Cloth filter adversarial and graph cut roi based segmentation for intelligent detection of defects in monument using UAV

ABSTRACT

A nation’s cultural heritage inspires people to learn about their history and innovate. It helps build social capital linking economic progress to social ideals and increasing quality of life. Monuments depict distortion and collapse against complex architectural patterns. Heritage monument fault identification must be resilient to architectural types. Human-taken photographs with erroneous data acquisition lower defect detection accuracy. While object detection is constantly evolving numerous deep learning approaches with excellent accuracy have been used to detect monument defects from UAVs to aid archaeological surveys and scholars. However, UAV photos with poor contrast and brightness require a new monument flaw identification algorithm. This study proposes UAV-based Cloth Simulated Filter Adversarial Network and Graph Cut (CSFAN-GR) ROI-based segmentation. Pre-processing and ROI-based segmentation comprise the CSFAN-GR monument flaw identification method. UAV photos are pre-processed using Cloth Simulation Filter and Generative Adversarial Network model. These two methods improve UAV image denoising using Generative Adversarial Network on raw images. Using image denoised data as input, Dijkstra Graph Cut ROI-based Segmentation model addresses structural abnormalities. Simulations will evaluate the proposed CSFAN-GR ROI-based segmentation approach using Matlab Simulink and training, testing and validation will examine and validate monument defect detection performance parameters, precision, recall, F1-score and PSNR.

Keywords:
Monument defect detection; Cloth simulated filter; Generative adversarial network; Dijkstra Graph Cut; ROI based segmentation.

1. INTRODUCTION

Cultural monuments are considered as both the tangible and intangible heritage of a specific group of society transmitted from generation to generation. Cultural monuments and historical artifacts appear as the combined memory of societies, warranting future generations to get a sight of into the past. Societies require to outlook cultural monuments as a harmonizing element and organize safeguarding and maintaining cultural monuments to future generations as a consideration of their past.

Photogrammetry-based Digital Twin was proposed in [¹] using three-dimensional (3D) geometric documentation and different instruments and procedures were also involved in the design to aid the overall procedure of crack detection. Both photogrammetry and geodetic methods were integrated to generate accurate form of 3D geometric documentation. Also results with traditional methods for analyzing automatically detected cracks resulted in improved accuracy. Pyramid Scene Parsing Network (PSP_R50) was proposed in [²] for automatic detection of damp on history masonries. Here the segmentation models were introduced that in turn ensured improved accuracy. To consider Portuguese glazed strips, quality was enhanced. Hence early defect detection remains major aspects to be resolved.

A UAV based routing and scheduling using mixed-integer linear programming was proposed in [³] for engineering managers with the intent of providing valuable insights into architectural monuments. A survey of deep learning technique for monument defect detection was investigated in [⁴]. Complementary cloth simulation technique was applied in [⁵] to filter out the noisy portions in building in the process of defect detection. Yet another You Only Look Once (YOLO) method was employed in [⁶] for detecting defects in cultural heritage buildings with improved accuracy. In [⁷], deep residual network was proposed to focus on structural integrity.

An accurate assessment of buildings is a demanding task. Safety evaluation on the basis of both qualitative and quantitative data is required prior to the interpretation of any data. Quantitative data necessitates complicated mechanisms that enable professionals make both time and money consuming. Hence, it is very paramount to employing sophisticated mechanisms with the intention of measuring probable risk of buildings as a preliminary step towards preservation.

To analyze external building structure conditional survey protocol matrix was employed in [⁸] with the intention of analyzing structural failure in an efficient manner. A case study of visual inspection with the aid of micro UAV was investigated in [⁹] to not only obtain defect information but also performing detailed visual inspection for both maintenance and conservation precisely. Nevertheless, the defects in concrete surfaces cannot be steadfastly and efficiently identified employing conventional crack detection methods. An intelligent model for crack detection employing image processing techniques was designed in [¹⁰].

Inheritance structure is beneficial significant compositions for enormous amount of attention from international community. In [¹¹], surface defect was determined by Swin Transformer as well as YOLOv5. Also by using this resulted in the highest accuracy level promoting surface defect detection. Yet another deep learning technique called, U-Net and Feature Pyramid Network (FPN) architecture was proposed in [¹²] to detect manufacturing defects. A non-destructive testing method was designed in [¹³] for assessing bridge condition in monument. The method was not only proved to be accurate but also ensured low cost.

1.1. Related works

Culture in any facet plays a significant role in country overall growth. The Culture denotes the objectives and overall nation values. With India possess a large cultural diversity, its monuments plays a paramount part in imparting perceptions of its rich cultural heritage and architectural past. Hence monument defect detection remains a major part to be focused to sustain cultural heritage.

Gold foil flaking of stone artistic inheritance was recognized [¹⁴] via Mask R-CNN. Also, with aid of R-CNN model the defects were identified even in the presence of large interference factors with high average precision. Yet another deep learning technique was proposed in [¹⁵] for both classification and detection of defects with improved precision and accuracy. An integration of K-means clustering and multilayer neural networks were presented in [¹⁶] to detect the decaying factor involved in the monuments for analysis. By using this hybrid mechanism resulted in the improved accuracy level.

A novel vision transformer-based autoencoder that included region-attentive patch projection for Intelligent Cultural Heritage Monitoring was proposed in [¹⁷]. Here, Region-attentive patch projection technique was applied for defect detection and additionally complex patterns were recognized employing structural similarity index measure (SSIM). A systematic review of fault detection in building employing deep learning was investigated in [¹⁸]. Yet another crack detection method employing improved U-Net network was proposed in [¹⁹].

An overview of image segmentation method employing computer vision based deep learning for defect detection was investigated in [²⁰]. Hyperspectral images were employed in [²¹] where deep and tensor based learning techniques were applied for inspection of cultural images in an automatic fashion. Also using these techniques ensured improved accuracy in addition to robustness against overfitting. A systematic review on defect detection employing deep network was investigated in [²²].

A novel channel attention mechanism in addition to faster R-CNN method was presented in [²³] for ensuring automatic defect detection in an accurate and precise fashion. Cultural heritage was preserved [²⁴] with digital twin mechanism. Class activation maps along with explainable artificial intelligence were presented in [²⁵] for assessing defect evolution over time. Forest areas classification was carried out [²⁶] via DL. With gathered UAV images, forest coverage area is evaluated with the performance of deep learning algorithm along with stacked Auto-encoder. It ensures better detection accuracy on classification algorithm than other machine learning algorithms.

The CNN based automated weed detection system were developed in [²⁷] using UAV image. The designed CNN is a type of DL model that used to prevent the safety of crops and environmental stress by identifying most essential image pixels. The classification of pictures for weed detecting is conducted by utilizing a real dataset with UAV images. Planet scope nano satellites image classification using machine learning was portrayed in [²⁸] for detecting continuous PS nano satellite vegetation. It accurately classifies vegetation images with higher accuracy. Based on aforementioned methods, hybrid Cloth Simulated Filter Adversarial Network and Graph Cut (CSFAN-GR) ROI-based segmentation using UAV developed.

1.2. Research gap

Aim of intelligent detection of monument defects using UAV remains in defining certain distinct assessment of negative impact on the monument that is potential to disclose the noise connected with sample real-time images obtained via drone under test thus improving PSNR with improved accuracy in the near future. Using Photogrammetry-based Digital Twin [¹] method though ensured accurate form of 3D geometric documentation however, the removal of noise in the image was not focused. Pyramid Scene Parsing Network (PSP_R50) [²] of recall was unable to consider for crack detection.

1.3. Contributions

To handle this problem, PSNR, accuracy is concentrated for monument defect detection. Proposed CSFAN-GR ROI-based segmentation using UAV is designed. Contribution is given by,

• CSFAN-GR ROI-based segmentation is to intelligent monument defect detection

• Floor points divided as of non-ground points using Cloth Simulation Filter function and mapping relationship between noisy and clean images are performed using min-max game theory via Generative Adversarial Network to raw sample intact and crack temple images. In this way, PSNR is increased.

