A Review of Artificial Intelligence as a Tool for Damage Detection in Paintings: Challenges and Limitations for Contemporary Paintings

Abstract

The degradation of works of art constitutes a significant problem for the preservation of cultural heritage. In the case of paintings, the observed alterations can be physical, chemical, or visual, affecting both the integrity and appearance of the artworks. Degradation compromises the authenticity, aesthetic legibility, and historical value of paintings, making the early monitoring of such issues, as well as the development of appropriate conservation and restoration strategies, essential. For an effective approach, the characterisation of the materials and techniques used by the artist, as well as the degradation processes inherent in the materials used, proves to be crucial. In this context, the application of artificial intelligence (AI) emerges as a non-invasive solution capable of detecting and predicting degradation in works of art. This bibliographic review aims to explore existing studies in this field in depth, with special attention to contemporary paintings considered as case studies. The methodology involved a systematic review of peer-reviewed studies, theses, and interdisciplinary databases, using keywords related to the topic under investigation (e.g., “degradation detection,” “artificial intelligence,” “craquelure segmentation”). The results indicate that artificial intelligence enables the early detection of degradations that may not yet be visible to the naked eye while also improving objectivity and consistency in the analysis of complex and irregular patterns typical of paintings. It became evident that there is a significant gap in the literature, regarding studies addressing the potential of AI for degradation detection specific to contemporary paintings. However, these could be a valuable case study given their potential material and technical heterogeneity, as well as their differences from traditional easel paintings.

1. Introduction

The degradation of works of art constitutes a major challenge for the preservation of cultural heritage at multiple levels. In the case of paintings, it is essential to consider their inherently stratified structure, composed of multiple layers with distinct and specific physical and chemical properties that interact with one another [1]. These internal interactions, combined with external factors such as relative humidity and temperature, contribute to the development of various forms of deterioration [2].

Within this framework, degradation phenomena in paintings can be broadly classified into three main categories [3]: (1) chemical damage, resulting from the natural chemical evolution of the materials that constitute the painting and their mutual interactions [4]; (2) mechanical damage, associated with alterations in internal and external forces that lead to deformation of the artwork, often induced by fluctuations in temperature and relative humidity, as well as by impacts or vibrations [3]; and (3) biological damage, which occurs when paintings are affected by biological agents that exploit the organic materials of the artwork as a source of nutrients [5]. As such, early identification and the application of appropriate conservation strategies are essential.

The most frequently observed types of degradation in paintings include the following:

• Colour alterations: caused by photodegradation processes, leading to damage such as fading, darkening, or blanching [6].
• Craquelure: networks of fine cracks resulting from the drying of paints and mechanical stresses between the different layers of the painting. They can also be caused by impact, improper storage conditions, or excessive stretching of the canvas (Figure 1) [7].
• Support deformations: deformations of the painting’s structure caused by elongation or shrinkage, which weaken the overall integrity of the artwork (Figure 2) [8].

The observed alterations result from multiple contributing factors, including the materials and artistic techniques used by the painter (inherent vice), the environmental conditions to which the work is exposed, and improper handling or conservation [2]. A detailed study of degradation allows for early identification of vulnerable areas, allows for the anticipation of potential damage, and supports the planning of specific conservation strategies for each work.

The study of paintings through non-destructive examination and analytical methods allows for the acquisition of information about the artwork without causing damage, enabling the characterisation of the materials used and the identification of the pictorial techniques employed [9].

Following this approach, the combination of non-destructive examination and analytical techniques (e.g., X-radiography, ultraviolet imaging, and infrared reflectography) with the innovation of artificial intelligence methods can support both the detection of degradation phenomena and the material characterisation of artworks by utilising digital images, high-resolution photographs, or hyperspectral data. The automation of these processes contributes to reducing the time and workload required for analysis, while simultaneously increasing efficiency and consistency [10,11].

Additionally, artificial-intelligence-based methods enable the successful identification and classification of pigments and other materials used in paintings by leveraging data obtained from other techniques, such as hyperspectral imaging [12,13].

Another advantage of artificial intelligence is its ability to enable the virtual reconstruction of paintings, restoring them as close as possible to their original appearance. Examples include the study by Sizyakin et al. [14], which employed deep learning to inpaint craquelure patterns, and the study by Parvathaneni et al. [15], which used areas protected by the frame to regenerate the original colours in Van Gogh’s paintings.

