Sage Journals: Discover world-class research

Abstract

Although picture extraction is challenging, the murals at Dunhuang are historically significant and offer rich content. The work suggests an image segmentation model based on the Mean Shift algorithm and an area salience prioritisation model to extract the cultural aspects in the Dunhuang murals for landscape design. First, an image segmentation model based on the Mean Shift algorithm is established, and then a region salience value calculation method and a region prioritisation method are designed to establish a region salience prioritisation model. The outcomes showed that a segmentation model built using the Mean Shift algorithm in the study processed a 405175 image with a processing time of 3.18 seconds, an edge integrity rate of 88.9%, an accuracy rate of 87.4%, an F-value of 88.7%, and a total of 302 regions. The segmented Dunhuang image featured few noise points and a distinct shape. Salient region transfer path is more regular and more in line with the human visual transfer mechanism thanks to the research design of the region saliency value calculation method, which also improves saliency detection performance. The highest correct rate when dividing the image is 0.97, the highest check rate is 0.8, and the highest F1 value is 1. In conclusion, the study’s methodology has some favourable implications for landscape design and may be effectively used to extract cultural components from photographs.

Keywords

Mean shift algorithm Image extraction landscaping salient region

1. Introduction

One of the most crucial roles of landscape design (LD) is to create a comfortable living environment, but as living standards rise and societal interest in traditional culture grows, landscape design must meet more humanistic and artistic demands in addition to aesthetic ones [1, 2]. The Dunhuang Mogao Caves are one of the greatest artistic treasures in the world, with a long and rich history. If applied to LD, these cultural elements can give it more historical significance, but after countless years of weathering, changes in the environment, and the influence of various biomes, the colours of Dunhuang’s murals have undergone significant changes and fading [3]. The Mean Shift (MS) algorithm is a common clustering algorithm that converges to a local optimum under certain conditions to achieve accurate localisation of moving objects, and is often used in clustering, image segmentation, object contour checking and target tracking [4, 5]. Based on the MS method, this paper builds an image segmentation model and an area saliency prioritisation model for this situation, in order to extract clear cultural elements of Dunhuang murals and apply them to the design of garden landscapes. This paper introduces two key breakthroughs, the first of which is the development of an image segmentation model based on the MS method. The second point is how area saliency is prioritised. The main structure of the study is divided into four parts. The first part analyses the current state of relevant research; the second part constructs an image segmentation model and a region saliency prioritisation model based on the MS algorithm; the third part analyses the application effect of the proposed model; and the last part concludes the whole study.

2. Related works

The MS algorithm locates the moving body precisely by using a non-parametric probability density estimation technique that converges to a local optimal position. By taking into account the complexity of multiple objects in scene images and using a benchmark dataset, Ahmed et al. [6] proposed an effective multi-class object classification method for indoor and outdoor scene classification of landscape images. This method is an illustration of the fuzzy C-mean and MS algorithms, which can infer multi-objective segmentation in complex images. When used on UAV blue and red band images to show the area and extent of various bare land fields, Vlachopoulos et al. [7] reported that two machine learning algorithms – a random forest non-parametric supervised classifier and an unsupervised non-parametric MS clustering algorithm – performed extremely well. Bahraini et al. [8] proposed an improved MM algorithm based on minimum Bayesian risk to address the issue of labeling errors in hyperspectral image classification and eliminate the impact of errors during the classification process. Ma [9] studied a fixed point tracking model for English reading texts based on mean shift and multi feature fusion. For instance, to estimate the underlying low-dimensional master curve embedded in a high-dimensional space, Ghassabeh and Rudzicz [10] first introduced a modified version of the MS algorithm and then combined it with various variants of the subspace constrained mean shift algorithm, proving that the modifications to the MS algorithm ensured its convergence. Xie et al. [11] proposed a unilateral adaptive truncation index weighted moving average scheme for fast detection of upward or downward mean shifts. The truncation method used helps improve the sensitivity of the recommended scheme to simultaneously detect small and large mean shifts.

