Evaluation and improvement of the creativity of works of art by integrating artificial immune algorithm

Abstract

The creativity of artistic works encompasses multifaceted elements such as color, shape, and texture, characterized by their complexity and subtlety. Current evaluation methods are predominantly subjective, lacking the ability to objectively and comprehensively capture and quantify these dynamic attributes. To address this limitation, this study proposes an innovative approach for evaluating artistic creativity by integrating an artificial immune algorithm (AIA). Initially, features related to color, shape, and texture are extracted from the artworks, and these feature vectors are input into the AIA as antigens. Subsequently, by defining antigen-antibody matching rules, the features of the artworks are compared with creative reference antibodies generated by the algorithm to derive creativity evaluation results. Finally, leveraging a dynamic adjustment mechanism for bidirectional crossover mutation probabilities, the diversity of the antibody population is optimized through a clonal selection strategy, enhancing the model’s adaptability to diverse artistic styles. Experimental results demonstrate that the proposed improved AIA achieves significant enhancements in evaluation metrics, with average Precision and Recall values increasing by 8.45% and 11.21%, respectively, compared to the baseline AIA. This study concludes that the AIA-integrated creativity evaluation method effectively improves the objectivity and accuracy of artistic assessments, offering a novel perspective for the intelligent evaluation of artworks.

Keywords

Creative evaluation artificial immune algorithm artistic creation clonal selection mechanism feedback and iteration

Introduction

Under the background of global integration and rapid development of science and technology, artistic creation and dissemination have ushered in unprecedented opportunities.^1,2 The works of art are no longer limited to traditional display platforms. The Internet and social media provide a larger stage for them. At present, the creative evaluation method of works mainly relies on the aesthetic judgment of professionals and the subjective feelings of the audience. It is difficult to eliminate the influence of factors such as individual preferences and cultural differences, making the evaluation results not objective and consistent.^3,4 Moreover, most traditional methods focus on qualitative research and are difficult to provide specific quantitative data. This not only restricts the fairness and scientificity of work evaluation, but also prevents artists from getting effective creative guidance. In recent years, the combination of artificial intelligence (AI) and bionic computing has enabled AIA (artificial immune algorithm) to make great progress in real-world applications.^5,6 It simulates the immune mechanism of organisms and has functions such as memory, diversity maintenance, and adaptive regulation. An evaluation model based on AIA is constructed, and the memory database of AIA is used to characterize the creative features of the works. The antibody-antigen binding method is used to measure the creativity of the works, which is of great significance to promoting the intelligent development of art education and innovative practice.

Intelligent technology is increasingly being used in the analysis of artworks.⁷ Existing research on the creative evaluation of works of art focuses on the extraction and analysis of creative features. To objectively evaluate the aesthetic value of creative products, Liang Dan used the BP (back propagation) network to establish an aesthetic value evaluation model. A segmentation adaptive strategy was used to improve the field of view and step size of the algorithm. The simulation results showed that as part of the adaptive improvement method, the accuracy of the algorithm’s optimal solution was improved, which can be said to provide objective guidance for the evaluation of the aesthetic value of the product.⁸ To enhance the recognition of works of art, Chen Ruhua introduced an AI-based art recognition method by combining advanced feature extraction technology with a customized CNN architecture and using key dataset features to verify that the proposed method performs well in cross-attribute classification and evaluation of works of art.⁹ Gangadharbatla Harsha explored the role of AI attribution knowledge in the evaluation and reception of artworks. The results showed that AI attribution knowledge can connect figurative art with humans and abstract art with machines to achieve objective evaluation of artworks.¹⁰ Cetinic Eva used AI for art analysis and explored the application of object detection and similarity retrieval in the evaluation of digital artworks. By discussing the practical and theoretical aspects of AI in art evaluation, she demonstrated the effectiveness of AI in helping people understand artistic creation.¹¹ Pitt Christine S demonstrated a unique hybrid method to analyze unstructured qualitative data on the characteristics of artworks. She first collected semi-structured interviews with art collectors and then conducted quantitative analysis through automatic text analysis. She used a hybrid method to explore these characteristics to develop a typology of art consumers.¹² Current research helps to clarify the evaluation criteria for the innovation of artworks and enrich the analytical dimensions of artworks. However, most of them are based on static data and cannot meet the dynamic optimization needs in creative evaluation, resulting in a large gap between the intelligence of evaluation and optimization guidance.