• To enhance both quality as well as quantity, ROI-based segmentation model called, Dijkstra Graph Cut ROI-based Segmentation is employed.

• Experiments evaluated by CSFAN-GR as well as traditional methods for intact and crack temple images.

1.4. Road map

It organized by, literature review described in Section 2 for filtering as well as ML. CSFAN-GR is explained in Section 3. Simulation outcome is presented in Section 4. Comparison of CSFAN-GR as well as traditional schemes is portrayed in Section 5. Conclusion is shown in Section 6.

2. METHODOLOGY

Vision-based regular exploration of stone monument decay over a period of time is both considered to be laborious and time consuming. Specifically in the event of wide-reaching immovable cultural heritage, the workforce is enlarging in a considerable manner, in addition to the probability of error contingency. Any camouflaged in deterioration can lead to irreversible structural issues in monuments and hence it is imperative to design possible exploration methods. Computer-vision methods denote an efficient mechanism to discard both human errors and problems in the field. Therefore this work aims to adopt the hybrid method called, Cloth Simulated Filter Adversarial Network and Graph Cut (CSFAN-GR) ROI-based segmentation using UAV as shown in Figure 1.

Figure 1
Structure of cloth simulated filter adversarial network and graph cut (CSFAN-GR) ROI-based segmentation.

As shown in the above figure, a hybrid method called, Cloth Simulated Filter Adversarial Network and Graph Cut (CSFAN-GR) ROI-based segmentation using UAV is proposed. CSFAN-GR for monument defect detection divided for pre-processing and ROI based segmentation. First, Cloth Simulation Filter function is applied to the raw images with the intent of separating ground points from non-ground points. Following which, mapping relationship among noisy as well as clean images is traced by Generative Adversarial Network. Image denoised outcomes are achieved with loss. Second, ROI-based segmentation employing Dijkstra Graph Cut is applied to strengthen boundary portions in an accurate and precise manner.

2.1. Data description

Eswaran Temple drone images contain about 317 annotated images of fractured as well as whole walls of historical building. By DL, goal of method is to identify defects in monuments employing pre-processing and segmentation. Such methods are presently extensively utilized in the monument defect detection. Genuine images of past construction cracks are captured from Kottaya Easwaran Temple located in Erode fort area. It was built during the era of 1177 AD by the Konguchozhan Karikalan who ruled Dharapuram as Rasarasapuram. Eswaran Temple drone images dataset containing 317 annotated images of fractured as well as non-cracked exterior.

Proposed CSFAN-GR method performs monument defect detection by considering images using Eswaran Temple drone images dataset. Input images are extracted into training, testing and validation images. From that, 222 (70%) images are training images, 63 (20%) of testing images and 32 (10%) of validation images.

The proposed method is designed to work not only for the Eswaran Temple dataset but also for other heritage structures with different stone textures, colors, and architectural styles. Instead of depending on a single type of surface pattern, the model learns general structural and texture-related features that are commonly associated with defects such as cracks, spalling, discoloration, and erosion.

The preprocessing and filtering stages also help reduce the effect of lighting variations, noise, and surface irregularities, which improves the model’s ability to adapt to different heritage environments. Because of this, the method can identify damage patterns even when the appearance of the structure changes. However, since heritage sites can vary significantly in terms of material composition and environmental conditions, some performance variation may still occur when the model is applied to completely new datasets. To further improve generalization, future work will include testing the method on multiple heritage sites and using larger and more diverse training datasets.

2.2. Cloth simulation filter and generative adversarial network model

Image pre-processing is the initially step that had to be done prior to segmenting a region of interest (ROI) from complex images. The intent of pre-processing is to improve significant image features and boost their quality for further processing. Image preprocessing covers but is not constrained to illumination rectifications, filtering, noise removal etc. In our work Cloth Simulation Filter and Generative Adversarial Network-based pre-processing model is applied for noise removal and extraction of valuable information from UAV images via point cloud.

The Cloth Simulation Filter (CSF) was originally developed for LiDAR data to separate ground points from non-ground objects by simulating how a cloth settles over a surface. In this work, the same idea is adapted for 2D image denoising. Instead of using 3D point cloud heights, the method treats image pixel intensity values as surface variations.

During the filtering process, the simulated cloth moves smoothly over the image surface and reduces unwanted noise while preserving important structural details such as cracks, edges, spalling, and discoloration regions. Neighboring pixel information is used to control the smoothing process so that random noise is removed without affecting the actual defect patterns.

This adaptation allows the method to improve image quality before feature extraction and defect analysis. As a result, the proposed approach can achieve more accurate and stable defect detection, especially in complex structural images containing different types of surface damage.

The proposed method is capable of detecting not only simple cracks but also more complex defects such as spalling, discoloration, and surface texture changes. It can effectively analyze images containing multiple defect types at the same time by learning different visual patterns from the structural surfaces. During the evaluation, the method showed stable and accurate performance even in challenging conditions where defects appeared together or were visually irregular. This indicates that the proposed approach is robust and suitable for practical real-world structural inspection applications.

The point cloud in the setting of UAV indicates a group of 3D data points obtained from the drone, formulating a virtual depiction of object being analyzed basically imparting a comprehensive 3D map of area taken by the drone’s camera with each point in the cloud denoting a particular location permitting for exact measurements and analysis.

The Orthorectification of real input UAV images is a basic standard preprocessing stage to provide high quality UAV image. For Orthorectified images, input images are processed to identify Camera tilt and lens distortion. The determined geometrical information provides orthorectified UAV images with uniform scale and consistent distance.

The image lighting variation is occurred due to changes in shadows, sun angle and exposure variations. The lighting significantly affects GANs performance of segmentation tasks. For improved GAN based image denoising, image lighting is normalized by adjusting image pixels values. Apart from that, UAV images are vary in scale due to altitude changes or perception. When the changes occurred in image altitude or terrain height, scale differences is achieved and it is handled by performing photogrammetric reconstruction phase. The scale differences is normalized based on ground sampling distance, image resizing. Patch extraction and data information. This supports to detect scale differences and ensure uniform spatial image resolution. Thus, the normalization of lighting variation and scale differences are attained with the combination of radiometric calibration, relative normalization techniques and consistent flight planning.

UAV flight parameters such as altitude, camera angle, and image resolution are used for affecting performance of proposed method with higher quality and structural defects visibility. For example, images captured from higher altitudes may reduce the visibility of small cracks, while different camera angles can introduce distortion or lighting variations. Similarly, low-resolution images may make it difficult to identify fine surface details.

To handle these challenges, the proposed method includes preprocessing and feature extraction steps that help improve image quality and reduce the impact of noise and illumination changes. The model was also evaluated using images captured under different viewing conditions, which helped improve its ability to handle moderate variations in UAV settings.

First, the images acquired from UAV are subjected to pre-processing using Cloth Simulation Filter and Generative Adversarial Network model. During the initial analysis of monument defect detection model using the images acquired from UAV several cracks are found to be occluded by vegetation, making defect detection a laborious and time consuming process. This in turn do not exhibit clear planarity. To address on this issue, two steps are performed during preprocessing called, Cloth Simulation Filter and Generative Adversarial Network. These two steps are applied to the raw images acquired from UAV using Cloth Simulation Filter. Noisy as well as spotless images are mapped via loss function through min-max game theory. Generative Adversarial Network is to attain denoised outcome.

To classify original points into ground points, Cloth Simulation Filter technique obtains final cloth shape and bias is utilized. Here, ground points carry out for procedure of filtering data by drone with the intent of identifying and separating points from actual ground surface. The ground surface points are specified as vegetation that includes various plants and trees in the temple region.

CSF process classifies ground surfaces as vegetation, true ground and monument structure. The vegetation noise specified as irregular points, true ground denotes continuous base layers of surfaces and monument structure is preserved based on geometric differences. From that, smooth surfaces are reconstructed for ensure points labeled as ground, while vegetation noise and monument structures are effectively distinguished and preserved.