LD is the process of using engineering methods and garden art, such as reshaping the terrain, planting trees, constructing buildings, and arranging garden paths, to create lovely natural settings, livable communities, and outdoor leisure places within a specific geographic region. In an effort to explore the applicability of cutting-edge mapping methods and tools to public relations, Liu and Nijhuis [12] conducted semi-structured open-ended interviews with 11 practitioners with backgrounds in landscape architecture. They argued that spatial design is at the core of landscape architecture and that mapping spatial-visual features is important for landscape architects to interpret and talk about space. Sasmal et al. [2] summarized the reported synthetic microcrystalline strategies and revealed the impact of dimensions on their physical and chemical properties and applications, which helps to improve the quality of landscape architecture. In the case of several heritage and legacy projects launched by the South Island Ministry of Arts and Culture to remember those who lost their lives in the conflict, Hami and Abdi [13] believe that students have different preferences for the landscape of open learning areas and leisure places, and proposes to find suitable campus landscape models based on students’ preferences. Ba et al. [14] developed a new landscape design structure based on fused deposition modelling techniques and proposed a laser polishing method for FDM fabricated PLA mechanical components in order to address the high surface roughness that frequently occurs on the surfaces of 3D printed components in landscape architecture, necessitating additional post-treatment. Liu [15] evaluated the various components of agricultural theme park design based on the concepts of sustainability and agroeconomics, supported by the concept of ecology, and also examined the landscape design of agricultural theme parks with case studies. He also considered the need to improve the effectiveness of agricultural theme park landscape design in the context of the concept of agroecology. Loganathan et al. [16] studied corn yellow fiber and wheat yellow fiber to evaluate the reinforcing effect of the fibers in the stems, petioles, and leaves of these two plants in polymer composite materials, which has certain positive significance for landscape design. Geng and Zhu [17] utilized urban landscape design intelligence technology to study the management of municipal infrastructure in China and proposed an adaptive method to introduce China’s best policies and sustainable building planning. They revealed the main direction and residents’ preferences for introducing urban environmental design technology to achieve sustainable regional management in Chinese cities.

In conclusion, even if numerous earlier researchers and academics have supported the MS algorithm’s function in image processing, they have also discovered that LD has historical and humanistic relevance in addition to aesthetic value. However, it is quite uncommon for research to combine the extraction of cultural elements from Dunhuang murals with the application of MS algorithms for image segmentation and saliency prioritisation for LD. As a result, research into the application of cultural elements from Dunhuang murals based on MS algorithms to LD has some promise.

3. Extraction of cultural elements from dunhuang murals based on MS algorithm

Although the Dunhuang frescoes have endured for millennia, they are currently plagued by difficulties with permanent discoloration that make image processing challenging. They are also huge in scale and rich in cultural features. The paper suggests an MS algorithm-based image segmentation model for Dunhuang paintings as well as a region salience prioritisation model for Dunhuang frescoes in order to more precisely extract the cultural aspects in Dunhuang frescoes for LD.

3.1 MS algorithm-based image segmentation model for dunhuang murals

Image segmentation refers to the process of dividing an image into regions with various special properties and extracting interested targets. The main segmentation methods include clustering based segmentation, region based segmentation, and threshold based segmentation. The MS algorithm is a highly effective clustering iterative algorithm widely used in the field of image processing. Its kernel density estimation method, which estimates the density function of the points in the two-dimensional space of an image, is one of the more well-liked methods for estimating density, and the density function for the entire space is estimated by the known sample points. If there are $n$ sample points $x i$ in the d-dimensional space, the calculation of kernel density is shown in Eq. (1).

$\displaystyle f(x)=\frac{\sum\limits_{i=1}^{n}{w(xi)\textit{KH}(x-xi)}}{\sum% \limits_{i=1}^{n}{w(xi)}}$ (1)

In Eq. (1), $w(xi)$ denotes the weights, $H$ represents a d-dimensional parameter matrix and KH(x) denotes the kernel function in d-dimensional space, as shown in Eq. (2).

$\displaystyle\textit{KH(x)}=\left|H\right|^{-\frac{1}{2}}K(H^{-\frac{1}{2}}X)$ (2)

In Eq. (2), the efficiency of the function is greatly reduced if each element in the matrix is not zero, in which case $H$ is replaced with a diagonal matrix form or a unit scale matrix, as shown in Eq. (3).

$\displaystyle\left\{\begin{array}[]{l}H=\textit{diag}[h_{1}^{2},h_{2}^{2},% \ldots,h_{n}^{2}]\\ H=h^{2}I\\ \end{array}\right.$ (3)

In Eq. (3), diag represents the Diagonal matrix, $I$ represents the Identity matrix, and $h$ represents the elements in the matrix. If $H$ is a unit scale matrix and $w(xi)$ is a constant, then the kernel density is more easily calculated as shown in Eq. (4).