This literature review aims to analyse the advantages and limitations of using artificial intelligence for the detection of degradation in paintings. It begins with a general overview of currently employed strategies, followed by a discussion on its application on contemporary painting. The review seeks to explore the potential of artificial intelligence for the analysis of this specific type of artworks.

The validation of artificial-intelligence-based methodologies applied to conservation and restoration requires their verification in real-world contexts in order to assess the effectiveness and robustness of the models beyond controlled environments. In this regard, case studies play a fundamental role in identifying practical limitations, computational biases, and methodological shortcomings that may not be evident under experimental conditions, thereby enabling subsequent model refinement.

Contemporary painting thus provides a methodologically suitable framework due to the inherent challenges it presents. Contemporary paintings were chosen because of their composition of diverse, sometimes unconventional materials that deviate from traditional easel painting norms (such as the use of paint binders other than oil) [16], as well as their heterogeneous techniques, which result in complex stratigraphies [17].

In summary, the application of artificial intelligence helps reduce subjectivity in assessing the conservation condition and supports data-driven decision-making based on analytical and computational evidence, representing an advantage in the field of conservation and restoration, with multiple benefits that will be analysed and discussed in this review.

2. Methodology

This review was guided by the following research question: how can artificial intelligence be applied to the detection of degradation phenomena in contemporary paintings, considering their structural, material, and technical differences in comparison to traditional easel paintings? Additionally, this study aims to identify the main methodological approaches reported in the literature, as well as their advantages and limitations.

With the aim of analysing the use of artificial intelligence in the detection of degradation in paintings, the methodology of this review was based on the consultation of relevant and interdisciplinary scientific databases, covering the fields of conservation and restoration, materials science, and artificial intelligence, with particular emphasis on peer-reviewed research. The main databases consulted were Web of Science, Scopus, SpringerLink, and Google Scholar. In addition, academic and institutional repositories were consulted to access theses and dissertations, which were used primarily for contextualisation and thematic framing. To complement the research, artificial intelligence tools such as ChatGPT ((OpenAI, San Francisco, CA, USA, https://chat.openai.com/, last accessed 14 May 2026) and Perplexity (Perplexity AI, Inc., San Francisco, CA, USA, https://www.perplexity.ai/, last accessed 14 May 2026) were also employed.

These artificial-intelligence-based tools were employed in a supportive manner, namely to organise and summarise information and to assist in improving the clarity and linguistic quality of the text. However, all sources and information included in this review were independently consulted, analysed, and verified without the use of artificial intelligence tools, ensuring the scientific reliability and relevance of the material considered.

The bibliographic search was conducted using combinations of keywords related to the research topic and object of study, namely artificial intelligence, painting degradation, machine learning, contemporary painting, and conservation. The selection criteria included publication date, prioritising recent studies, thematic relevance, and methodological suitability for the scope of the review. Furthermore, it should be noted that although the focus of this study is on contemporary painting, with a case study centred on Paula Rego, the selection of sources for in-depth analysis was not restricted by geographical criteria. Relevant studies were considered regardless of their geographical context, based solely on their relevance to the research topic and whether they were published in, or accessible through, international databases.

The review process was developed in successive and structured phases, including the identification of relevant research—namely studies combining artificial intelligence methods applied to the detection of degradation, with approaches that may be adaptable to contemporary painting—followed by a preliminary selection of the most relevant articles based on titles and abstracts, according to the defined selection criteria.

Subsequently, a thematic organisation of the selected studies was carried out. This approach enabled the identification of key research directions and the selection of the most relevant contributions, guiding the focus of this review towards artificial intelligence applications in degradation detection within the context of contemporary painting.

Furthermore, this process informed the selection of the key themes explored in Section 3 and Section 4.

3. AI in Conservation

Artificial intelligence emerges as an innovative tool based on computational techniques that enable the execution of tasks that would normally require human intelligence, including pattern recognition and complex data analysis [18]. According to Wang et al. [19], artificial intelligence can be defined as “adaptation with insufficient knowledge and resources,” reflecting these systems’ ability to learn and adapt even when information is incomplete or limited [19].