Original point cloud has rotated for down. Cloth drops to inverted surface from above. By exploring the relationships between nodes of cloth and the subsequent UAV carrying LiDAR points, the final cloth shape is obtained. It utilized as a basis in classifying original points to ground or non-ground points. By utilizing a UAV denotes to the procedure of filtering data acquired by drone with the intent of identifying and separating the points denoting actual ground surface from points denoting objects to name a few being, vegetation, including plants and trees typically those in the temple region.

In the proposed Cloth Simulation Filter technique discrete floating hemisphere and square canopies as given below are initially simulated for sample images.

(1) $$zH=SI\left[\sqrt{r^2-x^2-y^2}\right]$$

(2) $$zC=SI\left[r*\cot \theta \frac{\sqrt{x^2+y^2}}{\tan \theta }\right]$$

From the above equations (1) and (2) the floating hemisphere ‘zH’ and square canopies ‘zC’ of the corresponding sample images obtained from Eswaran Temple drone images are first simulated. ‘(x, y, zH)’ denotes simulated point coordinates, ‘r’ and ‘θ’ represents base radius as well as angle among slant height (i.e. distance from square apex to base edge) and square height (i.e., perpendicular distance from square apex to base center). Pre-processing includes noise eradication, height normalization and grid. Eswaran Temple drone information eliminates maximum or minimal noise points. Secondly, to extract ground point’s, noise eliminated images are used for estimating relation height above ground in all point. Finally, extracted ground points’ are used in obtaining height normalized points.

Proposed Cloth Simulation Filter is based on the simulation of cloth-touching objects, a straight forward pre-processing procedure. Let us consider that the cloth piece is positioned above a height normalized landscape and this cloth piece owing to gravity drops progressively. On the basis of this technique, the Cloth Simulation Filter employed in our work obtains ground points of sample images via UAV or drone to acquire data. To confirm cloth shape at a definite time (i.e., the sample images obtained at a define time), the cloth particles (i.e. the sample images) positions in the 3D space are measured. Each particle position via UAV is obtained by both external and internal force. Association among particles location (i.e. the sample image position) and internal/external forces is measured via Newton’s second law.

(3) $$M\frac{\partial S{I}_{Pos}(t)}{\partial t^2}=F_{ext}(S{I}_{Pos},t)+F_{int}(S{I}_{Pos},t)$$

From the above equation (3), ‘M’ refers to the mass of cloth particle or the (mass of sample images) with ‘SI_Pos’ denoting the position of sample image at time ‘t’ and association or the relationship between particles position (i.e. the sample image position) obtained via external force ‘F_ext’ and internal force ‘F_int’ respectively. Displacement vector ‘${\overrightarrow{S{I}_{Pos}(\mathrm{\Delta }t)}}_{ext}$’ of each pixel in sample image following single time step ‘∆t’ obtained below external force and is represented as given below.

(4) $${\overrightarrow{S{I}_{Pos}(\mathrm{\Delta }t)}}_{ext}=vel\mathrm{\Delta }t\cdot {\overrightarrow{norm}}_{ext}$$

From the above equation (4) with ‘vel’ denoting the velocity and ‘${\overrightarrow{norm}}_{ext}$’ representing normalized vector pointing to the movement directions, the displacement vector under the external force is obtained for the corresponding sample image. Displacement vector ‘${\overrightarrow{S{I}_{Pos}(\mathrm{\Delta }t)}}_{int}$’ of each pixel in sample image, ‘∆t’ achieved under internal force and is represented as given below.

(5) $${\overrightarrow{S{I}_{Pos}(\mathrm{\Delta }t)}}_{int}={\displaystyle {\sum }_{i=1}^{8}Const}\left(S{I}_{Pos}{}_{{}_{adj}}-S{I}_{Pos}{}_{{}_{curr}}\right){\overrightarrow{norm}}_{int}$$

From the above equation (5) ‘Const’ is referred to as a constant with ‘0’ denoting the pixel absence (i.e. where a pixel is missing or has not data value at a specific location within a sample image) whereas ‘1’ denoting the pixel presence (i.e. presence of individual pixel within a sample image) respectively. In addition, ‘$S{I}_{Pos}{}_{{}_{curr}}$’, ‘$S{I}_{Pos}{}_{{}_{adj}}$’ refers to the height value of current pixel and height value of adjacent pixel and ‘${\overrightarrow{norm}}_{int}$’ denoting normalized vector pointing to the movement directions, the displacement vector under the internal force is obtained for the corresponding sample image. Finally, the displacement vector ‘$\overrightarrow{S{I}_{Pos}(\mathrm{\Delta }t)}$’ of each pixel in sample image after one time step ‘∆t’ is measured as given below.

(6) $$\overrightarrow{S{I}_{Pos}(\mathrm{\Delta }t)}={\overrightarrow{S{I}_{Pos}(\mathrm{\Delta }t)}}_{ext}+{\overrightarrow{S{I}_{Pos}(\mathrm{\Delta }t)}}_{int}$$

(7) $$GP=\overrightarrow{S{I}_{Pos}(\mathrm{\Delta }t)}$$

(8) $$NGP=1-\overrightarrow{S{I}_{Pos}(\mathrm{\Delta }t)}$$

To start with association or the relationship between particles position, denoting initial state of intact, crack image, displacement by external force of intact, crack image, displacement by internal force of intact, crack image and final state of intact and crack image are obtained. Earth points are speared in an efficient manner.

The proposed method classifies raw sample input images into intact (unbroken) and crack temple images. By performing preprocessing phase, image pixels are segmented to identify the relationship between particles position, initial state of intact, crack image, displacement by internal force of intact and final state of intact and crack image. The performance of classification task labels ground points from non-ground points in an efficient manner. The inter-annotator agreement is conducted to verify and ensures reliability for high-precision segmentation. Next, Generative Adversarial Network with min-max game theory is employed for generating spotless images.

The main purpose of selecting Generative Adversarial Networks (GANs) architecture in filtering process is to achieve underlying structure of data. GAN captures UAV image patterns, textures and relationship among various images, audio, or video. The captured images help to separate region of interest from complex images.

The min-max game theory is carried to perform loss function for better UAV image denoised results. GANs comprise Generator as well as Discriminator. Here, generator makes information as well as discriminator acts classifier determining if information is genuine (as of training set) or false (produced with generator). It achieves effective data classification of real or fake and provides removal of fake data with higher accuracy.

Owing to reason that original GAN is unbalanced, noise combined with generative ‘Gen’ at output images to design a deeper structure with generating denoised UAV images.

‘Gen’ acquires random noise as well as generates new crack images for resemble real crack data. On other hand, discriminator network ‘Dis’ for distinguish among real crack images as of separated ground and non-ground points sample image and generated crack images produced by ‘Gen’.

Here two players, discriminator network ‘Dis’ and generator network ‘Gen’ are involved where the generator network ‘Gen’ strategy is ‘Prob(Ω)’ denoting the probability set ‘µ_Gen’ on ‘Ω’ and the generator network ‘Gen’ strategy is ‘µ_Gen = Ω →Prob(0,1)’. In this manner by generating synthetic cracks with varied sizes and orientations, GANs assist in learning to identify cracks. Then, the GAN game with the objective function is formulated as given below.

(9) $$L({\mu }_{Gen},{\mu }_{Dis})=E_{GP\sim {\mu }_{ref}{}_{,NEI}\sim {\mu }_{Dis}(GP)}\left[\text{In}NEI\right]+E_{GP\sim {\mu }_{GAN,NEI}\sim {\mu }_{Dis}(GP)}\left[\text{In}1\right]$$

On one hand, Gen reduces above objective role. ‘Dis’ improves above objective task.

Min-max game theory for image denoising is presented in a Generative Adversarial Network (GAN). GAN process is carried by G and D network. For enhanced performance of noisy-to-clean image, generator network obtains noisy crack or monument surface images into clean and denoised images. On other hands, discriminator network differentiate between real crack images from sample images.

Then, a discriminative model ‘Dis’ that fine tunes in an alternative fashion by solving min-max game theory is formulated as given below.