$\displaystyle\mathord{\buildrel\lower 3.0pt\hbox{$\scriptscriptstyle\frown$}% \over{f}}h,K(x)=\frac{ck,d}{nh^{d}}\sum\limits_{i=1}^{n}K\left(\left\|{\frac{x% -xi}{h}}\right\|^{2}\right)$ (4)

In Eq. (4), $c k, d$ represents the normalized constant and $K(x)$ represents the multivariate kernel. The three major steps of the MS algorithm are an iterative algorithm. The density of points is estimated for the entire space in the first phase using a data window that has been chosen. The data centre of that data window is determined in the second phase using an estimate of the density of the points. The third step is to check if the window’s centre overlaps the data centre; if it does, the iteration is over; if not, the window’s centre is moved towards the data centre, and the window is built using the data centre, returning to step one. Figure 1 depicts the MS algorithm’s workflow.

Figure 1.

MS algorithm process.

Image segmentation, the process of dividing an image into multiple unrelated regions, can be broadly classified into three categories: edge-based segmentation methods, threshold segmentation methods and region growing segmentation methods. However, with the vast amount of information and complexity of images, no one type of method is suitable for all types of images. The result of image segmentation is generally used as input for image target recognition and feature extraction, and the result will affect further analysis of the image, which is a key step in image processing. The role of image segmentation in the overall image processing is shown in Fig. 2.

Figure 2.

The role of image segmentation in the entire image processing.

The basic principle of the MS algorithm-based image segmentation model is to compute convergence points to represent the centroids of each region through multiple iterations, and then cluster the regions according to their centroids to obtain the segmented region of the image. It consists of three main steps. The first step is to obtain image feature data. For positional features, individual pixel points on a 2D image are represented using $x$ -coordinate values and $y$ -coordinate values. For colour features, images saved using RGB values are transformed and calculated, and LUV values are used for colour characterisation. The second step is to perform Kernel density estimation for all pixels. Set the feature of any point $(i,j)$ on an image of size $m\times n$ to be a vector $\left\{{xi,j,i=1,2,3,\ldots,n,j=1,2,3,\ldots,n}\right\}$ with the kernel density shown in Eq. (5) according to Eq. (4).

$\displaystyle\mathord{\buildrel\lower 3.0pt\hbox{$\scriptscriptstyle\frown$}% \over{f}}h,k(x)=\frac{ck,d}{nh^{d}}\sum\limits_{j=1}^{m}\sum\limits_{i=1}^{n}k% \left(\left\|\frac{x^{L}-x_{ij}^{L}}{hL}\right\|\right)k\left(\left\|{\frac{x^% {C}-x_{ij}^{C}}{hC}}\right\|\right)$ (5)

In Eq. (5), $L$ represents the label of the i-th pixel of the segmented image, and $C$ represents the classification set of pixel points. Let $\left\{{xi,j,i=1,2,3,\ldots,n,j=1,2,3,\ldots,n}\right\}$ and $\left\{{yk,k=1,2,\ldots}\right\}$ be three-dimensional feature vectors, and the iteration function of the MS algorithm is shown in Eq. (6).

$\displaystyle yk+1=\frac{\sum\limits_{j=1}^{m}{\sum\limits_{i=1}^{n}\textit{% xig}\left(\left\|\frac{y_{k}^{C}-x_{ij}^{C}}{hC}\right\|^{2}\right)g(\left\|{% \frac{y_{k}^{L}-x_{ij}^{L}}{hL}}\right\|^{2})}}{\sum\limits_{j=1}^{m}\sum% \limits_{i=1}^{n}g\left(\left\|\frac{y_{k}^{C}-x_{ij}^{C}}{hC}\right\|^{2}% \right)g\left(\left\|\frac{y_{k}^{L}-x_{ij}^{L}}{hL}\right\|^{2}\right)}$ (6)

In Eq. (6), $yk=xij$ . Step 3, an appropriate clustering rule is defined and all eligible points are clustered to obtain the segmented region, thus completing the image segmentation process based on the MS algorithm.