The AIA algorithm can effectively solve the problems of high complexity and lack of dynamic optimization guidance in existing methods by simulating the dynamic learning and adaptability of the biological immune system.^13–15 To overcome the shortcomings of large computational complexity, low convergence accuracy, and slow convergence speed in the function optimization process, Meng Yafeng applied multiple adaptive immune operators and proposed an improved adaptive AIA. In the classic AIA, the number of iterations was applied to adaptively design the antibody incentive calculation operator, and the average incentive of the population antibody and the antibody incentive were applied to adaptively design the operator. Finally, nine typical test functions were selected as experimental objects. The results showed that the improved adaptive AIA was effective and superior in solving function optimization problems.¹⁶ To improve the dynamic optimization effect, Qi Yutao constructed a cooperative co-evolution model based on a new subtask division method, applied the model into the AIA system, and proposed a cooperative immune multi-objective optimization algorithm based on the cooperative model. The experimental results showed that the algorithm performed well in both solution quality and convergence speed. Especially for multi-objective optimization problems with nonlinear correlation of decision variables, the performance was particularly outstanding.¹⁷ The dynamic adaptive ability of AIA enables it to effectively deal with the complex data problems in creativity evaluation, and provides a systematic solution for the creativity evaluation of works. However, how to further improve the objectivity and efficiency of dynamic capture and quantification of work features is still an urgent problem to be solved.

To enhance the accuracy and efficiency of artistic creativity evaluation and better capture the nuanced creative characteristics of artworks, this study integrates the artificial immune algorithm (AIA) to establish a novel evaluation model. The primary objective of this research is to develop a robust, objective, and adaptive framework for assessing artistic creativity, addressing the limitations of traditional subjective methods. By utilizing feature vectors extracted from artworks as antigens, the proposed model employs antigen-antibody matching and clonal selection mechanisms to evaluate creativity and provide actionable feedback for improvement. Experimental results demonstrate significant advancements over the baseline AIA: the improved AIA achieves an 8.45% increase in average Precision and an 11.21% increase in average Recall. Furthermore, the enhanced model exhibits superior diversity, with an average diversity value of 1.657, representing a 34.83% improvement over the baseline. In terms of computational efficiency, the improved AIA reduces the average processing time to 1.323 seconds across various artistic styles, marking a 14.81% reduction compared to the basic AIA. These findings underscore the effectiveness of the AIA-integrated creativity evaluation model in improving both the accuracy and efficiency of artistic assessments, while also offering valuable insights for refining creative processes.

Creativity evaluation and improvement method of works of art based on AIA

Data collection and preprocessing

This paper achieves an objective summary of the creativity of works by comprehensively extracting and accurately quantifying the multidimensional visual features of the works. The data collection and preprocessing of this paper are divided into three stages: image acquisition, feature extraction, and construction and normalization of feature vectors.

Image acquisition

This paper uses the public WikiArt dataset as the main data source. WikiArt covers more than 80000 works of art, including more than 1000 artists and 27 art styles. It has a large time span and includes works of various styles from the Renaissance to modern art. Using it as the main dataset can ensure the authority and diversity of the data source and provide a reliable data basis for creativity evaluation.

Feature extraction

Due to the different resolutions of the original images, there are certain differences in color modes, and the background interference has a greater impact on the data quality, so the original images are normalized first. All images are adjusted to a uniform 1024×1024 pixel size, and then converted into RGB (Red, Green, Blue) color space. In terms of irrelevant background removal, a model based on the U-Net framework is used to achieve efficient separation of the main content of the image through its encoder-decoder structure. Assuming that the probability map output by U-Net is $P (Ξ)$ , $Ξ$ represents each pixel in the input image¹⁸:

P (Ξ) = σ (W^{T} f (Ξ) + b)

(1)

W

represents the weighting matrix;

f (Ξ)

represents the feature mapping function;

σ

represents the sigmoid activation function;

b

represents the bias term. Then, the image is thresholded to obtain a binary foreground area template.

The specific feature extraction is shown in Figure 1.

Figure 1.

Feature extraction of works.