(10) $$\text{DI}\underset{Gen}{\text{=}}\text{min}\underset{Dis}{\max }{E}_{S{I}^G\sim Dis[S{I}^G]}\left[\text{log}\left(\text{1}-Dis(S{I}^G)\right)\right]-E_{S{I}^{RC}\sim Dis[S{I}^{RC}]}\left[\text{log}Dis(S{I}^{RC})\right]$$

From the above equation (10) by solving min-max game theory employing, real clean image ‘SI^RC’, ‘SIG’ is generator clean image, ‘Dis(SI^RC)’, ‘Dis(SI^G) is distribution of real as well as generated clean image. Generative method generates better images in such a way that the distribution of ground points ‘GP’ of sample images ‘SI’.

Outcome of denoised images using min-max game theory is formulated in GAN model. For enhanced image denoising, the generator receives a noisy image as input and produces a cleaned image and discriminator evaluates whether the generated output corresponds to a real clean image conditioned on the same noisy input.

On the other hand, the distributive model differentiates simulated distribution of ground points ‘GP’ of sample images ‘SI’ synthesized via ‘Gen’. This in turn assists in obtaining better UAV image denoised results for higher PSNR. PSNR is a primary metric for denoising image quality because it quantitatively measures how closely the reconstructed (denoised) image matches the original clean reference image at the pixel level. While performing crack detection and heritage monument surface analysis, preserving image structural details andrestraining noise is significant. The estimation of PSNR indicates noise images and provides a direct indication of reconstruction reliability. The perceptual metrics like Structural Similarity Index (SSIM) are sometimes for structural preservation. In proposed research work, SSIM metrics are not considered. The estimation of PSNR attains enhanced image denoising result by maintaining accurate crack detection with minimized information loss. Pseudo code representation of Cloth Simulation Filter and Generative Adversarial Network described by.

As given in the above algorithm, with the Eswaran Temple drone images are subjected to pre-processing employing Cloth Simulation Filter and Generative Adversarial Network model. First, with the presence of ­several cracks are found to be occluded by vegetation, causes defect detection a time consuming process, therefore compromising clear planarity. Therefore, the raw images are subjected to Cloth Simulation Filter for partition ground points. Loud as well as spotless images are mapped by utilizing loss function. Generative Adversarial Network obtains denoised outcomes, therefore improving the overall PSNR.

2.3. Dijkstra graph cut ROI-based segmentation

Image segmentation is one of the most paramount paces as far as image processing is concerned. It plays pivotal part just about in every application, to name a few being medical image analysis, monument crack detection and so on. In this section Dijkstra Graph Cut ROI-based Segmentation algorithm is applied with the objective of segmenting a denoised image into multiple segments so that any region of interest is said to be obtained easily with accurate and precise defect detection results.

The Dijkstra Graph Cut algorithm enhances boundary description in crack regions with the combination of graph-based segmentation and optimal path finding. The identification of optimal path ensures weaknesses of standard graph cut for better detection. Here, Graph Cut identifies the regions belong together and Dijkstra decides the most accurate path with optimal boundaries along the crack. As a result, Dijkstra Graph Cut algorithm is significantly more effective for performing crack detection tasks.

The proposed Dijkstra Graph Cut ROI-based Segmentation algorithm is a way to split a denoised image into different regions of interest (ROIs). The proposed Dijkstra Graph Cut ROI-based Segmentation algorithm works by utilizing the principles of graph cut theory and energy minimization. By creating a graph representation of the denoised image, with pixels representing the nodes and their relationships denoting edges, the algorithm aspires to split the denoised image into foreground and background regions. Using the concept of energy, the algorithm reduces a cost function that summarizes both local and global image features. This model results in comprehensive segmentation results, accurately and precisely describing structural irregularities from the complex background of symmetrical design images. Figure 2 shows the structure of Dijkstra Graph Cut ROI-based Segmentation model.

Figure 2
Structure of Dijkstra graph cut ROI-based segmentation model.

From figure, Dijkstra Graph Cut ROI-based Segmentation algorithm takes an input Denoised Image ‘DI’ consisting of UAV image denoised results. Let us consider a graph ‘G(V,E)’ where each pixel in Denoised Image ‘DI’ is considered a vertex ‘V’ and connection between adjacent pixels is denoted as edges ‘E’ respectively. The input Denoised Image ‘DI’ is denoted as a grid of pixels where each pixel represents a node in the graph. This is mathematically denoted as given below.

(11) $$V=DI\left\{v_{ij}\right\},i=1,2,\mathrm{...},H;j=1,\mathrm{2...},W$$

From the above equation (11) the vertex ‘V’ represents each pixel in Denoised Image ‘DI’ with ‘H’ denoting the height and ‘W’ representing the width respectively. Next, for each Denoised Image ‘DI’, a seed point ‘SP’ is generated.

The proposed Dijkstra Graph Cut ROI-based Segmentation algorithm utilizes the principles of graph cut theory and energy minimization. By creating a graph representation of denoised image, image pixel representation and edge relationships re provides to split the denoised image into foreground and background regions. The foreground and background regions are classified in Dijkstra Graph Cut algorithm to ensure boundary between pixels.

These seed points symbolize as initial pointers for the Dijkstra Graph Cut ROI-based Segmentation process to provide initial information about the region of interest, written as given below.

(12) $$S{P}_i=(x_i,y_i),i=1,2,\mathrm{...},n$$

Following which the strength of the connection between nodes (i.e. edge weights) for each seed points is arrived at on the basis of intensity differences as given below.

(13) $$W_{ij,kl}=\text{exp}-\left(\frac{{(D{I}_{ij}-D{I}_{kI})}^2}{2\sigma D{I}^2}\right)$$

From the above equation results (13) the connection between nodes (i.e. edge weights) results ‘W_ij,kl’ are obtained on the basis of intensity differences ‘DI_ij – DI_kl’. Next, two special nodes, i.e., source (i.e. starting point of denoised image) and sink nodes (i.e. ending point of denoised image) are added to the graph denoting the foreground and background respectively. Moreover the weight or cost of each edge is disclosed by its thickness. The cost of a cut for graph is equal to the associated energy values. Correspondingly, a cut with minimum cost infers energy minimization and leads to optimal segmentation of foreground and background regions. Intelligent defects in monuments are identified. Accordingly, graph-cut segmentation function is applied with energy minimization objective of identifying optimal cut that divides the foreground and background regions, written as given below.

(14) $$En={\displaystyle {\sum }_{i}D{C}_i}(D{I}_i)+{\displaystyle {\sum }_{ij}P{C}_{ij}}(D{I}_i,D{I}_j)$$

From the above equation (14), the energy minimization function ‘En’ results are arrived at based on the data cost (i.e. cost associated with assigning a label to a single node on the basis of its individual features (i.e. pixel intensity) ‘DC_i’ while pair-wise cost ‘PC_ij’ denotes the cost of assigning labels to two neighboring nodes (i.e. homogeneous labels for adjacent nodes with homogeneous features, hence contributing to smoothness in segmentation result). Finally the segmented structural irregularity images (i.e. unusual geometric shapes) are detected within the Denoised Image ‘DI’ utilizing minimum cut function. An ‘x – y’ cut ‘C = (x, y)’ is vertex ­division ‘V’ (‘x ∈ X ’ and ‘y ∈ Y ’). In other words, an ‘x – y’ cut is a division of vertices into two parts, with source node (i.e. starting point) in one part and the sink node (i.e. ending point) in the other part. Then, the ­cut-set ‘DI_C’ of a cut ‘C’ denotes the set of edges that connect source part of the cut to the sink part of the cut and is written as given below.

(15) $$D{I}_c=\{(i,j)\in E:i\in X,j\in Y\}=(X*Y)\cap E$$

(16) $$MC(D{I}_c)=min{\displaystyle {\sum }_{(i,j\in E)}{C}_{ij}}{E}_{ij}$$

From the above equations (15) and (16) by formulating minimum cut ‘$min{\displaystyle {\sum }_{(i,j\in E)}{C}_{ij}}$’ sharp boundaries between different regions in a denoised image are obtained across specific set of edges ‘E’. Finally, ROI-based segmentation or the segmented structural irregularity results are obtained using refined segmentation (i.e. minimum cut results) as given below.