3.2 Regional significance prioritization model construction for dunhuang murals

When examining an image, the human eye is most drawn to the part of the image known as the salient region since it most accurately reflects the image’s content. As a result, processing complicated image data can be made simpler and less challenging by using prominent regions to characterise an image [18]. Based on the human eye’s visual attention process, methods for identifying salient areas in images typically use one of two computer model types: object-based or spatial-based [19]. The study uses an object-based computational model for significant area detection of images of Dunhuang murals because the spatial-based computational model cannot be further analysed at a macro level in the image through this saliency calculation method, which has some limitations. This method is more similar to how the human eye naturally observes the environment and is more conducive to maintaining the independence and integrity of cultural elements in Dunhuang murals. The process of detecting salient regions in images based on object-based computational model is shown in Fig. 3.

Figure 3.

The process of significant region detection in Dunhuang mural images.

When an image is segmented, several regions are obtained, and by comparing the various features between the regions, the saliency values of the image regions can be obtained. The most important features affecting the saliency of an image are colour, shape and location features, which are normalised because they have different measures, scales and ranges of values. The location-based features were used in the calculation of the visual transfer equation, and three colour-based features and eight shape-based features were normalised. Setting each of these 11 features as $f1,f2,\ldots,f11$ and transforming them, then for a region $r k$ , its significant value is defined by calculating its contrast value with the feature quantity of other regions of the image, as shown in Eq. (7).

$\displaystyle ri.s(rk)=\sum\nolimits_{rk\neq ri}{w(rk)}\textit{Dr}(rk-ri)$ (7)

In Eq. (7), $w(rk)$ denotes the weight, and to emphasise that larger regions have greater influence, $w(rk)$ is set to the number of pixel points in region $r k$ , and the sum of the differences between region $r k$ and region feature vectors is shown in Eq. (8).

$\displaystyle D(r1,r2)=\sum\limits_{i=1}^{n}{g(i)\left|{f1,i-f2,i}\right|}$ (8)

In Eq. (8), $g(i)$ denotes the weight of feature $f1$ and $f k, i$ denotes the $i$ th feature value in region $r k$ . After obtaining the saliency values of each region of an image, it is still not straightforward to determine which regions are more or less salient, while the human eye is quick enough to identify and respond to the important parts of a picture containing a large amount of information, so it is possible to simulate the human visual attention shifting mechanism for prioritisation. Visual attention shifting refers to the fact that when looking at a pair of images, a person’s point of view always shifts between different targets in the image according to a certain pattern, which can be divided into implicit attention and explicit attention. The visual attention shifting mechanism also involves a return inhibition mechanism and a location enhancement mechanism. The return inhibition mechanism means that when the human eye shifts the visual focus, it will inhibit the area that has already been attended to and exclude the area that has already been attended to when selecting the next shifted area. Set the set of regions of Dunhuang mural images after segmentation to $\left\{{r1,r2,\ldots,rn}\right\}$ . The return suppression factor is calculated as shown in Eq. (9).

$\displaystyle\textit{I(ri)}=\left\{{{\begin{array}[]{ll}0,&\textit{ri}\in S\\ 1,&\textit{otherwise}\\ \end{array}}}\right.$ (9)

In Eq. (9), $S$ represents the set of salient regions that have been selected. The location enhancement mechanism of viewpoint shift refers to the fact that the closer the human eye is to the current target area when shifting visual focus, the more likely it is to become the next salient target for viewpoint shift. The location enhancement factor is calculated as shown in Eq. (10).

$\displaystyle L(ri)=\exp\left(-\frac{D(ri,rk)}{\sigma^{2}}\right)$ (10)

In Eq. (10), $\sigma^{2}$ controls the weight of the location enhancement factor, the larger the weight the smaller the effect of the location enhancement factor, $r k$ indicates the current salient region, $r i$ indicates the other regions, and $D(ri,rk)$ indicates the Euclidean distance between the current region and the other regions. The principle of exclusivity means that only one region can be selected at a time for visual focus shifting, and the significant shift value of BB region is calculated as shown in Eq. (11).

$\displaystyle\textit{f(ri)}=\textit{I(ri)L(ri)s(ri)}$ (11)

In Eq. (11), $s(ri)$ denotes the significant value corresponding to region $r i$ . The next significant region for the focus shift is shown in Eq. (12).

$\displaystyle\textit{rx}=\arg\textit{ri}\max\textit{f(ri)}$ (12)

A salient region shift termination condition was set to determine whether the salient region shift should be terminated. After calculating all the shifted regions by Eq. (12), hypothesis $S=\left\{{a1,a2,\ldots,an}\right\}$ , the visual focus shift condition is shown in Eq. (13).

