Exploration and Practice of Artificial Intelligence Generative Art in Environmental Public Art

When an image is subjected to line extraction, its first task is to find the set of pixel points with step change or roof change. Whether it is threshold segmentation or detecting edges based on high and low frequency components of gray-scale images, the filtering operation of maximization is usually performed first, and then the line extraction operation is completed by differential operators in the smooth region with clear line features according to the image intensities (grayscale, luminance). In this study, a color gradient algorithm is designed for line extraction, which first uses the normalized calculation of the plane gradient of the RGB three-channel to achieve the highlighting of the edge information in the smooth inner region, and then uses the maximum variance queue value to perform the queue segmentation to obtain a smooth and clear line contour map.

In this study, the steps for segmenting the image based on line extraction with color gradient are as follows.

Step 1: Input the color original image with R, G and B channels.

Step 2: A continuous clear line contour image is acquired with color gradient algorithm.

Step 3: Take the extracted contour image as a mask image, do a subtraction operation with the grayscale image of the input image, perform image segmentation, and obtain a background image.

Among them, the specific process of color gradient algorithm is as follows:

Input: color image img;

Output: contour image counterlmg;

Extract img of three channels imgR, imgG, imgB;

Create nonlinear filter f;

Filtering operations are performed in the x and y directions of imgR, imgG and imgB with f to obtain imgRX, imgRY, imgGX, imgGY, imgBX and imgBY, respectively;

Calculate the angle of radians according to equation (1) equation (4) A; 1 xy=imgRX*imgRY+imgGX*imgGY+imgBX*imgBY 2 xx=imgRX*imgRX+imgGX*imgGX+imgBX*imgBX 3 yy=imgRY*imgRY+imgGY*imgGY+imgBY*imgBY 4 A=0.5*(arctan(2*xy(xx−yy)))

The normalization matrix Q is obtained from Eq. (5) Eq. (7); 5 Q1=0.5*((xx+yy)+(xx−yy)*cos(2*A)+2*xy*sin(2*A)) 6 Q2=0.5*((xx+yy)+(xx−yy)*cos(2*(A+π/2))+2*xy*sin(2*(A+π/2))) 7 Q=max(Q1,Q2)

for s = 1 to 3 do

Calculate the gradient of each channel plane as Q

end for

Get the gradient transformation matrix X

Maximum variance thresholding based on the edge information to get the contour image counterImg

retrun counterImg

Given a target image, we use the methods of “image partitioning and flow field generation” and “eye-movement based visual attention modeling” to obtain the flow field and visual attention distribution map of the image, respectively, and we try to generate a set of chunks for each region. The region partitioning problem is a typical mathematization problem, and its goal is to place a set of chunks, equal or unequal in size, into the whole region without overlapping each other. For image collage applications, we consider a chunk as a square whose side lengths are determined by the region and the visual attention distribution map, which can be formally described as follows:

Given a region Ω ∈ ℝ² and n flow field lines SL={sli}i=1n, we seek to find an optimal configuration C for all chunks P={pj}j=1m within the region such that their chunks do not overlap each other and are distributed along the direction of the flow field lines. At the same time, we introduce a coverage τ ∈ ℝ, τ > 0 ensuring that all chunks can approximately cover the entire region. Thus, the region partitioning problem can be formalized as the following maximization problem: 10 Maximize∑j=1mArea(pj)≥τ*Area(Ω)Subjectto{pi⊆Ω,i∈{1,…,m}pi∩pj=∅,i,j∈{1,…,m},i≠j

A configuration of the set of chunks P={pj}j=1m can be denoted as C(L,D,S) = {(l₁,…,l_m),(d₁,…,d_m),(s₁,…,s_m)} where l_j is the center coordinate of chunk p_j, d_j is the orientation of chunk p_j, and s_j is the size of chunk p_j. We say that the configuration is legal if all chunks satisfy the above two constraints. Since a chunk is viewed as a closed square, the first constraint can be computed from the size of the chunks; the inclusion and intersection relations in the second constraint can be obtained from the relations between the pixel points it covers.