(17) $$Reg=DI\odot MC(D{I}_c)$$

From the above equation (17) the segmented regions results ‘Reg’ are obtained so that any region of interest is said to be obtained accurately and precisely.

The performance of ROI-based segmentation in proposed method accurately handle cracks that are partially obscured by shadows or have low contrast by combining regional localization with local enhancement. The image segmentation approach is carried based on two different processes such as region of interest (ROI) segmentation and image pixel-level segmentation. ROI-based segmentation handles shadows or low contrast cracks by focusing on crack areas rather than the entire image, enhancing contrast before analysis and graph based inferences are used to fill cracks images.

The hyperparameters used in the proposed model were carefully selected to achieve stable training and improved detection performance. Similar to the strategy discussed in Deep Learning based modeling approaches for groundwater storage change, the hyperparameters in this study control the learning behavior, convergence speed, and generalization capability of the network. At proposed framework, main hyperparameters considered such as learning rate is 0.001, batch size is 32, number of epochs, is 100, dropout rate is 0.5 as well as loss function is Cross-Entropy. It was chosen via validation performance as well as training stability. These hyper parameters helped the proposed model achieve improved accuracy and reliable convergence during defect detection and classification tasks. The pseudo code representation of Dijkstra Graph Cut ROI-based Segmentation is given below.

As given in the above algorithm, four steps are formulated for strengthening the boundary portions with image denoised results. First, the denoised image results are subjected to graph representation with each pixel denoting a node and edges connecting adjacent pixels. Second, seed points along with the edge weights are determined via pixel intensity differences. Third, graph-cut segmentation with minimum cut is applied to avoid boundary and region balance problem issue by strengthening the boundary portions with image denoised results. Finally, region of interest is obtained with maximum precision and accuracy.

2.4. Case study

Monument defect detection using UAV Eswaran Temple drone images are simulated by applying Cloth Simulated Filter Adversarial Network and Graph Cut (CSFAN-GR) ROI-based segmentation method. Figure 3 given below shows the sample intact (a) to (d) and crack images (e) to (h).

Figure 3
Eswaran Temple drone images (a)–(d) normal or intact images (e)–(h) crack images.

To perform simulation two intact Eswaran Temple drone sample images and two crack Eswaran Temple drone sample images are used as shown in figure. Figure 4 illustrates hemisphere canopy a 100 m * 100 m square scene consisting of 50 hemisphere canopies and with radius as well as height range put to 5–6 and 8–11 m. On other hand the square canopy a 150 m × 150 m square scene consisting of 70 squares canopies with radius as well as tallness rate is 4–7 and 20–60 m.

Figure 4
Constructed 100 m * 100 m Eswaran Temple drone image scene (a) hemisphere canopy Eswaran Temple drone image scene (b) square canopy Eswaran Temple drone image scene.

Figure 4 given above shows different proportions of Eswaran Temple drone data pits arbitrarily created in hemisphere and square canopies. By splitting the different proportions of Eswaran Temple drone data pits arbitrarily employing hemisphere and square canopies any type of presence of vegetation in the images are said to be identified accurately. Also by employing both hemisphere canopy and square canopy, both curved rounded top portions and flat, square portions are filtered making the primary difference the shape of canopy’s overhead surface assisting in separate ground points from non-ground points. Immaterial image part eradicated with higher precision. Accurate and precise division of ground points is to mapping relationship among noisy and clean images. Figure 5 illustrates association or the relationship between particles position, denoting initial state of intact image (a), initial state of crack image (e), displacement by external force of intact image (b), displacement by external force of crack image (f), displacement by internal force of intact image (c), displacement by internal force of crack image (g) and final state of intact image (d) and crack image (h) respectively.

Figure 5
Cloth simulation filter process.

Figure 5 given above illustrates the cloth simulation filter process for intact and crack image. First, displacement vector are obtained with every pixel under external as well as internal force therefore evolving better image denoised results. This in turn by differentiating between presence of vegetation and structural dissimilarities in sample images aids in improving overall PSNR in an extensive manner. Figure 6 given below shows the results of Dijkstra Graph Cut ROI-based Segmentation results for the corresponding intact and crack denoised images.

Figure 6
Dijkstra Graph Cut ROI-based Segmentation results (a) sample denoised intact image (b) segmented intact image results (c) sample denoised crack image (d) segmented crack image results (e) crack area with highlighted portions.

Figure 6 given above shows the segmented denoised results of sample intact and crack images by applying Dijkstra Graph Cut ROI model. Here first, the denoised sample intact and crack images were segmented into multiple segments with the objective of obtaining even minute region of interest.

By energy concept, local as well as global image features are merged for structural irregularities from complex background of symmetrical design images. These structural irregularities caused by non-uniform distribution of mass is handled using based on Newton’s second law via relationship between particles position (i.e. the sample image position) and internal/external forces. Mass structural irregularity via Dijkstra Graph Cut ROI model was handled for boosting precision.

3. EXPERIMENTAL SETUP AND DISCUSSION

Proposed intelligent detection of defects in monument using UAV with the aid of CSFAN-GR ROI-based segmentation investigated. Simulations were performed in MATLAB on computer Intel(R) by RAM of 32 GB. Parametric different metrics considered for the evaluation of intelligent detection of defects in monument. CSFAN-GR, conventional [¹] and [²] were compared where same Real-time drone images from Eswaran Temple were used for validating the performance metrics.

Photogrammetry-based Digital Twin and PSP_R50 were chosen as baseline methods because they are widely used and well-recognized approaches in structural defect detection and image analysis. The Photogrammetry-based Digital Twin method is effective for structural surface modeling and inspection, while PSP_R50 is a popular deep learning model known for its strong segmentation performance in crack and damage detection tasks.

These methods were selected to provide a meaningful comparison with the proposed approach in terms of defect detection capability, feature extraction, and overall performance. In addition, both methods are commonly used in related studies, making them suitable benchmark models for evaluation.

The authors also recognize that several newer deep learning and transformer-based detectors have recently been proposed for structural damage analysis. However, due to limitations related to computational resources and dataset compatibility, only the most relevant baseline methods were included in this study. Future work will include comparisons with more recent state-of-the-art models to further validate the effectiveness and generalization ability of the proposed method.

3.1. PSNR

Validation of PSNR is discussed in detail. PSNR represents ratio between highest pixels of sample image value. MSE denotes difference among unique corrupting noise mage and noise-free image. The PSNR is mathematically evaluated as given below.

(18) $$PSNR(dB)=10lo{g}_{10}\frac{{255}^2}{MSE}$$

(19) $$MSE=\frac{1}{MN}{\displaystyle \sum {}_{i=1}^M}{{\displaystyle \sum {}_{j=1}^N\left[ac{t}_{i,j}-pre{d}_{ij}\right]}}^2$$

From the above equation (18), ‘PSNR’ is evaluated using mean square error ‘MSE’. Square dissimilarity amid pixel range of actual sample temple image ‘CI_ij’ and detected (crac/intact) sample temple image ‘SI_ij’. PSNR is portrayed in Table 1 using CSFAN-GR and existing. Photogrammetry-based Digital Twin [¹] and PSP_R50 [²].

Table 1
Tabulation of PSNR using CSFAN-GR, Photogrammetry-based Digital Twin [¹] and PSP_R50 [²].

SAMPLE IMAGE SIZE	PSNR (dB)
	CSFAN-GR	PHOTOGRAMMETRY-BASED DIGITAL TWIN [1]	PSP_R50 [2]
8.19	45.72	35.57	30.32
9.05	41.65	32.6	29.21
8.50	47.22	39.4	35.76
9.76	43.09	37.5	32.62
9.07	49.3	40.15	36.54
11.6	45.51	39.42	33.57
9.29	40.32	30.33	29.66
6.25	53.82	44.45	39.62

Figure 7 given above illustrates the comparative analysis of peak signal to noise ratio by taking into consideration different sizes of real-time drone images from Eswaran Temple images as input. Horizontal direction indicated as real-time drone images as well as vertical axis represented as PSNR results. PSNR obtained with separating ground points from non-ground points and mapping relationship between noisy and clean images from the original real-time drone images from Eswaran Temples. For boosting overall clear planarity, Cloth Simulation Filter and Generative Adversarial Network employed. PSNR is enhanced by 23% as well as 37% than [¹] and [²].