In the context of cultural heritage conservation, AI has been progressively applied to the analysis of artworks [20]. Van Vijle et al., 2025 [10], organise the use of AI in cultural heritage into five main areas: enhancement of scientific images, pigment analysis, damage detection, damage prediction, and virtual restoration [10]. However, it is important to note that the five categories outlined by Van Vijle et al. [10] are a guiding framework and should not be understood as mutually exclusive, as significant overlaps exist between them For example, virtual restoration may involve damage detection [21], while damage prediction, depending on the type of degradation being addressed, may require pigment analysis [22]. AI applications have enabled significant advances, offering advantages in terms of speed, the ability to process large volumes of data, and objectivity in information analysis [11].

Artificial intelligence (AI) has proven to be a valuable tool in enhancing scientific images. It is capable of improving the quality of images acquired through various analytical techniques, removing noise, enhancing contrast, correcting distortions, and generally facilitating the identification of details that would be invisible to the naked eye. Van Vijle et al. [10] demonstrated that neural networks can enhance hyperspectral images of pigments, enabling the detection of microcracks and invisible layers [10].

In pigment analysis, AI uses spectral patterns derived from analytical data, such as XRF or Raman spectroscopy, to automatically identify materials and pigment compositions, reducing reliance on manual or invasive analyses. Chen et al. [23] applied convolutional neural networks to distinguish modern and historical pigments in paintings by Amadeo de Souza-Cardoso with high accuracy.

Computer vision models are also instrumental in damage detection, analysing digital images to identify and quantify both visible and invisible degradation. This includes fissures, flaking, and stains. One relevant example of advanced deep learning applied to anomaly detection is the DRAEM model [24], originally developed for surface anomaly detection in images. This model has since been applied to artworks including to X-ray pictures of historical paintings to highlight visual anomalies [25].

Beyond detection, AI models can also be used to predict future degradation based on existing data, such as environmental conditions and material composition. Predictive modelling approaches, including regression algorithms within a machine learning framework, have been identified as a promising direction for estimating future damage risks in artworks and heritage object [10].

Finally, AI plays a crucial role in virtual restoration, enabling the simulation of interventions or the digital reconstruction of lost or severely degraded areas without physical contact. Generative neural networks can produce digitally restored images, testing hypotheses of colour, texture, and pattern reconstruction. Parvathaneni et al. [15] applied an artificial intelligence technique to digitally reconstruct the original colours of Van Gogh’s paintings, using areas preserved under the frame as references for the original coloration [15].

Additionally, several studies have demonstrated the potential of AI to support conservation decision-making and preventive conservation, providing objective insights into the condition of artworks and informing intervention strategies. These findings highlight the growing and consolidated adoption of AI in cultural heritage practice [26,27].

3.1. Damage Detection

Degradation in paintings encompasses a range of changes that occur over time and may compromise both the structural integrity and the visual readability of the artwork. As previously discussed, these alterations can be of a chemical or mechanical nature, affecting the materials that constitute the painting and their interactions [3,4]. Early detection of degradation phenomena, together with systematic environmental monitoring, is therefore crucial to ensure the long-term preservation of artworks [28]. Among the most frequently observed types of damage are cracks, flaking, paint losses, colour alterations, stains, and support deformations [6,7,8]. Monitoring these degradation processes constitutes a fundamental component of preventive conservation, as it enables conservation interventions to be planned and prioritised based on scientific evidence regarding the condition of the artwork [28].

Traditionally, damage detection in paintings relies on visual inspection performed by trained experts, often complemented by scientific analytical techniques such as Raman spectroscopy, X-ray fluorescence (XRF), Fourier-transform infrared spectroscopy (FTIR), infrared reflectography, or thermography [29]. Although these approaches are well established and effective, they present several limitations, including observer subjectivity, difficulties in identifying subtle or early-stage degradation patterns, and reduced efficiency when analysing large volumes of data or extensive collections [11].

Within this context, the application of artificial intelligence (AI) to degradation detection has emerged as an innovative and transformative approach. Machine learning and deep learning techniques, as subfields of AI, enable the automated analysis of visual and spectral data, supporting tasks such as the classification of damaged versus intact areas, the segmentation of cracks, stains, or losses, and the identification of complex degradation patterns that may escape conventional visual analysis [10]. These methods offer significant advantages over traditional approaches, notably increased processing speed, the capacity to handle large datasets, and a reduction in subjectivity in condition assessment [10,11,30].