The color distribution is described using the HSV (Hue, Saturation, Value) color space method. First, the histogram distribution of hue $H$ , saturation $S$ , and brightness $V$ is calculated by transforming RGB to HSV; assuming that the total number of pixels is $N$ , the HSV value of each pixel is $(H_{i}, S_{i}, V_{i})$ , and the expression of the color histogram is^19,20:

P_{k} = \frac{\sum_{i = 1}^{N} δ (H_{i} \in k)}{N}, k = 1, 2, \dots, φ

(2)

δ

is the indicator function. If

H_{i}

is in the

k

-th color interval,

(H_{i} \in k) = 1

, otherwise it is 0. The distribution values of each color interval are merged to obtain the color feature vector

C = [P_{1}, P_{2}, \dots, P_{M}]

The shape feature is extracted by combining edge detection and shape moment analysis. First, the edge extraction method of Canny image^21,22 is used to obtain the boundary distribution of the image, and the Hu moment is used as the overall shape descriptor. Hu moment is a seven-invariant feature based on the grayscale moment of the image, and its formulas are:

η_{u v} = \frac{m_{u v}}{m_{00}^{γ}}

(3)

γ = \frac{u + v}{2} + 1

(4)

Among them, $m_{u v}$ is the grayscale moment of the image, and its expression is:

m_{u v} = \sum_{x} \sum_{y} x^{u} y^{v} I (x, y)

(5)

Here,

I (x, y)

refers to the pixel value. After nonlinear transformation, the Hu moment obtains seven invariants. Finally, the shape feature vector is obtained:

S = [h_{1}, h_{2}, \dots, h_{7}]

The texture feature is extracted by Gray-Level Co-Occurrence Matrix (GLCM). On a grayscale image, the distance of a certain pixel and the grayscale co-occurrence probability matrix $p (i, j, d, β)$ in a certain direction are calculated. Among them $i, j$ are grayscale values; $d$ is the distance between pixels; $β$ is the direction. The texture feature extraction based on GLCM is expressed as^23,24:

(1) Contrast:

C o n t r a s t = \sum_{i = 0}^{L - 1} \sum_{j = 0}^{L - 1} {(i - j)}^{2} p (i, j)

(6)

(2) Energy:

E n e r g y = \sum_{i = 0}^{L - 1} \sum_{j = 0}^{L - 1} p {(i, j)}^{2}

(7)

(3) Homogeneity:

H o m o g e n e i t y = \sum_{i = 0}^{L - 1} \sum_{j = 0}^{L - 1} \frac{p (i, j)}{1 + | i - j |}

(8)

The final texture feature vector is: $T = [C o n t r a s t, E n e r g y, H o m o g e n e i t y]$ .

Feature vector construction and standardization

Finally, by integrating the $C, S, T$ feature vectors, the comprehensive feature vector $X$ of the work is obtained. To eliminate the scale differences between the feature vectors, the Z-score method is used to normalize them, and its expression is:

X_{i}^{'} = \frac{X_{i} - μ (X)}{σ (X)}, i = 1, 2, \dots, n

(9)

Among them,

μ (X)

represents the average value of the feature value

X_{i}

;

σ (X)

represents the standard deviation; and

n

represents the total number of samples. Finally, the normalized feature vector is used as the input of the model.

Evaluation mechanism based on AIA

The theoretical mechanism of AIA is to imitate the body’s recognition, memory, and response to exogenous substances to adapt to changes in the external environment.^25,26 When evaluating the creativity of a work, its visual features are mainly used to measure its originality. These features are converted into feature vectors in mathematical form. The feature vector represents the “antigen” characteristics of the work, and the preset creativity evaluation criteria correspond to the “antibody,” as shown in Figure 2.

Figure 2.

Evaluation mechanism based on AIA.

In Figure 2, the evaluation process is to find the antibody that can best reflect the creativity of the work, that is, to find an antibody that best matches the antigen. This process depends not only on the quality of the antibody library, but also on the dynamic changes of the antibody library. Therefore, the antibody library needs to be dynamically adjusted to match the ever-changing artistic trends.

Definition of antigens and antibodies

First, the features extracted from the works of art are defined as antigens. In the data collection and preprocessing stage, image processing is used to extract the multidimensional features of color, shape, and texture, and these features are fused into a high-dimensional feature vector, which is represented as $X = [x_{1}, x_{2}, {\dots, x}_{n}]$ . Here, $n$ is the feature dimension. The feature vector of each work is an independent antigen.

Antibodies are defined by the evaluation criteria generated by AIA. On this basis, a method based on expert evaluation and the principle of statistics are used to generate a series of representative antibodies. Assuming that $A = [a_{1}, a_{2}, \dots, a_{n}]$ , $a_{j}$ is the evaluation reference value for the $j$ -th dimension feature quantity. $A_{1}, A_{2}, \dots, A_{m}$ represent the antibody population, where m represents the number of antibodies.

The generation of antibodies is divided into three main steps: initialization, training, and verification:

(1) Initialization: a set of initial antibodies is randomly generated.