As mentioned earlier, in order to keep the chunks distributed along the direction of the flow field lines, we need to search each flow field line sequentially along the direction of the flow field lines. In addition, we adopt a top-down approach to maximize the chunk size while minimizing the number of chunks to increase the search efficiency. That is, we first consider covering the region with the largest chunks, and if we find a location where the largest chunks can be placed, we halve the chunk size and search again, and so on, until the two constraints are satisfied.

Although the above method can obtain a set of chunks that satisfy the objective constraints, however it does not satisfy that the larger the chunks should be placed in more prominent locations, for this reason we include a visual attention distribution map in the chunk search process. A luminance threshold ξ(s) is used to determine whether a chunk of size s can be placed at the current location. The luminance threshold can be defined as: 11 ξ(s)=lightnessMAX−λ*log2sizeMAXs where lightness_MAX represents the maximum brightness value of the region, size_MAX is the maximum size of the chunk, s is the size of the current chunk, and λ is used to control the effect of the chunk size on the brightness threshold.

Once a chunk is found, we first calculate the average luminance value of the chunk. If the average luminance value is higher than the luminance threshold at its corresponding size, it means that the chunk size satisfies the current level of visual attention, so we add the chunk to the set of target chunks P, and also update the centroid coordinates, size, and orientation of the chunk to the configuration C. Otherwise, we discard the chunk and proceed to find the next chunk.

However, the set of chunks obtained by using only the above steps tends to be visually quite confusing because neighboring chunks of the same size from different flow field lines may have intersecting extension lines along the direction of the flow field lines.

To solve this problem, we need to selectively search the flow field lines. In other words, after traversing a flow field line, we need to selectively decide which flow field line to traverse next, instead of selecting the next flow field line sequentially. The selection criterion is whether there is a new chunk in the previously traversed flow field line. If there is a new chunk, the flow field line is divided into two subsets, one of which falls within the region covered by the extension line of the new chunk along the direction of the flow field line, and the other subset falls outside of the region. We then continue the search for the first subset by half the size of the added chunk, and the second subset by the size of the added chunk. If no new chunks were added, the search continued for other flow field lines according to the size of the previous chunks. We found that these steps are quite similar, except that the chunk size and the set of flow field lines are different for each search, so we further introduced a partitioning algorithm to solve the region partitioning problem.

Although the set of area chunks obtained by the partitioning algorithm and the visual attention distribution map P can reflect the visual attention level of the viewer well, due to the complexity of the flow field and the different luminance values of the visual attention distribution maps obtained from different target pictures, the luminance thresholds corresponding to different chunk sizes are difficult to be determined in advance, and the overall effect of the set of chunks obtained by empirically setting the luminance threshold may appear to be scattered. The overall effect of the set of chunks obtained by empirically setting the luminance thresholds may appear to be scattered. Therefore, by dynamically adjusting the size of λ in the luminance threshold, we calculate the set of chunks obtained under different luminance thresholds, construct an energy equation through “Gestalt rule modeling”, as shown in Eq. (12), calculate the degree of Gestalt visual perception of each set of chunks, and then select the best chunk set as the area chunks by minimizing the energy equation. The best chunk collection is selected as the result of regional chunking by minimizing the energy equation.

where ω_similar, ω_closure, are the weights of the similarity rule, closure rule, and regularity ω_regular rule, respectively, and generally ω_similar + ω_closure + ω_regular. E_i is the ith clip, and N is the number of clips.

With a collection of chunks, our goal is to find a suitable set of images for that collection. In general, choosing images with similar themes to the target images is more appealing to the viewer’s eye and reflects the theme of the collage better than choosing irrelevant images to generate the image collage. For this reason, we consider the high-level semantic features and the underlying color features of the images to select suitable images for image collage synthesis. To achieve this goal, this paper adopts a new semantic-color similarity measure to find the most suitable picture for each chunk by combining two different features.