Despite these advantages, the application of AI in the detection of painting degradation also presents important challenges. These include the scarcity of annotated and standardised datasets, which limits the generalisation of models across different artistic techniques, materials, and historical contexts, as well as the continued need for expert validation to ensure the correct interpretation of results [10,11]. Consequently, the implementation of AI-based methods in conservation practice remains an evolving field, with substantial potential for further methodological development and research [10,11].

Within this framework, the present literature review focuses specifically on the detection of craquelure, paint losses, and colour alterations, given their high prevalence and their critical relevance in assessing the conservation state of paintings.

3.1.1. Generalist Damage Detection Models

This section aims to explore four different approaches to damage detection, identifying methodological similarities, as well as their strengths and limitations, rather than providing an exhaustive description of individual studies. In this way, it seeks to critically evaluate and compare generalist AI-based damage detection methods.

In this section, we will address degradation detection methods used in a more generalist way, based on the analysis of articles dedicated to general damage detection, without predefined categories. These studies primarily aim to identify visual or structural anomalies, detecting any pattern that deviates from the expected and “normal” behaviour of the painting, regardless of the type of damage. The priority of these models is robustness across a wide spectrum of damages, rather than detailed classification, which will be discussed later in this review, in the chapter on craquelure and lacunae detection. Although these studies are more general, they all help reduce subjectivity, accelerate degradation detection and diagnostic processes, and systematically monitor paintings. These approaches are particularly useful when pathologies coexist or overlap, or when the type of damage is not predefined.

Moradi et al. [29] developed an approach capable of identifying superficial and internal defects in paintings and other artworks through the integration of spatiotemporal information and infrared thermography. Detecting heating and cooling patterns over time allowed for the identification of discontinuities or damages in the works. Additionally, the study developed a practical tool to virtually reconstruct the affected regions.

This study employed a spatiotemporal deep neural network that analysed temporal variations, associating them with the spatial information of preprocessed thermal images. It thus combined the extraction of spatiotemporal features with damage reconstruction, beyond simple detection. The network results demonstrated higher sensitivity in detecting subtle damages, greater accuracy compared to traditional thermographic analysis methods, and good performance across different temperature and material contexts.

With the same goal of identifying damages and degradations in paintings, but using X-ray images, Mezina et al. [25] developed an unsupervised deep learning technique based on convolutional autoencoders. In this architecture, convolutional neural networks are integrated into autoencoders with two main components: compression (encoders) and reconstruction of the original image (decoders). The model learns to reconstruct the painting’s “normal” appearance and interprets poorly reconstructed or anomalous areas as damaged zones.

This technique proved useful for degradation detection, showing favourable results in locating degraded regions and highlighting subtle damages visible only in X-rays. It also demonstrated robustness across diverse paintings and applicability in real contexts; however, it has limitations in precise damage identification, requiring expert validation for final interpretation.

Comparing these two studies, it can be observed that both rely on reconstruction-based architectures for damage identification. However, their approaches differ substantially in terms of input data and learning strategy. Moradi et al. [29] combine temporal dynamics with thermal imaging within a supervised framework, whereas Mezina et al. [25] adopt an unsupervised model based on X-ray structural information. Thus, although both methods demonstrate strong performance in damage detection, robustness, and sensitivity to subtle defects, they also present important limitations in conservation contexts, particularly the need for expert validation when interpreting results, especially in cases where degradation patterns are ambiguous or particularly subtle.

Wu et al. [31] proposed an automatic method using a lightweight neural network to detect damages in grotto murals, such as cracks, detachments, and lacunae. Based on the YOLOv5 model, several improvements were added: a Ghost Conv for lightweight feature extraction, Double Attention Mechanisms to enhance feature extraction and accelerate convergence, and a Bi-directional Weighted Feature Pyramid to combine multi-scale information and detect damages of various sizes.

This study showed promising results for conservation and restoration, achieving a model simplification of approximately 34.4% and an improvement in real-time performance by around 53%. The accuracy reached 64.7%, suggesting that the model provides a useful alert for degradation detection but does not guarantee coverage of all degradations, especially the more subtle ones. Nevertheless, the lightweight design facilitates practical implementation in sites under conservation and restoration treatments.

Garcia-Moreno et al. [32] developed the ARTDET project, an automatic AI-based tool for detecting deteriorations in easel paintings, identifying and segmenting damaged areas at the pixel level, including lacunae and stucco restorations. The tool uses a pre-trained Mask R-CNN deep learning model and a dataset of high-resolution paintings manually annotated under controlled imaging conditions.