(2) On this basis, supervised learning of antibody parameters is performed through labeled training to maximize the applicability to the actual evaluation criteria. This step can be completed by minimizing the loss function $L (a_{j}, x_{i})$ ^27,28:

L (a_{j}, x_{i}) = \frac{1}{2} {‖ a_{j} - x_{i} ‖}_{2}^{2}

(10)

(3) Verification: the antibodies are cross-validated, and the antibody library is modified based on feedback to ensure its universality.

Antigen-antibody matching rule

After defining antigens and antibodies, determining the matching relationship between antigens (features of works) and antibodies (evaluation criteria) is an important step in the creative evaluation process. The similarity $S i m (x, a)$ between antigens and antibodies is defined as the matching degree, and $S i m (x, a)$ is obtained by reverse operation of the Euclidean distance:

S i m (x, a) = \exp (- \frac{{‖ x - a ‖}^{2}}{2 λ^{2}})

(11)

λ

is the regularization parameter that controls the decay rate of the matching degree. A higher matching degree indicates that the features of the work are closer to the evaluation criteria:

{‖ x - a ‖}^{2} = \sum_{i = 1}^{n} {(x_{i} - a_{j})}^{2}

(12)

For each work, the similarity score between it and all antibodies is calculated, and the group with the highest score is the final evaluation result.

Feedback and iterative improvement

The core of the feedback and iteration mechanism lies in its ability to refine the model’s creativity through continuous evaluation feedback and corrective adjustments. Simultaneously, the AIA-based evaluation model is dynamically adapted through iterative processes to accommodate diverse art forms and the integration of new artworks. This mechanism comprises two key components: feedback loops and self-updating mechanisms. The feedback loop leverages quantitative evaluation results to optimize the creative quality of artworks, establishing a cyclical process that enhances creativity based on ongoing assessments. The self-updating mechanism focuses on improving the model’s adaptability, enabling it to assimilate new information from incoming data and evolve over time. Together, these components ensure that the model remains responsive to changing artistic contexts and continuously improves its evaluation accuracy and relevance.

Feedback loop

The construction of the feedback loop is based on the evaluation results output by the model, and provides specific feedback for the creation of the work through the creative evaluation and matching results generated by the model, as shown in Figure 3.

Figure 3.

Feedback loop mechanism.

Based on the output of the AIA evaluation model, the matching degree $S i m (x, a)$ between the work features and the reference antibody is extracted. According to the matching results, the specific results of the work in color, shape, and texture dimensions are analyzed. If the matching degree with a certain feature is not high, the corresponding adjustment strategy is proposed in the feedback result, such as improving color contrast or enhancing the complexity of the structure.

Iterative improvement

Fixed evaluation criteria are difficult to adapt to different art styles and creative characteristics, resulting in a lack of universality and fairness in the evaluation results. When dealing with high-dimensional feature spaces, traditional methods are susceptible to data bias and fall into local extreme values. They cannot efficiently explore the multidimensional creative space, resulting in a single evaluation standard and cannot objectively reflect the creativity of the work.

To ensure the continuous evolution of the model, an iterative improvement mechanism is constructed. As shown in Figure 4, its core is to dynamically adjust the antibody library so that it can reflect the current artistic trends and the development of personal style. Whenever a new work is added, it is evaluated, and the results are fed back to the model. When the existing antibody library cannot effectively identify some novel creative elements, the clonal selection program is started to generate new antibodies to fill this gap.

Figure 4.

Iterative improvement mechanism.

During the algorithm iteration process, since the crossover probability and mutation probability of the operator are generally fixed, the optimization effect and accuracy are affected during the iteration process. On this basis, the bidirectional crossover and mutation probability is used to dynamically control the crossover and mutation probabilities, and its expressions are:

P_{c}^{n e w} = \min (P_{c}^{\max}, P_{c}^{b a s e} + δ_{c} \cdot \frac{{F i t}_{a v g}}{{F i t}_{b e s t}})

(13)

P_{m}^{n e w} = \max (P_{m}^{\min}, P_{m}^{b a s e} + δ_{m} \cdot \frac{{F i t}_{a v g}}{{F i t}_{b e s t}})

(14)

The formula variables are shown in Table 1.

Table 1.

Formula variables.