In order to construct a mapping relationship between the set of chunks and the set of pictures, so that the larger the chunks are, the more the semantics of the pictures corresponding to the chunks conform to the theme of the target picture, so that the user can understand the theme expressed by the collage faster; the smaller the chunks are, the more the color of the pictures corresponding to the chunks conform to the color of the region where the chunks are located, so as to facilitate the preservation of the visual characteristics of the target picture. We introduce the semantic item weight ω_s and the color item weight ω_c, and the calculation of their weights is as follows: 13 ωs(s)=λ*ssizeMAX 14 ωc(s)=1−ωs(s) where λ is an adjustable parameter in the range of [0,1] to control how much the chunk size affects the semantic items. s is the chunk size, and size_MAX is the maximum size of the chunk.

Further, we define a new semantic-color similarity measure as follows: 15 Ematch(Pi,Ij)=ωs*DEJ(W(Pi),M(Ij))+ωc*DINT(H(Pi),H(Ij)

where P_i and I_j are the ird chunk and the jth picture respectively. W(P_i) is the semantic vector of the ith chunk, M(I_j) is the semantic vector of the jth picture. H(P_i) is the HSV histogram of the ith chunk, M(I_j) is the HSV histogram of the jth picture. The larger the value of E_match, the higher the semantic-color similarity between the two.

On the premise of ensuring the neighborhood diversity of each chunk in the set of chunks, we find for each chunk the image with the highest semantic-color similarity to be placed in the region where the chunk is located, thus generating the entire image collage.

The goal of the experiment was to test the system’s ability to generate paintings of art outside the realm of human experience, i.e., human subjects thought the painting was created by a top human artist but were unable to distinguish between the stylistic types of the artwork. In this experiment, each subject was asked after a single 5-second observation whether the creator of the image was human or computer, and how much he or she preferred the image.

The results showed that 21 of the 50 subjects believed that the image was computer-generated, suggesting that this group of subjects quickly sensed the difference between the image and the existing human artwork, and thus suggesting that for them the image lost the premise of the discussion of artistic creativity. Another 29 subjects believed that the image was created by a human, indicating that the generated image used in this experiment successfully deceived 58% of the subjects. The mean and standard deviation statistics of the scores for the two questions in Experiment I are shown in Table 1. In that experiment for the optimized generative AI and the traditional generative AI, the mean values of the veracity scores were 0.64 and 0.43, respectively, which indicates that the optimized image generation system in this study has stronger functionality compared to the pre-optimization image generation system.

The subjects’ preferences for the generated image are shown in Figure 1. Four subjects, or 8% of the total, gave scores of 1 or less, indicating that the generated image had an extreme effect on some subjects when stimulating arousal potential and that overstimulation was not completely avoided. Forty percent of the subjects scored below 2.5, and 60% scored above 2.5, indicating that the generated image has a 20% advantage over the hedonic stimulus with lower aesthetic preference when delivering hedonic stimulus with higher aesthetic preference, reflecting that the generated image is more in line with the aesthetic concepts of the subject population, and it can bring them a certain degree of aesthetic enjoyment. As for the subjects’ categorization of the image’s artistic style and the visual illusion effect, the experiment could not draw any obvious conclusions for the time being.

Figure 1.

Liking score

To further confirm the results of Experiment I, in the second round of experiments, the subjects were first presented with questions that guided them to purposefully construct the image semantics starting from a random point and attribute the source of differences to the image styles. The mean and standard deviation statistics of the scores for the six problems in Experiment II are shown in Table 2, and the results of similar experiments of the optimized generative AI are compared with those of the mainstream generative AI techniques GAN (Generative Adversarial Networks), DGGAN (Dual Generative Adversarial Networks), CAN (Creative Adversarial Networks), and NST (Neurostyle Migration).

Overall, it seems that with the development of generative deep learning neural network, the scores from Q3 to Q7 do not have a big change, but the fondness, novelty and ambiguity are in an increasing trend, which indicates that the learning of ambiguous visual illusion type features by this system is successful, and with the increase in the ambiguity of the generated image, it also improves the novelty and the degree of being fondness of the image to some extent. And the result of Q8 shows that Experiment II is 5% more than Experiment I, which indicates that the successive pursuit of the source of the image under the guidance of the evocative potentiality feature improves the difficulty of discriminating the authenticity of the image, i.e., the increase of creativity causes the increase of the difficulty of recognition.