The model’s results were favourable. Automated deterioration detection achieved an average recall of 80.4%, and identified areas had an average confidence score of 99%. Nevertheless, detection was limited to visible and traditionally observed damages, such as lacunae or old restorations; therefore, not all damages were detected, making the identification of less superficial damages challenging.

Although these two studies employ substantially different methods, they reflect a broader pattern in AI-based conservation research, revealing a tension between developing models that are fast and efficient—and therefore more suitable for real-world conservation and restoration settings—and models that are highly precise in damage detection. While lightweight models prioritise real-time applicability and on-site deployment, more complex architectures achieve higher accuracy and finer segmentation under controlled conditions. This dichotomy also highlights the importance of high-quality and well-curated datasets, as well as the continued need for expert interpretation in validating AI-generated results.

In reflecting on the studies presented, it is clear that artificial intelligence has had a positive impact in reducing subjectivity in degradation detection. Moreover, generalist damage detection models have proven to have potential in complex contexts where multiple types of deterioration may coexist. However, this reduction in subjectivity does not eliminate uncertainty, as the performance of these models remains strongly dependent on the data on which they are trained and may therefore reproduce systematic errors and biases. These limitations are not always immediately identifiable, particularly when models perform well on benchmark datasets but fail under more complex or variable conservation conditions. Furthermore, even when AI systems successfully identify potential areas of damage, their outputs do not replace material interpretation, which remains essential in conservation and restoration practice for understanding the nature, causes, and significance of degradation phenomena. Finally, it is important to acknowledge additional limitations, including the need for expert validation and the challenges associated with artworks produced using heterogeneous materials and techniques, as is often the case in contemporary art.

The different approaches allow each technique to be adapted to specific contexts. For example, infrared thermography is effective in detecting internal defects, whereas X-ray imaging is suitable for identifying overlaps or damages invisible to the naked eye. Lightweight models, such as those proposed by Wu et al. [31], are particularly valuable for practical implementation in conservation and restoration settings. Furthermore, it should be emphasised that the generalisation achieved in damage detection relies heavily on the diversity of training data and the complexity of the materials and techniques employed by the artists.

Another important aspect to consider concerns the nature and quality of the datasets used for training and evaluation of the models. Regarding the datasets employed in the reviewed studies, Moradi et al. [29] use paintings and artworks analysed through time-dependent infrared thermography, Mezina et al. [25] rely on X-ray images of paintings, Wu et al. [31] focus on historical grotto murals, and Garcia-Moreno et al. [32] use easel paintings acquired under controlled imaging conditions. In most cases, these datasets are limited in size and constructed under controlled or highly curated conditions. This raises concerns regarding their representativeness of real conservation and restoration scenarios, where imaging conditions, material heterogeneity, and degradation patterns are significantly more variable. Consequently, although these datasets are valuable for model development and benchmarking, they represent only a limited subset of the complexity observed in real-world degradation phenomena.

Looking ahead, the application of models like those described will require the development of hybrid approaches that integrate multiple techniques, maximising the available information to produce more comprehensive and reliable results. Future specialisation in specific types of degradation (as will be discussed in the following section) will be essential for precise damage detection.

While these models are robust and effective for general degradation detection, they operate broadly without focusing on any specific type of damage. The detection of complex phenomena such as craquelure, however, requires specialised analysis due to its structural intricacy, high prevalence, and unique patterns that function as a “fingerprint” of the painting. The following section will therefore focus on craquelure detection, enabling the exploration of more detailed techniques specifically adapted to this type of degradation.

Table 1. Synthesis of the studies addressed regarding generalist degradation detection models.

