynamically adjusted based on its historical gradients, thereby improving the efficiency and stability of training. Additionally, the cross-entropy loss function is employed to evaluate the model’s performance in multi-class classification tasks, aiming to effectively reduce the error between the model’s predicted categories and the true categories, increasing classification accuracy. The cross-entropy loss function is described in equation (19):

(19)

N refers to the batch size; C is the number of categories; denotes the true category label of the i th sample; represents the model’s prediction probability that the i th sample belongs to category c.

The Adaptive Moment Estimation (Adam) optimization algorithm is adopted to improve the model’s training efficiency and convergence speed. Adam combines the advantages of Momentum and RMSprop to accelerate convergence by adaptively adjusting the learning rate of each parameter. The updates of the ADAM optimization algorithm are as follows:

(20)

(21)

(22)

(23)

refers to the parameter; and are the first-order and second-order momentum estimation, respectively; ; ; is the small constant to prevent the dividing error; α represents the learning rate.

Table 2 lists the main hardware parameters and training configurations of the experiment.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. The primary hardware parameters and training configuration of the experiment.

https://doi.org/10.1371/journal.pone.0313946.t002

4.1 Performance evaluation of algorithms

The DKSE-Net model’s overall performance on verification and test sets is plotted in Fig 4.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Overall performance of the DKSE-Net model on verification set and test set.

https://doi.org/10.1371/journal.pone.0313946.g004

Fig 4 displays the performance of the DKSE-Net model on the training, validation, and testing sets of four different datasets (A, B, C, D). In dataset A, the classification accuracy on the training set reaches 93.5%, while the accuracy on the validation and testing sets are 91.2% and 92.7%, respectively. This indicates that the model can effectively learn image features during training and demonstrate strong generalization capabilities on unseen data, particularly with its performance on the testing set. For dataset B, the accuracy of the training set is 91%, but it drops to 89.5% and 90.5% on the validation and testing sets, respectively. This decline may reflect the complexity and diversity of the dataset, posing certain challenges for the model when processing these samples. Dataset C shows relatively high accuracy for both the training and validation sets, at 92% and 90.8%, respectively, with an accuracy on the testing set of 91.3%, demonstrating the model’s strong adaptability to this dataset. Lastly, dataset D performs the best in all stages, with a training set accuracy as high as 94.5%, with accuracies on validation and testing sets of 92.8% and 93.2%, respectively. This result suggests that the image features of dataset D are highly compatible with the feature extraction capabilities of the DKSE-Net model, leading to particularly outstanding classification performance on this dataset.

To further verify the DKSE-Net model’s superiority, it is compared with several traditional and advanced CNN models, encompassing VGG16, ResNet50, InceptionV3, and MobileNetV2. The comparison of classification performance between DKSE-Net and the baseline model is illustrated in Fig 5.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Comparison of classification performance between DKSE-Net and baseline models.

https://doi.org/10.1371/journal.pone.0313946.g005

Fig 5 provides a detailed comparison of the classification performance of different models in the task of environmental art image classification, with specific metrics including accuracy, precision, recall, and F1 score. Through this data, the strengths and weaknesses of each model can be more clearly assessed. Firstly, in terms of accuracy, the DKSE-Net model achieves an accuracy as high as 92.7%, significantly surpassing all other benchmark models. Specifically, the VGG16 model has an accuracy of 85.3%, ResNet50 achieves 89.4%, InceptionV3 reaches 90.8%, and MobileNetV2 is 88.7%. These data indicate that the DKSE-Net is effective at capturing more precise feature information when dealing with complex environmental art images, demonstrating a stronger classification capability. Secondly, regarding precision, the DKSE-Net’s precision attains 92.4%, also higher than the other models. The precisions of VGG16, ResNet50, InceptionV3, MobileNetV2 are 84.8%, 88.9%, 90.2%, and 88.4%. Precision measures the proportion of actual positive samples among all those classified as positive by the model. A higher precision implies that the DKSE-Net is more accurate in predicting positive classes, reducing the incidence of misclassifications, which is significant for reducing erroneous categorizations in practical applications. Next, in terms of recall performance, the DKSE-Net’s recall is 92.6%, again outperforming the other models. The recalls for VGG16, ResNet50, InceptionV3, and MobileNetV2 are 85.1%, 89.2%, 90.5%, and 88.6%, respectively. Recall reflects the model’s ability to identify all actual positive samples. A higher recall indicates that the DKSE-Net can more effectively capture important features in environmental art images, ensuring that the majority of positive samples are correctly classified. Finally, regarding the F1 score, the DKSE-Net continues to lead the other models with a score of 92.5%. The F1 scores for VGG16, ResNet50, InceptionV3, and MobileNetV2 are 84.9%, 89%, 90.3%, and 88.5%, respectively. The F1 score is the harmonic mean of precision and recall, taking into account both the precision and recall capabilities of the model. The DKSE-Net’s excellent performance in this regard indicates that it can accurately identify positive samples and effectively reduce false negatives and false positives, fully demonstrating its comprehensive advantage in the task of environmental art image classification. The DKSE-Net model performs exceptionally well across all metrics, including recall, accuracy, precision, and F1 score, far exceeding traditional models. These data not only validate the effectiveness and stability of the DKSE-Net in classifying environmental art images with complex cultural backgrounds but also offer a strong reference for future research in this field.