In order to verify the affective nature of the evocative potentiality trait on the aesthetics of environmental public art, the third round of experiments presented subjects with an interrogation of themselves during the process of observing the generated images, and simulated the process of transferring image information from the ideal visual cognitive mode to artistic discernment within the human brain by guiding subjects to consider the content of the generated images in terms of deliberate, structural, communicative, and inspirational aspects. The mean and standard deviation statistics of the scores of the four questions in Experiment III are shown in Table 3, and the results of the performance of the human art paintings Abstract Expression, Art Basel 2016, Conceptual Art, and Minimalism in similar experiments are compared.

Usually, AI art, due to its lack of intentionality, unconsciousness, and social exclusion, is not sufficient to indicate that the level of AI in the current era has reached the level of generating creativity as human artwork, and the experimental expectation did not believe that the present system would lead to better performance. However, unexpectedly, the results of this experiment show that the paintings produced by the optimized generative AI outperform human art paintings across the board in all of the above four aspects. From a prudent point of view, this result may be due to the magnification of individual cases caused by the small test sample size, and thus a single investigation containing only one image generated by this system is not statistically significant and is not suitable for direct comparison with human artworks with sufficient samples. However, it is certain that the system’s introduction of creativity and ambiguity to the generated images was able to be recognized by the subjects and resulted in a positive aesthetic experience.

A total of 608 valid samples were collected in this study, of which 92.8% of art audiences were exposed to AI paintings through social media, 62.2% were generated or previewed by themselves through AI generation platforms, 42.1% and 10.5% of art audiences were exposed to AI paintings through art websites and offline exhibition channels, and 3.0% were exposed to AI paintings through other channels, including private mutual sharing, advertising campaigns, lectures, corporate training and other channels of exposure to AI painting works, the specific data is shown in Figure 2.

Figure 2.

Distribution of contact channels of AI paintings

It can be concluded that the online medium is the main communication channel for AI paintings, but about 10% of AI paintings also enter the art audience in the form of exhibitions offline. As offline exhibitions are guarded by art curators, and as offline exhibitions are a common way of exhibiting and selling works in the art industry, art audiences recognize up to 80% of the exhibited works. In addition, because offline exhibitions involve paying to visit or buy exhibits, the dissemination of AI paintings through this channel also indicates that they have entered commercialization in the art industry.

In addition, among other channels, private sharing and advertising reflect the effectiveness of AI painting in social activity and media content production, while lectures and corporate training also show the importance of AI painting in academic and corporate settings.AI painting work generation toward the trend of building a professional training system, the art audience’s access to the channel in the future may be more systematic and professional.

This paper proposes seven influences on the willingness to accept art generated by artificial intelligence, which are perceived quality, perceived value, perceived cost, perceived risk, social influence, technology anxiety, and media barriers. All variables are operationally defined in relation to the research questions of this study, and measurement question items are identified.

The descriptive statistics of the observed variables in this study were analyzed in terms of the mean and standard deviation of each measurement question item and the mean of each variable as a whole, as shown in Table 4.A 7-point Likert scale was used for each of the observed variables, and the option scores corresponded to the descriptions as follows: not at all conforming - 1 point; relatively not conforming - 2 points; somewhat not conforming - 3 points ; average - 4 points; somewhat conforming - 5 points; more conforming - 6 points; and fully conforming - 7 points.Typically, the mean distribution gives a general indication of the degree of agreement with the description of the observed variable, and the standard deviation is used to analyze whether there is a significant difference between the data. When the mean value is greater than 4, it indicates that the survey sample agrees with the viewpoint, and the higher the mean value, the higher the degree of agreement, while the mean value is less than 4, it indicates that the survey sample does not agree with the viewpoint.

The data results show that the mean values of perceived quality, perceived value, perceived cost, perceived risk, social impact, media barriers, and willingness to accept are above 4 points, indicating that most of the art audience agrees with the views in this variable, in which the mean value of one question item each of perceived value and social impact is lower than 4 points, but they are all higher than 3.5 points, and the degree of deviation from 4 points is not very high and is within the acceptable range. Technology anxiety has the lowest mean value of 3.832 and is below 4 points, indicating that the arts audience does not quite agree that AI technology will replace humans.