$\displaystyle\frac{\sum\limits_{j=1}^{k}{\textit{f(aj)}}}{\sum\limits_{i=1}^{n% }{\textit{f(ai)}}}\leqslant\varepsilon$ (13)

In Eq. (13), $\varepsilon$ denotes the visual transfer inflection factor. The segmentation of key regions of the photographs will reveal variances as the style of the Dunhuang frescoes varies from time to time. As a result, it’s crucial to prioritise the areas of importance in order to research the traits of the cultural components in Dunhuang frescoes from various eras and apply them to LD. As shown in Fig. 4, the images of the Dunhuang frescoes were first segmented using an MS algorithm-based image segmentation model to create a group of segmented regions, and then the region saliency prioritisation model was used to determine the saliency value of each region and to prioritise the process.

Figure 4.

Priority division and processing process of Dunhuang mural images.

4. Analysis of the effectiveness of the cultural elements of dunhuang murals extraction method based on MS algorithm

The paper suggests an MS algorithm-based image segmentation model for Dunhuang murals and a region salience prioritisation model for Dunhuang murals in order to extract the cultural components of Dunhuang paintings for use in LD, however its effectiveness needs to be further confirmed. The study focuses on two analysis-related issues. The first section examines the performance of the region saliency prioritisation model for Dunhuang murals, and the second section examines the performance of the MS algorithm-based image segmentation model for Dunhuang murals.

4.1 Analysis of the effectiveness of dunhuang mural image segmentation model based on MS algorithm

The chromaticity domain (hr), the null domain (hs), and the threshold value for the number of minimum pixel points in the region are three crucial variables that affect the outcome of MS-based picture segmentation. The study conducted image filtering processing under different HR and HS conditions, and the results are shown in Fig. 5. From the figure, it can be seen that the images obtained by filtering under different parameters also vary, but Fig. 5(d) has the best overall effect, indicating that the two parameters hr and hs play a crucial role in this step, and hs has a greater impact on the smoothing results.

The more clearly the extracted image edges are, the lovelier the garden landscape will be. This is one of the requirements for LD. The results of edge extraction experiments on a 211 by 266-pixel image are displayed in Table 1 along with comparisons to common edge detection techniques including Canny, Gabor, and LSD. The MS algorithm described in the study has, as can be seen from the table, the highest completeness, accuracy, and F-value compared to the other three algorithms, with 88.9%, 87.4%, and 88.7% respectively, making the pictures processed using the MS algorithm more suitable for LD.

Table 1
Image edge extraction test results

Measure	Canny	LSD	Gabor	MS
Comp (%)	66.5	71.8	72.6	88.9
Corr (%)	58.4	69.2	65.1	87.4
F (%)	62.3	70.8	68.9	88.7

Table 2

MS image segmentation information

Image name	Image size	Segmentation time	Number of regions
(1)	405 $\times$ 205	3.37 s	150
(2)	405 $\times$ 175	3.18 s	302
(3)	395 $\times$ 405	4.43 s	402
(4)	405 $\times$ 215	3.44 s	360
(5)	405 $\times$ 345	4.36 s	716

Figure 5.

Results under different chromaticity and spatial domains.

The processing of the Dunhuang mural photographs is more challenging because they are more complicated and exhibit fading and peeling. Figure 6 displays the outcomes of the segmentation procedure performed on the mural images from Dunhuang. Figure 6 displays the outcomes of the segmentation procedure performed on the mural images from Dunhuang. The MS algorithm-based image segmentation model has a good segmentation effect, with distinct contours and minimal noise. The experimental results demonstrate certain application implications of the MS algorithm-based Dunhuang mural picture segmentation model.

Figure 6.

The segmentation processing results of Dunhuang mural images.

The segmentation information of Fig. 6(b) is shown in Table 2. As can be observed from the table, Fig. 2 has the shortest segmentation time of 3.18 s and the lowest number of regions of 302. Figure 3 takes the longest time to segment, 4.43 s, and has the highest number of regions at 402. The efficiency of the MS algorithm based image segmentation model is therefore affected by the size of the image.

In conclusion, the MS algorithm-based image segmentation model developed in the study is better at segmenting the target object in detail, effectively retaining boundary information, and improving segmentation effects for difficult-to-process Dunhuang images. However, segmentation efficiency of the model is influenced by image size; the larger the image, the longer the segmentation time.