The performance of DKSE-Net on different datasets is indicated in Fig 6.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. Performance of DKSE-Net on different datasets.

https://doi.org/10.1371/journal.pone.0313946.g006

Fig 6 compares the overall performance of DKSE-Net across four different datasets. The accuracy of dataset A is 92.47%, indicating that the environmental art image features of this dataset are well captured and classified by the model. In contrast, dataset B has an accuracy of 90.33%, suggesting that this dataset may contain more complex or diverse cultural backgrounds, making the classification task more challenging. Dataset C’s accuracy of 91.37% reflects the model’s ability to maintain high classification capabilities under various artistic styles. Dataset D’s performance is once again prominent, with an accuracy of 93.2%, showing the effective integration of the dataset’s image features with the model. Overall, the consistent performance of the DKSE-Net model across different datasets demonstrates its good adaptability and stability, effectively addressing environmental art images with different cultural backgrounds and styles.

Additional experiments are conducted to further verify the enhancement of the DKSE-Net model’s performance by transfer learning methods. In addition, comparisons are made between the original model, the model with data augmentation methods, and the model after transfer learning. Dataset D is introduced to the experiments, and the classification accuracies of different models on the four datasets (A, B, C, and D) are shown in Table 3. It can be found that the model’s performance on all datasets significantly improves after the introduction of data augmentation and transfer learning. Specifically, on dataset A, the classification accuracy of the DKSE-Net model after transfer learning reaches 94.0%, an increase of 1.3 percentage points compared to the original model. On dataset B, which includes images with complex cultural backgrounds, the model’s accuracy improves from 89.5% to 91.1%, indicating that transfer learning and data augmentation strategies effectively enhance the model’s generalization capabilities in complex cultural backgrounds. On dataset C, where the environmental art image style is relatively uniform, the model after transfer learning also attains a significant improvement, with an accuracy of 92.4%. It is particularly noteworthy that the model performs excellently on the newly added dataset D, achieving a classification accuracy of 93.2%. This result indicates that the introduction of transfer learning methods enables the DKSE-Net model to better handle images containing diverse artistic styles and cultural backgrounds, remarkably enhancing the model’s generalization capabilities and stability.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Classification accuracy of different models on four datasets (A, B, C, and D).

https://doi.org/10.1371/journal.pone.0313946.t003

To verify the influence of the lightweight strategy in Section 2.4 on the lightweight effect of the model, experiments are conducted to compare the performance of DKSE-Net with other lightweight models (such as MobileNet) in resource-constrained environments. The results are listed in Table 4:

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Performance comparison of different models in resource-constrained environments.

https://doi.org/10.1371/journal.pone.0313946.t004

Table 4 shows that the lightweight-optimized DKSE-Net model outperforms traditional VGG16 and ResNet50 in terms of the number of parameters and inference time. Meanwhile, it still maintains a high level of classification accuracy (91.2%), only slightly lower than MobileNet (89.5%), but with improved accuracy. This indicates that DKSE-Net has effectively reduced computational complexity through optimization strategies, making it suitable for deployment on resource-constrained devices. Furthermore, to further reduce the model’s computational costs, future research can explore collaborative inference methods based on edge computing and cloud computing. Moreover, combining the lightweight optimization of DKSE-Net enables flexible scheduling of computational tasks between resource-constrained and cloud devices, thereby achieving more efficient deployment.

4.2 Ablation experiment analysis

Fig 7 demonstrates that each module’s contribution to the model’s overall performance is evaluated by removing or modifying the various components of the DKSE module (such as Dropout, dilated convolution, L2 regularization, etc.). Specifically, four ablation experiments are designed, including removing dilated convolution, L2 regularization, ReLU activation, Dropout, and batch normalization. The contribution of each module to DKSE-Net performance is drawn in Fig 7.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 7. The contribution of each module to DKSE-Net performance.

https://doi.org/10.1371/journal.pone.0313946.g007

Fig 7 illustrates that the performance of DKSE-Net is optimal. When dilated convolution is removed, the model’s classification accuracy drops to 90.1%, showing that dilated convolution effectively captures local details of images while expanding the receptive field, which plays a vital role in improving classification accuracy. After removing L2 regularization and Dropout, the model’s performance decreases by 1.4% and 1.2%, respectively. It indicates the contribution of L2 regularization in reducing overfitting and enhancing training stability, as well as the importance of Dropout in improving model generalization. The removal of ReLU activation and batch normalization has the most significant effect, with model accuracy dropping to 89.7%, suggesting that activation functions and batch normalization are critical in accelerating model convergence and facilitating classification accuracy.