Sequence	Variables	Definition
1	$P_{c}^{n e w}$	Dynamically adjusted crossover probability
2	$P_{m}^{n e w}$	Dynamically adjusted mutation probability
3	$P_{c}^{b a s e}$	Initial crossover probability
4	$P_{m}^{b a s e}$	Initial mutation probability
5	$δ_{c}$	Adjustment factor
6	$δ_{m}$	Adjustment factor
7	${F i t}_{a v g}$	Average fitness
8	${F i t}_{b e s t}$	Optimal individual fitness
9	$P_{c}^{\max}$	Upper limit of crossover probability
10	$P_{m}^{\min}$	The lower limit of mutation probability

By increasing the probability of crossover between populations in the early stage, the detection ability of the population is improved. In the late stage of population evolution, the probability of mutation gradually decreases, thereby achieving effective search of the population.^29,30

The antibody population update is performed by the dynamic adjustment sum of $P_{c}^{n e w}$ and $P_{m}^{n e w}$ in each iteration.

In the crossover operation, the two antibodies $A_{p}$ and $A_{q}$ cross each other to produce a new antibody $A_{n e w}$ :

A_{n e w} = α \cdot A_{p} + (1 - α) \cdot A_{q}

(15)

α

is the crossover coefficient, which takes a value in the interval

[0, 1]

The formula for the variation of the feature vector of each antibody is³¹:

A_{i}^{n e w} [k] = A_{i} [k] + ς \cdot r a n d (- 1, 1)

(16)

Among them, $ς$ is a randomly generated interference value, representing the variation amplitude.

The clonal selection theory believes that after antibodies specifically bind to antigens, they undergo large-scale self-replication and mutation, thereby forming new variants.

(1) Clone generation

In terms of dynamic optimization, antibodies that match the antigen to a higher degree than the threshold are first selected, and $k$ sub-antibodies are obtained through cloning:

A_{k}^{(j)} = A^{(j)} + ϵ, ϵ \sim γ (0, δ^{2})

(17)

ϵ

is the Gaussian noise used to apply diversity, and

δ

is the intensity of the noise.

(2) Clonal selection

A part of this dimension is randomly used to fine-tune the feature vector of each sub-antibody:

a_{j} = a_{j} + η, η \sim μ (- ρ, ρ)

(18)

here,

η

represents the uniform noise distribution, and

ρ

represents the control of variation amplitude.

(3) Adaptability evaluation

The matching degree between each sub-antibody is calculated, and the individuals with relatively large values are taken as the composition of the antibody library to replace the components with small matching degree. The negative selection mechanism is used to eliminate antibodies that cannot effectively recognize any antigen. The negative selection condition is³²:

\forall a \in A, \exists x \in X, S i m (x, a) > θ

(19)

When the work is updated, new data is collected, and its feature vector is extracted. The memory library of the model is updated:

{M e m}_{n e w} = ξ \cdot {M e m}_{o l d} + (1 - ξ) \cdot {D a t a}_{n e w}

(20)

{M e m}_{o l d}

is the old memory library;

{D a t a}_{n e w}

is the feature data of the new work; and

ξ

is the update weight of the memory library, which is in the range of

0 < ξ < 1

On this basis, new data is applied to adjust the parameters of the model, and the matching error is minimized by the gradient descent method:

\min_{F} L (F) = \frac{1}{m} \sum_{i = 1}^{m} {({S i m}_{i, p r e d} - {S i m}_{i, t r u e})}^{2}

(21)

In this process, the evaluation criteria of the model are repeatedly iterated to achieve the optimal value.

Experiment on the evaluation and improvement of creativity of works of art based on AIA

Experimental setting

Through experimental analysis, the performance of the AIA evaluation model is analyzed to verify its adaptability and effectiveness in evaluating the creativity of different types of works.

Data preparation

This paper selects the public WikiArt as the experimental dataset. To ensure the diversity and representativeness of the experiment, eight different types of works of art are randomly selected. Its basic composition is shown in Table 2.

Table 2.

Basic composition of samples.

Number	Artistic style	Number of samples
S-1	Abstract-expressionism	336
S-2	Neoclassicism	294
S-3	Realism	301
S-4	Impressionism	299
S-5	Romanticism	472
S-6	Symbolism	259
S-7	Surrealism	205
S-8	Baroque	192

In Table 2, the data is preprocessed, and the color features (24 dimensions), shape features (20 dimensions), and texture features (16 dimensions) of the work images are extracted to obtain a 60-dimensional feature vector.

Experimental grouping

The experiment is divided into a basic AIA group, an improved AIA group, and an expert evaluation group:

(1) Basic AIA group: the basic AIA is a traditional AIA algorithm without any improvement, which is used as a control group to compare with the model in this paper.

(2) Improved AIA group: the improved AIA applies a feedback and iteration mechanism.