4.2 Analysis of the effectiveness of a regional salience prioritization model for dunhuang murals

The study fixed the weight $g(i)$ of the eigenvalue $f i$ to 1 in order to test the efficiency of the region saliency computation approach it had suggested. To create a saliency grey-scale map of the image, the saliency values of each region segmented by the MS algorithm were calculated using Eq. (7), normalised, and multiplied by 255. A set of features was chosen to describe the salient characteristics of the image, and the saliency value of the region was defined by the degree of contrast of the amount of features in this set of regions and compared with the Frequency Tuning (FT) and Context Aware (CA) algorithms, and the results are shown in Fig. 7. The FT method has a clearer profile, as can be seen from the graph, and the CA algorithm performs the poorest. The FT algorithm is a space-based computational model of saliency detection, and the intermediate result calculated is the saliency value of pixel points, not the saliency value of regions. As a result, it is not applicable to this study, which requires the division of the saliency priority of regions. The computer model of the item serves as the foundation for the significant region detection method used in the study. The saliency value is determined for each segmented region once the image has been processed and segmented, ensuring that each segment has the same brightness. Comparing the three approaches, it can be said that, when saliency prioritisation is necessary, the saliency detection approach suggested in the study works better than the other two saliency detection approaches.

Figure 7.

Significant area detection results of Dunhuang murals.

The ROC curve was used to reflect the performance of the saliency map detected by the model, with the threshold set to 0–255, and the GBVS model saliency map, AIM model saliency map, and IG model saliency map were compared. The results are shown in Fig. 8 to further validate the performance of the region saliency calculation method suggested by the study. The model put out in the study is more similar to the visual information provided by human eyes than previous models, as can be seen in the picture.

Figure 8.

ROC curves for each significance model.

$\sigma^{2}$ was set to 0.4 and $\varepsilon$ was set to 0.9 to test the efficiency of the species image region prioritisation approach. The significant values of the segmented regions were calculated and transferred using the significant degree transfer function from large to small, and the results are displayed in Fig. 9 after selecting various periods of fiducial maps and segmenting them with the MS algorithm. The second image has a noticeably wider span of visual transmission on the figure, as can be observed from above the figure. In order to assess the evolution of the culture of Dunhuang wall murals, the study suggests a system for ranking image areas that indicates the range of visual transfer.

Figure 9.

Significant transfer results of Dunhuang mural images.

Figure 10 displays the comparison between the transfer method used in the study and the visual transfer method according to the magnitude of the significant value. As seen in the picture, compared to the way of transferring according to the importance value, the suggested visual transfer method has a more regular transfer path of significant regions. Due to the location enhancement mechanism included in the approach, the transfer path of salient regions can be made more regular and is also more consistent with the human visual transfer mechanism. As a result, the suggested visual transfer method is reasonable and efficient.

Figure 10.

Significant regional transfer path.

According on the degree of fading, the study split 1200 photographs of Dunhuang murals into three classes, including 400 images in each class. 100 photos were chosen as the test set for the prioritising process. The weighted voting technique and the general average method of the two algorithms were compared, and the results are displayed in Fig. 11 in terms of three metrics: correctness, completeness, and F1. The model proposed in the study, which has the highest correctness of 0.97, the highest completeness of 0.8, and the highest F1 value of 1, has the highest correctness, completeness, and F1 values for the classification of the three classes of images compared to the other two methods.

Figure 11.

Test results of three algorithms.

In conclusion, the area saliency prioritisation model for the mural paintings of Dunhuang developed in this work performs better in terms of saliency detection, is more appropriate for saliency prioritisation, and is more similar to the visual information provided by the human eye. The region prioritising approach represents the range of visual transfer, aids in the analysis of the evolution of the Dunhuang mural culture, has some justification and validity, and the major region transfer path is more regular and also more in line with the human visual transfer mechanism.