Figure 7
PSNR versus sample image size.

3.2. Precision, recall and accuracy

It is presented for intelligent detection of defects in monument.

(20) $$Pre=\frac{TP}{TP+FP}$$

(21) $$Rec=\frac{TP}{TP+FN}$$

(22) $$Acc=\frac{TP+TN}{TP+TN+FP+FN}$$

In (20), (21) and (22), ‘Pre’is precision, ‘Rec’ is recall, ‘Acc’ is accuracy, ‘TP’, ‘TN’ is true positive and negative, ‘FP’and ‘FN’ is false positive as well as negative. Precision, recall as well as accuracy revealed in Table 2.

Table 2
Tabulation of precision, recall and accuracy using CSFAN-GR, Photogrammetry-based Digital Twin [¹] and PSP_R50 [²].

SAMPLE IMAGES	PRECISION (%)			RECALL (%)			ACCURACY
	CSFAN-GR	PHOTOGRAMMETRY-BASED DIGITAL TWIN [1]	PSP_R50 [2]	CSFAN-GR	PHOTOGRAMMETRY-BASED DIGITAL TWIN [1]	PSP_R50 [2]	CSFAN-GR	PHOTOGRAMMETRY-BASED DIGITAL TWIN [1]	PSP_R50 [2]
30	96.29	93.7	92.5	98.11	96.19	95.47	95	91	89.33
60	95.15	91.15	86.75	97.35	87.55	82.35	93.15	83.15	78.35
90	93	89.25	84.35	95	85.35	80.15	90	80.25	75.55
120	92.45	88.35	83.15	93	83.25	78.55	88.55	78.55	73.25
150	90	86.55	81.25	90	80.55	75.35	85.35	75.35	70.15
180	87.15	83.25	78.55	90.15	80.35	75	88.25	78.25	73.25
210	91.35	87.65	82.35	94	84.25	79.25	91.35	81.55	76.55
240	94	90.15	85.55	95.65	85.35	80.55	93	83.55	78.65
270	95.25	91.35	86.25	97	87	82.35	94.55	84.54	79.25
300	97	94.55	89.15	98	88	83.15	95	85.25	80.55

Precision illustrated in Figure 8 for 300 dissimilar crack as well as intact images. For 300 images, TP is obtained as 260, 253 and 250 and FP is 10, 17 and 20 for three methods. Precision using the three methods were observed to be 96.29%, 93.70% and 92.59%. CSFAN-GR provides better outcomes of precision.

Figure 8
Precision versus sample images.

Recall demonstrated in Figure 9 for three different methods, CSFAN-GR, Photogrammetry-based Digital Twin [¹] and PSP_R50 [²]. FN is obtained by 5, 10 and 12 using three methods. Recall is 98.11%, 96.19% and 95.47% for three techniques. CSFAN-GR offers superior outcomes of recall.

Figure 9
Recall versus sample images.

Figure 10 given above shows the accuracy measure for 300 different crack and intact images obtained from real-time drone images from Eswaran Temple using three distinct methods, CSFAN-GR, Photogrammetry-based Digital Twin [¹] and PSP_R50 [²] respectively. With 300 images, accuracy was achieved by 95%, 91% and 89.33% for CSFAN-GR, [¹–²].

Figure 10
Accuracy versus sample images.

Reason for higher precision, recall and accuracy is improved by Dijkstra Graph Cut ROI-based Segmentation algorithm. By applying this algorithm, the image denoised results were employed as input with the intent of addressing structural irregularities in images. With this objective, ROI-based segmentation was performed for segmenting an image into multiple segments to obtain region of interest easily. To achieve this objective minimum cut maximum flow using Dijkstra Graph Cut was applied that in turn avoided boundary and region balance issue by strengthening boundary portions. Precision enhanced via 4%, 10% than [¹–²]. Recall as well as accuracy increased using CSFAN-GR with 11%, 17%, 11% and 18% over [¹–²].

3.3. Case Scenario 3: F1-score

It has compute of forecasting performance for defect identification.

(23) $$F1-score=2\times \frac{Pre\times Rec}{Pre+Rec}$$

It has evaluated by precision as well as recall. F1-score portrayed in Table 3.

Table 3
Tabulation of F1-score using CSFAN-GR, Photogrammetry-based Digital Twin [¹] and PSP_R50 [²].

SAMPLE IMAGES	F1-SCORE (%)
	CSFAN-GR	PHOTOGRAMMETRY-BASED DIGITAL TWIN [1]	PSP_R50 [2]
30	0.97	0.94	0.93
60	0.96	0.89	0.84
90	0.93	0.87	0.82
120	0.92	0.85	0.80
150	0.90	0.83	0.78
180	0.88	0.81	0.76
210	0.92	0.85	0.80
240	0.94	0.87	0.82
270	0.96	0.89	0.84
300	0.97	0.91	0.86

F1-score exposed at Figure 11. This inference is made based on precision as well as recall range. Maximum probable F-score range is 1.0 as well as minimum value is 0. From above simulations performed, highest F-score is attained as 0.97, 0.94 and 0.93 for CSFAN-GR, two existing methods [¹–²]. F1-score of CSFAN-GR enhanced over [¹] and [²].

Figure 11
F1-score versus sample images.

3.4. Case scenario 4: computational complexity

Computational complexity is estimated as the amount of memory consumed for detecting monument defect according to total number of input sample images. It is measured in terms of megabytes (MB) and is evaluated by using following mathematical equation.

(24) $$Computationa{l}_c={\displaystyle {\sum }_{i=1}^n{S}_i}*Mem[DD]$$

From above equation (24), computational complexity ‘Computational_c’ is measured based on ‘S_i’ number of input sample images. Here, ‘Mem[DD]’ specifies memory taken to perform detection of defect in monument that includes both intact and crack images.

The Figure 12 describes the comparison graph for computational complexity versus the number of sample images with the range of 30 to 300 for 10 iterations. The effectiveness of overhead using proposed CSFAN-GR model is verified by comparing with existing Photogrammetry-based Digital Twin [¹] and PSP_R50 [²] respectively. During detection of different crack and intact images, the memory space utilized for defect detection is significantly minimized with the help of three methods. As compared to other methods, CSFAN-GR model attains better results of computational complexity for ROI-based segmentation and defect detection. By performing Cloth Simulation Filter and Generative Adversarial Network in proposed method, intact and crack temple images are separated with minimized complexity. With preprocessed image, Dijkstra Graph Cut ROI-based Segmentation is applied for intelligent monument defect detection. As a result, computational complexity is effectively minimized using proposed CSFAN-GR method by 5% and 11% when compared to existing Photogrammetry-based Digital Twin [¹] and PSP_R50 [²] respectively. Table 4 lists the comparative values of computational complexity.

Figure 12
Outline figure of computational complexity.

Table 4
Comparison values of computational complexity.

SAMPLE IMAGES	COMPUTATIONAL COMPLEXITY (MB)
	CSFAN-GR	PHOTOGRAMMETRY-BASED DIGITAL TWIN [1]	PSP_R50 [2]
30	35	37	43
60	38	42	46
90	45	48	51
120	50	53	55
150	54	57	59
180	58	61	64
210	63	65	68
240	67	69	74
270	72	75	78
300	75	78	82

3.5. Case scenario 5: runtime

The runtime is defined as the amount of time consumed to perform detecting the presence of defects in monuments. It measured in terms of milliseconds (ms). The runtime is mathematically expressed as given below.