Study	Moradi et al. 2022 [29]	Mezina et al. 2025 [25]	Wu et al. 2022 [31]	Garcia Moreno et al. 2024 [32]
Objective/focus	General detection of superficial and internal defects in paintings and artworks, without predefined damage categories.	Identification of general damages and degradations in paintings through the detection of structural anomalies not visible to the naked eye.	Automatic detection of damages in grotto murals (cracks, detachments, and lacunae), with a focus on practical implementation.	Automatic detection and pixel-level segmentation of visible deteriorations in easel paintings (lacunae and stucco restorations).
Method/model	Spatiotemporal deep neural network applied to infrared thermography; analysis of heating and cooling patterns over time; virtual reconstruction of damaged regions.	Unsupervised deep learning approach based on convolutional autoencoders (CNNs); learning of the painting’s “normal” appearance and interpretation of poorly reconstructed regions as damaged areas.	Lightweight neural network based on YOLOv5; Ghost Convolution; double attention mechanisms; bi-directional weighted feature pyramid for multi-scale damage detection.	Pre-trained Mask R-CNN deep learning model; pixel-level segmentation; training on high-resolution, manually annotated datasets.
Data/Case-study	Paintings and artworks analysed using time-dependent infrared thermography.	X-ray images of paintings.	Historical grotto murals.	Easel paintings acquired under controlled imaging conditions.
Contributions	Higher sensitivity to subtle damages; improved accuracy compared to traditional thermographic methods; robustness across different temperature and material contexts; integration of damage detection and reconstruction.	Effective localisation of degraded regions; highlighting of subtle damages visible only in X-ray images; good generalisation across diverse paintings; limitations in precise damage identification, requiring expert validation.	Model simplification of ~34.4%; real-time performance improvement of ~53%; accuracy of 64.7%; useful as a degradation alert system, though less effective for subtle damages.	Average recall of 80.4%; average confidence score of 99% for detected areas; effective for visible and traditionally recognised damages; limited capability for detecting less superficial or non-visible degradations.

summarises the studies analysed in this chapter. It presents a comparative synthesis covering the objectives of the studies, the methodologies used, the types of data analysed, and the main contributions to degradation detection. The table aims to facilitate reading and enable comparison of the approaches followed.

3.1.2. Detection of Craquelure Patterns Models

Craquelure is a common form of degradation in paintings, characterised by a network of fissures that may result from mechanical impacts, variations in tension within the support, drying processes of the pictorial layer, or the chemical evolution of the materials constituting the artwork. Each crack network presents a unique pattern, functioning as a kind of “fingerprint” of the painting. Given its high prevalence, craquelure is one of the most extensively studied degradation phenomena in the context of applying artificial intelligence (AI) techniques to damage detection in paintings.

The literature therefore reveals a wide diversity of computational methods that can be applied to the study of craquelure. These approaches, and their variants, range from deep learning models dedicated to crack segmentation [25] to models that use craquelure patterns as tools for artwork authentication [33]. There are also studies that use the physical structure of craquelure as a stable element for registration and matching tasks across different imaging modalities. Despite their differences, all these works demonstrate the strong potential of AI to analyse complex patterns with greater precision and consistency than traditional methods. This combined body of research enables the organisation of these works into several methodological categories.

The first methodological category to consider is the automatic segmentation and identification of craquelure, illustrated by the studies of Yuan et al., [34] and Sizyakin et al. [35] Yuan et al. [34] used a deep-learning-based model that employs a ResNet-50 residual structure for feature extraction and integrates it into the skip-connection module of U-Net. This combination allows for the fusion of features at different hierarchical levels and produces more accurate crack boundary detections. The study demonstrates that deep segmentation models can successfully differentiate between various types of degradation, further validating practical applicability through their implementation on polychrome mural paintings from the Imperial Palace in Beijing.

In the same segmentation category, Sizyakin et al. [35] apply CNNs specifically trained for crack detection, demonstrating high performance in identifying craquelure across different imaging modalities, including visible photography, infrared, and X-radiography. Their work thus establishes an initial benchmark for craquelure detection using CNN-based approaches.

The second category involves the extraction and analysis of craquelure patterns as an artwork’s signature. In this domain, the work of Zabari et al. [36] is particularly notable, combining image processing with convolutional neural networks to analyse geometric crack patterns. The author focuses on craquelure as a distinctive characteristic of the artwork, proposing a method that not only detects fissures but interprets their configuration as historical information.

Regarding the use of craquelure as a structural element for multimodal correspondence and image registration, Sindel et al. [37] propose the CraquelureNet framework, designed to use crack patterns as reference structures to relate images obtained through different techniques. The system ensures accurate alignment across visible images, infrared reflectography, X-radiography, and UV fluorescence, demonstrating that craquelure can serve as an extremely stable structural marker for computational registration. This innovative approach expands the use of craquelure beyond detection, positioning it as a tool for data integration.