5. Conclusion

Higher historical and cultural connotation requirements are placed on LD. The skin sections of the characters in the long-standing and culturally significant Dunhuang paintings have irreversibly faded, making image processing difficult. The work suggests an area saliency prioritisation model and an MS algorithm-based picture segmentation model for the challenging task of extracting cultural features from Dunhuang murals. The results show that the MS algorithm-based image segmentation model constructed in the study processed the highest image edge integrity rate, accuracy rate and F-value with 88.9%, 87.4% and 88.7% respectively. 3.18 s were taken to process a 405 $\times$ 175 image segmentation and the number of regions obtained was 302, and the larger the image, the longer the segmentation time. The research design of the region significant value calculation method significance detection performance is better, the region priority division method can reflect the span of visual transfer, with a certain degree of rationality and validity, significant region transfer path more regular, and also more in line with the human visual transfer mechanism. Its highest correct rate is 0.97, the highest check-all rate is 0.8, and the highest F1 value is 1. In conclusion, the study’s built-in model has some validity and plausibility. The scale of the image has an impact on the effectiveness of the image segmentation approach utilised in this study, and the features that are extracted are not necessarily representative. An adaptive algorithm must be created to process the images and choose more representative features that are better suited for the extraction of cultural elements from Dunhuang murals because this may affect the practical application of extracting cultural elements from Dunhuang murals for LD.

References

Gulpinar Sekban

Acar

. Determining usages in post-mining sites according to landscape design approaches. Land Degrad Dev. 2021; 32(8): 2661-2676.

Sasmal

Mahato

Majumder

Banerjee

. Landscaping covalent organic framework nanomorphologies. J Am Chem Soc. 2022; 144(26): 11482-11498.

Zeng

Sun

Yin

Shen

. Mobile visual search model for Dunhuang murals in the smart library. Libr Hi Tech. 2022; 40(6): 1796-1818.

Bhat

Subudhi

Veerakumar

Caterina

DGJ

. Soraghan

. Target tracking using a mean-shift occlusion aware particle filter. IEEE Sens J. 2021; 21(8): 10112-10121.

Stelling-Wood

Poore

AGB

Gribben

. Shifts in biomass and structure of habitat-formers across a latitudinal gradient. Ecol Evol. 2021; 11(13): 8831-8842.

Ahmed

Jalal

Kim

. A novel statistical method for scene classification based on multi-object categorization and logistic regression. Sensors. 2020; 20(14): 3871.

Vlachopoulos

Leblon

Wang

Haddadi

LaRocque

Patterson

. Delineation of bare soil field areas from unmanned aircraft system imagery with the mean shift unsupervised clustering and the random forest supervised classification. Can J Remote Sens. 2020; 46(4): 489-500.

Bahraini

Azimpour

Yazdi

. Modified-mean-shift-based noisy label detection for hyperspectral image classification. Comput Geosci-UK. 2021; 155: 104843.

. Fixed-point tracking of English reading text based on mean shift and multi-feature fusion. J Intell Fuzzy Syst. 2021; 40(4): 6315-6325.

10.

Ghassabeh

Rudzicz

. Modified subspace constrained mean shift algorithm. J Classif. 2021; 38(1): 27-43.

11.

Xie

Castagliola

Sun

. One-sided adaptive truncated exponentially weighted moving average x¯ schemes for detecting process mean shifts. Qual Technol Quant M. 2022; 19(5): 533-561.

12.

Liu

Nijhuis

. The application of advanced mapping methods and tools for spatial-visual analysis in landscape design practice. Sustainability. 2021; 13(14): 7952-7952.

13.

Hami

Abdi

. Students’ landscaping preferences for open spaces for their campus environment. Indoor Built Environ. 2021; 30(1): 87-98.

14.

Wen

. Mechanical characterization and thermodynamic analysis of laser-polished landscape design products using 3D printing. Materials. 2021; 14(10): 2601-2601.

15.

Liu

. Landscape design of agricultural theme park based on ecological concept. Acta Agr Scand B-S P. 2021; 71(9): 794-805.

16.

Loganathan

Sultan

MTH

Jawaid

Ahsan

Naveen

Perumal

. Characterization of new cellulosic Cyrtostachys renda and Ptychosperma macarthurii fibers from landscaping plants. J Nat Fibers. 2022; 19(2): 669-684.

17.

Geng

Zhu

. Municipal infrastructure management using smart city landscape design and sustainable architecture technologies. Open House Int, 2021; 46(6): 563-577.

18.

Feng

Fan

. Saliency-based image correction for colorblind patients. Comput Visual Media. 2020; 6(2): 169-189.

19.

Liu

Guo

Tian

Bao

. An improved YOLOv5 model based on visual attention mechanism: Application to recognition of tomato virus disease. Comput Electron Agr. 2022; 194: 106780.