(25) $$Runtime={\displaystyle {\sum }_{i=1}^m{S}_i}*t(DD)$$

With the help of above equation (25), runtime is estimated based on number of sample images indicated as ‘S_i’ and time consumed for detecting defects ‘t(DD)’ and the same is tabulated in Table 5.

Table 5
Tabulation for runtime.

SAMPLE IMAGES	RUNTIME (ms)
	CSFAN-GR	PHOTOGRAMMETRY-BASED DIGITAL TWIN [1]	PSP_R50 [2]
30	11	15.5	17
60	11.55	16.85	19.25
90	13.15	19.8	25.35
120	18	25.63	30.15
150	21.35	31.25	36.5
180	25.85	37.6	39.45
210	28.25	42.15	45.25
240	32.5	45.87	48.8
270	36.58	49.2	51.27
300	38.5	51.5	53.5

Figure 13 represents the experimental results of runtime with respect to different number of sample images in the range of 30 to 300 images using proposed and existing methods. Figure provides the comparison result of proposed CSFAN-GR method ith existing methods namely, Photogrammetry-based Digital Twin [¹] and PSP_R50 [²]. As demonstrated in figure, proposed method provides minimum time for achieving detection of defects in monument. The proposed technique carries image preprocessing and segmentation process to enhance detection of defects in monuments using UAV images with minimum runtime. Thus, intact and crack temple images are effectively detected with minimum runtime. As a result, proposed CSFAN-GR method reduces runtime by 30% and 37% when compared to existing methods such as Photogrammetry-based Digital Twin [¹] and PSP_R50 [²] correspondingly.

Figure 13
Experimental result of runtime.

3.6. Ablation study

The ablation study is carried out in CSFAN-GR ROI-based segmentation to find out the contribution of preprocessing and image segmentation process to enhance overall performance of intelligent monument defect detection. At first, Cloth Simulation Filter (CSF) function is carried out to perform efficient ground point’s separation. Then, CSF is combined with the Generative Adversarial Network (GNN) to map the relationship between noisy and clean images for identifying crack temple images. Finally, Dijkstra Graph Cut ROI-based Segmentation is integrated for efficient monument defect detection with higher accuracy. The assessment is carried out to improve the monument defect detection accuracy with minimum time consumption.

Table 6 shows the performance of individual CSF, GAN, and Graph Cut components to detect defects in monuments using UAV images. Cloth Simulation Filter (CSF) function achieves accuracy of 94.5% respectively. After that, CSF with GNN improves accuracy by 95.4 for enhanced defect detection. Finally, proposed CSFAN-GR method attained higher accuracy of 96.5% on performing Dijkstra Graph Cut ROI-based Segmentation process. CSFAN-GR across different samples UAV images of effectiveness is enhanced.

Table 6
Ablation study.

METHODS	ESWARAN TEMPLE DRONE IMAGES DATASET
	DEFECT DETECTION ACCURACY (%)
Cloth Simulation Filter (CSF) function	94.5
CSF function + Generative Adversarial Network (GNN)	95.4
CSF + GNN + Dijkstra Graph Cut ROI-based Segmentation	96.5

3.7. Paired t-test results

CSFAN-GR performs superior than existing Photogrammetry Digital Twin and PSP_R50 methods. CSFAN-GR achieved the highest average performance and maintained consistent results throughout the evaluation. Paired t-test outcome improvements are statistically significant with p-values below 0.001. Overall, the study proves that CSFAN-GR is a more reliable and effective method compared to the other approaches.

Table 7 and Figure 14 shows the Paired t-test results for proposed CSFAN-GR method. In this figure, x-axis specifies three methods and y-axis represents performance metrics values using proposed and existing methods.

Table 7
Paired t-test results.

COMPARISON	MEAN (%)	MEAN (%)	Δ MEAN	T-STATISTIC	P-VALUE	SIG. LEVEL	SIGNIFICANT?
CSFAN-GR vs Photogrammetry Digital Twin [1]	91.42	82.14	+9.28	15.7480	7.39e-08	***	Yes ✓
CSFAN-GR vs PSP_R50 [2]	91.42	77.49	+13.93	15.0652	1.09e-07	***	Yes ✓
Photogrammetry Digital Twin [1] vs PSP_R50 [2]	82.14	77.49	+4.66	13.7300	2.43e-07	***	Yes ✓

Figure 14
Outline of paired t-test results.

4. CONCLUSION

We investigated intelligent detection of defects in monuments using UAV employing Cloth Simulation Filter and Dijkstra Graph Cut ROI-based Segmentation. Pre-processing as well as segmentation executed for intelligent detection of defects in monuments. To deal with filter as well as noise, Cloth Simulation Filter function and Generative Adversarial Network was utilized via min-max game theory with maximum PSNR as well as minimum time. Dijkstra Graph Cut ROI-based Segmentation is designed for accurate detection of defects in monuments. Outcome of CSFAN-GR was finer performance with several factors over other schemes.

Although, proposed method achieved good overall performance, a few failure cases were observed during testing. In some images, normal surface regions containing shadows, stains, rough textures, or weathering effects were mistakenly detected as defects. On the other hand, a few very small or unclear cracks were not detected correctly, especially when the defect boundaries were weak or blended with the surrounding surface.

These issues mainly occurred because some natural patterns on heritage structures look very similar to actual damage. Changes in lighting conditions, image noise, low contrast, and UAV image quality also affected the detection performance in certain cases. In addition, defects captured from longer distances sometimes lacked enough visual detail for accurate identification. However, such cases were limited compared to the overall performance of the model. Future work will focus on improving the training dataset and using more advanced feature learning techniques to reduce false detections and improve the identification of small or complex defects.

5. DATA AVAILABILITY

Data will be made on available on reasonable request from the authors.