(3) Expert evaluation group: five professional painters and critics are invited to evaluate and score the creativity of the same work, and this is used as a reference for the real results.

The AIA model parameters are set as shown in Table 3.

Table 3.

AIA model parameters.

Sequence	Parameter	Range
1	Initial population size	100–300
2	Maximum number of iterations	500–1000
3	Cloning factor	0.1–0.5
4	Mutation probability	0.01–0.2
5	Cross probability	0.4–0.8
6	Memory bank size	10–50
7	Diversity threshold	100–200
8	Stop condition	0.3–0.5
9	Weight factors of fitness function	The weights of color, shape, and texture features are 0.3, 0.4, and 0.3, respectively

The feature vector of each style of works of art is used as the antigen input, and the evaluation criteria are set. The model uses the antigen-antibody pairing and clone screening principles to achieve creativity evaluation. Finally, the evaluation results of each model in each work are statistically analyzed and compared with the expert scoring results.

Experimental results

Precision

The Precision precisely measures the ratio of the model to the actual consistency of the expert evaluation results in the task of evaluating highly creative works. Its calculation is shown in Formula 22:

P r e c i s i o n = \frac{T P}{T P + F P}

(22)

Among them, $T P$ represents the number of highly creative works correctly identified by the model, and $F P$ represents the number of low-creativity works incorrectly identified.

The score of the expert evaluation group is used as the benchmark, and the works with the top 20% of the expert scores are defined as highly creative works. The basic AIA group and the improved AIA group, respectively, evaluate the creativity of the same feature vector set and compare it with the results of the expert evaluation group. The final Precision results are shown in Figure 5.

Figure 5.

Precision results.

As can be seen in Figure 5, the Precision results under the improved AIA are generally higher than those under the basic AIA. Among them, the precision of basic AIA in various styles of works reaches up to 0.648, and the average precision of matching is about 0.580; the precision of improved AIA in various styles of works reaches up to 0.676, and the average precision of matching is about 0.629. In the comparison of the average precision results, the improved AIA group improves by 8.45% compared with the basic AIA group. In terms of specific styles, the artworks in the impressionist and symbolist styles perform better in the evaluation of creativity, mainly because their color, texture, and other characteristics have higher complexity and consistency. Compared with the basic AIA, the improved AIA can better reflect the creative characteristics of the works. In the antibody group, by applying a dynamic adjustment mechanism of the probability of bidirectional crossover mutation, the model has a stronger local optimization ability, and on this basis, its adaptive ability in the evaluation task is further improved.

Recall

Recall is used to measure the ratio of highly creative works that are correctly identified by the model. Its calculation is shown in Formula 23:

R e c a l l = \frac{T P}{T P + F N}

(23)

Among them, $F N$ refers to the number of highly creative works that are not identified by the model.

Consistent with the experimental settings of Precision analysis, the Recall analysis still uses the top 20% of highly creative works in the expert evaluation group as the benchmark and analyzes the final evaluation results of each model. The comparison is shown in Figure 6.

Figure 6.

Recall results.

In Figure 6, the improved AIA has a significant advantage in the matching Recall results in the evaluation of the creativity of works of various styles. In the specific comparison results, the maximum matching Recall value of the basic AIA is 0.553, and its average result is 0.535; the maximum matching Recall value of the improved AIA in the evaluation of works of various styles is 0.640, and its average result is 0.595, which is 11.21% higher than the average Recall of the basic AIA. The matching Recall results show that the algorithm model in this paper has a stronger ability to recognize highly creative works, can more precisely capture different types of artistic creativity, and provide more reliable evaluation and guidance for its evaluation and creation guidance. At the same time, the diversity and adaptability of the antibody library are improved through the replication and screening mechanism, so that the model can better identify different types of creative elements, and realize the long-term evaluation of creativity through personalized feedback and iterative improvement.

Diversity of antibody population

The diversity of antibody population is used to evaluate the distribution breadth of antibody population, which is defined as the average value of Euclidean distance between antibodies:

D i v e r s i t y = \frac{1}{N (N - 1)} \sum_{i = 1}^{N} \sum_{j = i + 1}^{N} ‖ A_{i} - A_{j} ‖

(24)

By calculating the diversity of antibody population, the diversity differences between antibodies generated by the algorithm are compared. To ensure the stability of the results, each group of experiments is repeated 10 times, and then the average value is taken. The final comparison results are shown in Figure 7.

Figure 7.

Diversity results.