Finally, concerning generative models applied to structural analysis and authentication, Chirosca et al. [33] employs autoencoders to represent and analyse anomalies in artworks, including craquelure patterns. The study highlights the potential of generative models to detect subtle deviations from an artwork’s typical structural configuration. To achieve this, the author develops a modified convolutional neural network based on the VGG19 architecture and applies it to high-resolution greyscale images, aiming to assess the model’s capacity to extract both global and local features from craquelure patterns observed in paintings.

Although these studies adopt distinct approaches, they all demonstrate the relevance of craquelure and its computational value as an indicator of conservation state or material signature of the artwork. Consistently, results show that deep learning models—such as CNNs and U-Nets—significantly outperform traditional crack-segmentation methods, both in precision and robustness. These investigations act as catalysts for future work, paving the way for increasingly specialised models, better adapted to real case studies such as museum paintings, and confirming the versatility of craquelure as a resource for multiple tasks (damage segmentation, multimodal registration, and authentication).

Despite these convergences, relevant differences remain between the studies. While Yuan et al. and Sizyakin et al. [35] focus on practical craquelure detection with direct applications in conservation, Zabari and Sindel et al. [37,38] explore its potential as a structural element for scientific analysis and authentication. They also diverge methodologically, encompassing a spectrum that includes deep segmentation (U-Net, CNN), structural correspondence across modalities (CraquelureNet), and generative models based on autoencoders for compact pattern representation.

Comparing the categories of craquelure detection methods, it is important to note that they are not equivalent. While segmentation models (U-Net and CNNs) are effective in detecting craquelure with pixel-level precision, they are limited to visual detection and do not interpret the structural or historical significance of the cracks themselves. From another perspective, registration-based methods (CraquelureNet) focus on aligning images across different modalities and on the structural stability of patterns; however, they may be less sensitive to small and subtle localised damage. Finally, autoencoder-based approaches generate compact representations of crack patterns but introduce a higher risk of indirect interpretation and may produce features that do not always correspond to the physical reality of the artwork. This supports a transversal critical view that each approach addresses a different problem, and none is sufficient in isolation. Ultimately, there is a tension between local precision, structural interpretation, and material representativeness.

Although the models tested and presented in this review have shown promising results in craquelure detection, they still present several limitations. Most approaches rely on high-resolution, carefully curated datasets. In the case of Yuan et al. [34], the dataset consists of polychrome mural paintings from the Imperial Palace in Beijing; Sizyakin et al. [35] use visible, infrared, and X-radiography images; Zabari et al. [36] rely on historical paintings; Sindel et al. [37] use visible, infrared, X-ray, and UV fluorescence images; and Chirosca et al. [33] employ high-resolution greyscale images of paintings. Similar to the datasets discussed in Section 3.1.1, these collections present limitations in terms of variability in imaging conditions, surface textures, and degradation patterns, which in real-world conservation contexts are considerably more complex and heterogeneous. In addition, the performance of these models is strongly influenced by the quality and representativeness of the training data, which are often limited to specific collections or controlled acquisitions. This raises concerns regarding their generalisability when applied to museum collections or real-world conservation conditions.

In these paragraphs, it becomes evident that craquelure patterns are no longer interpreted solely as a pathological phenomenon in paintings, but rather as a rich and relevant computational object in the context of applying artificial intelligence to degradation detection. Craquelure patterns therefore exhibit two fundamental natures: (1) an indicator of the conservation state of the artwork and (2) an intrinsic material and structural signature of the painting itself. The combined analysis of the studies described in this chapter demonstrates that deep learning models are particularly well suited to dealing with complex, irregular, and highly variable patterns, such as those found in craquelure. In summary, craquelure constitutes one of the clearest examples of the real potential of artificial intelligence within the field of conservation and restoration.

Craquelure is, in general, a highly informative and complex form of degradation. It is often used as a structural fingerprint, supports authentication processes, and is simultaneously considered a manifestation of material deterioration. However, this multifunctional role implies that it is not a fixed or singular phenomenon. As crack patterns are influenced by multiple factors, including material composition, ageing processes, and environmental conditions, craquelure cannot be treated as a universal descriptor applicable to all artworks in AI-based approaches, hindering the data retrieval process. In other words, the variability inherent to craquelure patterns is so extensive that their generalisation across different paintings becomes highly problematic.