6. BIBLIOGRAPHY

• ^[1] YIĞIT, A.Y., UYSAL, M., “Detection of cracks in cultural heritage buildings using UAV photogrammetry-based digital twin”, ACM Journal on Computing and Cultural Heritage, v. 18, n. 1, pp. 1–17, 2025. doi: https://doi.org/10.1145/3703634.
  » https://doi.org/10.1145/3703634
• ^[2] ALEXAKIS, E., DELEGOU, E.T., MAVREPIS, P., et al, “A novel application of deep learning approach over IRT images for the automated detection of rising damp on historical masonries”, Case Studies in Construction Materials, v. 20, pp. e02889, 2024. doi: https://doi.org/10.1016/j.cscm.2024.e02889.
  » https://doi.org/10.1016/j.cscm.2024.e02889
• ^[3] ZHEN, L., YANG, Z., LAPORTE, G., et al, “Unmanned aerial vehicle inspection routing and scheduling for engineering management”, Engineering (Beijing), v. 36, pp. 223–239, 2024. doi: https://doi.org/10.1016/j.eng.2023.10.014.
  » https://doi.org/10.1016/j.eng.2023.10.014
• ^[4] PRUNELLA, M., SCARDIGNO, R.M., BUONGIORNO, D., et al, “Deep learning for automatic vision-based recognition of industrial surface defects: a survey”, IEEE Access : Practical Innovations, Open Solutions, v. 11, pp. 43370–43423, 2023. doi: https://doi.org/10.1109/ACCESS.2023.3271748.
  » https://doi.org/10.1109/ACCESS.2023.3271748
• ^[5] CAI, S., ZHANG, W., LIANG, X., et al, “Filtering airborne LiDAR data through complementary cloth simulation and progressive TIN densification filters”, Remote Sensing (Basel), v. 11, n. 9, pp. 1037, 2019. doi: https://doi.org/10.3390/rs11091037.
  » https://doi.org/10.3390/rs11091037
• ^[6] KARIMI, N., MISHRA, M., LOURENÇO, P.B., “Deep learning-based automated tile defect detection system for Portuguese cultural heritage buildings”, Journal of Cultural Heritage, v. 68, pp. 86–98, 2024. https://doi.org/10.1016/j.culher.2024.05.009.
  » https://doi.org/10.1016/j.culher.2024.05.009
• ^[7] MAKHANOVA, Z., BEISSENOVA, G., MADIYAROVA, A., et al, “A deep residual network designed for detecting cracks in buildings of historical significance”, International Journal of Advanced Computer Science and Applications, v. 15, n. 5, pp. 583, 2024. doi: https://doi.org/10.14569/IJACSA.2024.0150558.
  » https://doi.org/10.14569/IJACSA.2024.0150558
• ^[8] MOKHTAR, M., MAT NOR, A.H., KAAMIN, M., et al, “Inspection analysis of historic buildings with aid of unmanned aerial vehicle (UAV)”, International Journal of Innovative Technology and Exploring Engineering, v. 8, pp. 317–320, 2019.
• ^[9] KAAMIN, M., IDRIS, N.A., BUKARI, S.M., et al, “Visual inspection of historical buildings using micro UAV”, In: MATEC Web of Conferences, 2017. doi: https://doi.org/10.1051/matecconf/201710307003.
  » https://doi.org/10.1051/matecconf/201710307003
• ^[10] SHALABY, Y.M., BADAWY, M., EBRAHIM, G.A., et al, “Condition assessment of concrete structures using automated crack detection method for different concrete surface types based on image processing”, Discover Civil Engineering, v. 1, n. 1, pp. 81, 2024. doi: https://doi.org/10.1007/s44290-024-00089-5.
  » https://doi.org/10.1007/s44290-024-00089-5
• ^[11] FU, X., ANGKAWISITTPAN, N., “Detecting surface defects of heritage buildings based on deep learning.”, Journal of Intelligent Systems, v. 33, n. 1, pp. 20230048, 2024. doi: https://doi.org/10.1515/jisys-2023-0048.
  » https://doi.org/10.1515/jisys-2023-0048
• ^[12] ABOULHOSN, Z., MUSAMIH, A., SALAH, K., et al, “Detection of manufacturing defects in steel using deep learning with explainable artificial intelligence”, IEEE Access: Practical Innovations, Open Solutions, v. 12, pp. 99240–99257, 2024. doi: https://doi.org/10.1109/ACCESS.2024.3430113.
  » https://doi.org/10.1109/ACCESS.2024.3430113
• ^[13] FEROZ, S., ABU DABOUS, S., “UAV-based remote sensing applications for bridge condition assessment”, Remote Sensing (Basel), v. 13, n. 9, pp. 1809, 2021. doi: https://doi.org/10.3390/rs13091809.
  » https://doi.org/10.3390/rs13091809
• ^[14] HOU, M., HUO, D., YANG, Y., et al, “Using Mask R-CNN to rapidly detect gold foil shedding of stone cultural heritage”, Heritage Science, v. 12, n. 1, pp. 46, 2024. doi: https://doi.org/10.1186/s40494-024-01158-9.
  » https://doi.org/10.1186/s40494-024-01158-9
• ^[15] LO GIUDICE, M., MARIANI, F., CALIANO, G., et al, “Deep learning for the detection and classification of adhesion defects in antique plaster layers”, Journal of Cultural Heritage, v. 69, pp. 78–85, 2024. doi: https://doi.org/10.1016/j.culher.2024.07.012.
  » https://doi.org/10.1016/j.culher.2024.07.012
• ^[16] PERUMAL, R., VENKATACHALAM, S.B., “Non invasive decay analysis of monument using deep learning techniques”, TS. Traitement du Signal, v. 40, n. 2, pp. 639–646, 2023. doi: https://doi.org/10.18280/ts.400222.
  » https://doi.org/10.18280/ts.400222
• ^[17] KWON, D., YU, J., “Ravit-ae: unsupervised anomaly detection for intelligent cultural heritage monitoring using region-attentive vit autoencoder”, IEEE Access : Practical Innovations, Open Solutions, v. 12, pp. 180767–180780, 2024. doi: https://doi.org/10.1109/ACCESS.2024.3509988.
  » https://doi.org/10.1109/ACCESS.2024.3509988
• ^[18] ZHANG, F., SAEED, N., SADEGHIAN, P., “Deep learning in fault detection and diagnosis of building HVAC systems: a systematic review with meta analysis”, Energy and AI, v. 12, pp. 100235, 2023. doi: https://doi.org/10.1016/j.egyai.2023.100235.
  » https://doi.org/10.1016/j.egyai.2023.100235
• ^[19] ZHANG, Q., CHEN, S., WU, Y., et al, “Improved U-net network for asphalt pavement crack detection”, PLoS One, v. 19, n. 5, pp. e0300679, 2024. doi: https://doi.org/10.1371/journal.pone.0300679. PubMed PMID: 38820536.
  » https://doi.org/10.1371/journal.pone.0300679
• ^[20] DESHPANDE, S., VENUGOPAL, V., KUMAR, M., et al, “Deep learning-based image segmentation for defect detection in additive manufacturing: an overview”, International Journal of Advanced Manufacturing Technology, v. 134, n. 5, pp. 2081–2105, 2024. doi: https://doi.org/10.1007/s00170-024-14191-6.
  » https://doi.org/10.1007/s00170-024-14191-6
• ^[21] TZORTZIS, I.N., RALLIS, I., MAKANTASIS, K., et al, “Automatic inspection of cultural monuments using deep and tensor-based learning on hyperspectral imagery”, In: 2022 IEEE International Conference on Image Processing (ICIP), 2022. doi: https://doi.org/10.1109/ICIP46576.2022.9897527.
  » https://doi.org/10.1109/ICIP46576.2022.9897527
• ^[22] MA, Y., YIN, J., HUANG, F., et al, “Surface defect inspection of industrial products with object detection deep networks: a systematic review”, Artificial Intelligence Review, v. 57, n. 12, pp. 333, 2024. doi: https://doi.org/10.1007/s10462-024-10956-3.
  » https://doi.org/10.1007/s10462-024-10956-3
• ^[23] WANG, X., JIANG, G., HU, J., et al, “Automatic defect detection in infrared thermal images of ancient polyptychs using channel-attention Faster R-CNN”, Heritage Science, 2024. doi: https://doi.org/10.1186/s40494-024-01441-9.
  » https://doi.org/10.1186/s40494-024-01441-9
• ^[24] CASILLO, M., COLACE, F., GAETA, R., et al, “An IoT-based framework for protecting historical buildings”, Evolutionary Intelligence, v. 17, n. 5-6, pp. 3815–3831, 2024. doi: https://doi.org/10.1007/s12065-024-00959-y.
  » https://doi.org/10.1007/s12065-024-00959-y
• ^[25] CARDELLICCHIO, A., RUGGIERI, S., NETTIS, A., et al, “Physical interpretation of machine learning-based recognition of defects for the risk management of existing bridge heritage”, Engineering Failure Analysis, v. 149, pp. 107237, 2023. doi: https://doi.org/10.1016/j.engfailanal.2023.107237.
  » https://doi.org/10.1016/j.engfailanal.2023.107237
• ^[26] HAQ, M.A., RAHAMAN, G., BARAL, P., et al, “Deep learning based supervised image classification using UAV images for forest areas classification”, Photonirvachak (Dehra Dun), v. 49, n. 3, pp. 601–606, 2021. doi: https://doi.org/10.1007/s12524-020-01231-3.
  » https://doi.org/10.1007/s12524-020-01231-3
• ^[27] ANUL HAQ, M., “CNN based automated weed detection system using UAV imagery”, Computer Systems Science and Engineering, v. 42, n. 2, pp. 837–849, 2022. doi: https://doi.org/10.32604/csse.2022.023016.
  » https://doi.org/10.32604/csse.2022.023016
• ^[28] ANUL HAQ, M., “Planetscope nanosatellites image classification using machine learning”, Computer Systems Science and Engineering, v. 42, n. 3, pp. 1031–1046, 2022. doi: https://doi.org/10.32604/csse.2022.023221.
  » https://doi.org/10.32604/csse.2022.023221

Corresponding Author:

Prema Rajendhiran, [e-mail], [e-mail], [e-mail]

Publication Dates

• Publication in this collection
  21 Sept 2026
• Date of issue
  2026

History

• Received
  26 Jan 2026
• Accepted
  06 July 2026