In Figure 7, the maximum Diversity value of basic AIA is 1.301, and its average Diversity result is 1.229; the maximum Diversity value of improved AIA in the evaluation of works of various styles is 1.754, and its average result is 1.657, which is 34.83% higher than the average Diversity of basic AIA. The improved AIA obtains new antigen variants through large-scale repetition and random mutation of antigen-specific antibodies, thereby greatly enriching the diversity of the antibody library, making the model better suitable for the evaluation of works of various styles and objectively reflecting its creativity. With the addition of new works, the model can improve the diversity and adaptability of the evaluation of the creativity of works through the constantly updated antibody library.

Operation efficiency

This paper randomly selects the same number of works from various styles as evaluation samples, and calculates the operation efficiency of different models. The time taken from the input of the feature vector to the output of the final evaluation result is recorded. The results are shown in Figure 8.

Figure 8.

Operation efficiency results.

In Figure 8, a smaller operation time means that the model evaluation is more efficient. In the specific comparison, the average operation time of the basic AIA under various styles of works is 1.553 s; the average operation time of the improved AIA under various styles of works is 1.323 s. Compared with the basic AIA, the average operation time of the algorithm model in the evaluation of the creativity of works of art is reduced by 14.81%. Overall, the improved AIA has higher operation efficiency. By optimizing the algorithm structure, reducing unnecessary calculation processes, and accelerating the antigen-antibody comparison, the rapid evaluation of works can be realized faster, and timely feedback can be provided for the evaluation of the creativity of works, which improves the real-time performance of the algorithm.

Discussion

This study establishes a creativity evaluation model for artworks by integrating the artificial immune algorithm (AIA), enhancing its adaptive capabilities and computational efficiency through a robust feedback and iteration mechanism. Experimental results demonstrate significant improvements over the baseline AIA: the enhanced model achieves an 8.45% increase in Precision, an 11.21% increase in Recall, and a 34.83% improvement in antibody population Diversity, while reducing the average operation time by 14.81%. These findings validate the model’s effectiveness in identifying and evaluating highly creative artworks, as well as its ability to accurately reflect the unique creative characteristics of diverse works. By dynamically optimizing the AIA, this research addresses the issue of excessive randomness in traditional AIA operations and provides an efficient and flexible solution to complex creativity evaluation challenges. Through the integration of quantitative artistic feature analysis and AIA, the model enables objective feedback and scientific assessment of creativity in artworks. Furthermore, its adaptive update mechanism ensures dynamic responsiveness to evolving artistic styles, demonstrating its potential as a versatile tool for creativity evaluation in the arts.

Conclusions

This study integrates the artificial immune algorithm (AIA) to investigate the evaluation and enhancement of artistic creativity. By incorporating a feedback and iteration mechanism, an evaluation model is developed to address the limitations of traditional AIA operators, improving both the accuracy and real-time performance of creativity assessments. Experimental results demonstrate that the proposed algorithm outperforms traditional AIA in key metrics, including matching precision, population diversity, and computational efficiency, enabling objective and effective identification and evaluation of artistic creativity. The model’s adaptive feedback mechanism allows for continuous updates and optimization in response to new artworks and evolving artistic trends, providing targeted feedback to refine creative processes. However, certain limitations remain. Although the WikiArt dataset is extensive, the model’s robustness requires further validation due to the limited sample size and variability in data types. Future research could focus on enhancing the accuracy and intelligence of artistic evaluation by exploring advanced algorithms and investigating the integration of AIA with comprehensive optimization techniques for other art forms. Such advancements could drive the intelligent and personalized development of artistic creation, fostering innovation in the field.

Footnotes

ORCID iD

Xiuming Li

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

References

Wiratno

Brigitta

. Transformation of beauty in digital fine arts aesthetics: an artpreneur perspective. Aptisi Trans Technopreneurship (ATT) 2024; 6(2): 231–241.

Kharchenko

Chibalashvili

Savchuk

, et al. Technologies as a mediator between creator and audience in postmodern art practices. BRAIN: Broad Res Artif Intell Neurosci 2023; 14(1): 500–514.

Christou

Violet

Pinar

. Concepts and contexts of creative evaluation approaches. J Multidiscip Eval 2021; 17(40): 1–15.

Lewandowska

Michael

Emanuel

. Research quality criteria in the creative arts. Stud Higher Educ 2024; 49(4): 639–653.

Zhang

Tang

You

, et al. Network function heterogeneous redundancy deployment method based on immune algorithm. Chin J Netw Inform Secur 2021; 7(1): 46–56.

Cheng

Wei

Zhang

. Reliability allocation method based on immune genetic optimization. Ship Electron Eng 2022; 42(2): 152–156.