In this section, the computational complexity, number of parameters, and actual computational time of DKSE-Net are evaluated and compared with other popular models such as VGG16, ResNet50, and InceptionV3. Floating-point operations (FLOPs) and parameter counts are used to measure the model’s computational complexity, and each model’s training and inference time on the same dataset are employed to evaluate the computational cost. The comparison of computational complexity and cost between DKSE-Net and other models is suggested in Fig 8.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 8. Comparison of computational complexity and cost between DKSE-Net and other models.

https://doi.org/10.1371/journal.pone.0313946.g008

In Fig 8, DKSE-Net still has high performance while maintaining low computational complexity and cost. Specifically, the number of parameters in DKSE-Net is only 19.4 million, markedly lower than VGG16’s 138 million, and also below ResNet50 and InceptionV3. The FLOPs of DKSE Net are 4.8 GFLOPs, which is 12.3% lower than InceptionV3. Regarding computational time, the training time of DKSE-Net is 1.1 hours per epoch, while the reasoning time is only 6.8 milliseconds/sheet, which shows the efficiency and computational advantages of the model in practical applications.

This study aims to validate the innovation and practical application value of the DKSE-Net model in various environmental art image classification tasks. It comprehensively demonstrates the performance of DKSE-Net by expanding the range of datasets, introducing more comparative models, and conducting in-depth experimental analysis. The experimental results reveal that the DKSE-Net achieves a classification accuracy of 92.17% on the training set and maintains a high performance on the validation and testing sets, reaching 91.50% and 90.50%, respectively. Compared with classic models such as VGG16, ResNet50, InceptionV3, and MobileNetV2, DKSE-Net significantly improves in classification accuracy, with increases of 7.6, 3.5, and 2.2 percentage points, respectively. Additionally, transfer learning and data augmentation strategies further enhance DKSE-Net’s generalization capabilities on complex backgrounds and diverse images. The experimental results also indicate that under the optimization of different modules (such as dilated convolution, L2 regularization, Dropout, etc.), its computational complexity and inference time are markedly reduced, verifying its application potential in resource-constrained environments. These experimental results not only prove the innovation and practical value of DKSE-Net but also offer strong support for its widespread application in environmental art image classification.

4.3 Parameter sensitivity analysis

The effects of several key hyperparameters, such as learning rate, batch size, Dropout rate, on the DKSE-Net model’s performance are studied. The influence of hyperparameters on DKSE-Net performance is depicted in Fig 9.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 9. Effect of hyperparameters on DKSE-Net performance.

https://doi.org/10.1371/journal.pone.0313946.g009

Fig 9 analyzes the impact of hyperparameter settings on the performance of the DKSE-Net model. Through different settings of learning rate, batch size, and Dropout rate, the model’s accuracy, precision, recall, and F1 score are varied to different extents. Under Setting 1 (learning rate: 0.001, batch size: 32, Dropout rate: 0.5), the model’s accuracy is 92.7%, demonstrating the positive effect of this hyperparameter combination on model performance. In contrast, under Setting 2 (learning rate: 0.0001) and Setting 3 (learning rate: 0.01), the model’s performance significantly decreases, indicating that the choice of learning rate is crucial for the stability of the training process. In Setting 4, by increasing the batch size to 64, the model’s accuracy rises to 93%, showing that a larger batch can provide a more stable gradient estimation, thereby improving model performance. The results of Settings 6 and 7 show that appropriately reducing the Dropout rate (from 0.5 to 0.3 and 0.7) had little effect on model performance, but the slight decrease in Setting 7 also indicates the necessity of Dropout to prevent overfitting. Therefore, through systematic analysis of hyperparameters, it can be concluded that appropriate learning rates and batch sizes are key factors in ensuring the performance of the DKSE-Net model in Fig 10.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 10. The Influence of learning rates and batch sizes on the model’s classification results.

https://doi.org/10.1371/journal.pone.0313946.g010

To evaluate the effects of dilation rate and branch convolution kernel size on the DKSE-Net model’s classification performance, experiments are conducted to compare the classification results of models with different dilation rates and configurations of convolution kernel size, as demonstrated in Table 5.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 5. The Influence of dilation rates and convolution kernel size on the model’s classification results.

https://doi.org/10.1371/journal.pone.0313946.t005

Table 5 investigates the impact of dilation rate and branch convolution kernel size on the classification performance of the DKSE-Net model. When the dilation rate is set to 4, the model using a 3x3 convolution kernel has several parameters of 24.5M, a training time of 50 minutes, and a classification accuracy of 91.2%. However, when using a 5x5 convolution kernel, the number of parameters increases to 28M, the training time extends to 55 minutes, and the accuracy slightly improves to 91.7%. This result indicates that increasing the size of the convolution kernel can enhance the model’s accuracy and increase computational complexity and training time. When the dilation rate is raised to 16, the accuracy using a 3x3 convolution kernel decreases to 90.8%, while the accuracy with a 5x5 convolution kernel slightly rises to 91.3%. This suggests that under a high dilation rate, the choice of convolution kernel size significantly affects the model’s classification performance, thus a balance between precision and computational complexity needs to be found. Overall, the combination of dilation rate and convolution kernel size directly affects the model’s parameter and training efficiency, thereby significantly impacting the final classification outcome.