Yusa

IMM

Sovhyra

. Reflections on the use of artificial intelligence in works of art. J Aesthet Des Art Manag 2022; 2(2): 152–167.

Liang

. Aesthetic value evaluation for digital cultural and creative products with artificial intelligence. Wirel Commun Mobile Comput 2022; 1: 1–10.

Chen

Aghdam

MRG

Khishe

. Utilization of Artificial Intelligence for the automated recognition of fine arts. PLoS One 2024; 19(11): 1–16.

10.

Gangadharbatla

. The role of AI attribution knowledge in the evaluation of artwork. Empir Stud Arts 2022; 40(2): 125–142.

11.

Cetinic

She

. Understanding and creating art with AI: review and outlook. ACM Trans Multimed Comput Commun Appl 2022; 18(2): 1–22.

12.

Pitt

Bal

Kirk

. New approaches to psychographic consumer segmentation: exploring fine art collectors using artificial intelligence, automated text analysis and correspondence analysis. Eur J Market 2020; 54(2): 305–326.

13.

Zhang

, et al. Global digital compact: a mechanism for the governance of online discriminatory and misleading content generation. Int J Hum Comput Interact 2025; 41: 1381–1396. DOI: 10.1080/10447318.2024.2314350.

14.

Ren

Cui

Liu

. A scale-adaptive immune classification algorithm based on artificial immune system. Industr Control Comput 2020; 33(9): 40–44.

15.

Shi

Chen

Ding

, et al. Artificial immune dynamic multi-objective optimization algorithm based on offline search. Control Eng 2023; 30(8): 1467–1476.

16.

Yafeng

Wang

Zexi

, et al. Improved adaptive artificial immune algorithm for solving function optimization problems. J Beijing Univ Aeronaut Astronaut 2021; 47(5): 894–903.

17.

Liu

Ren

, et al. Collaborative immune multi-objective optimization algorithm based on cooperative model. Acta Electron Sin 2023; 42(5): 858–867.

18.

Jin

Zhou

, et al. An unsupervised multi-focus image fusion method based on Transformer and U-Net. IET Image Process 2023; 17(3): 733–746.

19.

Hassan

Gutub

. Improving data hiding within colour images using hue component of HSV colour space. CAAI Trans Intell Technol 2022; 7(1): 56–68.

20.

Othman

Abdulla

. Enhanced single image dehazing technique based on hsv color space. UHD J Sci Technol 2022; 6(2): 135–146.

21.

Sekehravani

Babulak

Masoodi

. Implementing canny edge detection algorithm for noisy image. Bulletin EEI 2020; 9(4): 1404–1410.

22.

Cao

Duan

. A new image edge detection algorithm based on improved Canny. J Comput Methods Sci Eng 2020; 20(2): 629–642.

23.

Sadique

Haque

. Content-based image retrieval using color layout descriptor, gray-level co-occurrence matrix and k-nearest neighbors. Int J Inf Technol Comput Sci 2020; 12(3): 19–25.

24.

Aouat

Ait-hammi

Hamouchene

. A new approach for texture segmentation based on the Gray Level Co-occurrence Matrix. Multimed Tool Appl 2021; 80(16): 24027–24052.

25.

Rezvanian

Vahidipour

Saghiri

. CaAIS: cellular automata-based artificial immune system for dynamic environments. Algorithms 2023; 17(1): 18–37.

26.

Wang

Zhang

. Artificial immune algorithm for constrained multi-objective interval optimization problems. J Guizhou Univ (Nat Sci Ed) 2022; 39(5): 94–99.

27.

Qiao

Yang

, et al. An adaptive hybrid evolutionary immune multi-objective algorithm based on uniform distribution selection. Inf Sci 2020; 512(3): 446–470.

28.

Lin

Zhong

. A survey of artificial immune algorithms for multi-objective optimization. Neurocomputing 2022; 489(7): 211–229.

29.

Zhu

. A hybrid artificial immune optimization for high-dimensional feature selection. Knowl Base Syst 2023; 260(25): 110111.

30.

Wang

Cui

. A novel adaptive genetic algorithm based on population diversity. J Terahertz Sci Electron Inf Technol 2023; 21(5): 671–676.

31.

Antipov

Buzdalov

Doerr

. Fast mutation in crossover-based algorithms. Algorithmica 2022; 84(6): 1724–1761.

32.

Sun

Wang

Liu

, et al. Self-updating continual learning classification method based on artificial immune system. Appl Intell 2022; 52(11): 1–27.