From a critical comparison of the reviewed studies, it can be observed that the segmentation models proposed by Yuan et al. and Sizyakin et al. [35] are the most directly applicable to conservation and restoration contexts, as they focus on the objective and precise detection of craquelure. In contrast, the structural and signature-based analysis approaches presented by Zabari et al. [36] extend the use of craquelure beyond simple detection, exploring its informational value as a distinctive feature of the artwork. Likewise, Sindel et al. demonstrate an innovative use of craquelure as a stable structural element for multimodal image integration. Nevertheless, several limitations common to these studies must be highlighted, namely the strong dependence on high-quality images, which are not always available, and the resulting limited validation in real museum contexts. Consequently, many of these works remain at the proof-of-concept stage, with restricted applicability as fully operational tools within conservation practice.

Despite the favourable results reported, most of the reviewed models are trained on relatively limited and poorly representative datasets, with insufficient attention given to the variability associated with artistic materials and techniques. In this regard, models developed primarily using historical paintings may exhibit reduced effectiveness when applied to contemporary art, due to significant differences in materials and working processes. Therefore, the future of these systems lies in the development of hybrid models that combine segmentation, structural analysis, and multimodal data, enabling a more comprehensive representation of the material complexity of artworks. In parallel, the integration of material and historical knowledge will be essential to adapt these models to specific case studies, rather than relying on universal solutions that are insufficiently sensitive to artistic and material context.

Following the discussion on craquelure, attention now turns to another prevalent form of degradation: paint loss. As mentioned above, craquelure is a specific form of degradation with complex structural patterns, and AI enables the precise identification of cracks, surpassing traditional methods. Despite the relevance of craquelure, paintings often exhibit other types of degradation. In the following chapter, we will discuss the application of AI to paint loss detection, which frequently results from previous forms of damage such as craquelure and flaking. In the same way that AI models can segment and analyse craquelure, they can also be trained to identify areas of paint loss, considering different visual and structural patterns.

Similar to the table presented in Section 3.1.1 regarding generalist degradation detection models,

Table 2. Synthesis of the studies addressed regarding the detection of craquelure pattern models.

Study	Yuan et al. 2023 [34]	Sizyakin et al. 2020 [35]	Zabari et al. 2021 [36]	Sindel et al. 2021 [37]	Chirosca et al. 2025 [33]
Objective/focus	Automatic segmentation and identification of craquelure	Develop and evaluate CNN-based models for the automatic detection of craquelure in paintings, demonstrating their robustness across multiple imaging modalities	CNN-based detection of craquelure across modalities	Extraction and analysis of craquelure as an artwork signature	Structural analysis and authentication using generative models
Method/model	Deep learning (U-Net with ResNet-50 residual structure)	Convolutional neural networks (CNNs)	Imagen processing + CNNs	CraquelureNet (structural pattern matching)	Autoencoders + modified CNN (VGG19-based)
Data/Case-study	Polychrome mural paintings, Imperial Palace, Beijing	Visible, infrared, and X-radiography images	Historical paintings	Visible, infrared, X-ray, and UV fluorescence images	High-resolution greyscale images of paintings
Contributions	Outperformed traditional crack segmentation and manual detection; captured subtle pigment loss patterns; demonstrated deep learning applicability in conservation.	Complete pipeline for paint loss detection and inpainting; reduced need for large annotated datasets; broader applicability across artworks.	Detects fissures and interprets geometric crack patterns as distinctive features; provides historical and structural information beyond simple detection.	Uses crack patterns as stable structural markers for alignment; expands craquelure application to data integration and registration.	Detects subtle deviations in craquelure patterns; extracts global and local features; demonstrates potential of generative models for authentication and anomaly detection.

comparatively presents the studies addressing the detection of craquelure pattern models.

3.1.3. Detection of Paint Loss Models

Paint losses in paintings are characterised by the absence of a pictorial layer, often resulting from the progression of other forms of degradation, such as flaking or cracking. These losses compromise the painting by interrupting its visual reading and altering the perception of the original artwork. Identifying their extent is essential to inform conservation interventions, while documenting the location and size of the losses allows for monitoring the degradation over time.

Traditional detection methods, such as visual inspection with the naked eye, raking light examination, or observation using an optical microscope or magnifying lenses, present limitations, namely being time-consuming, labour-intensive, and subject to observer bias. The introduction of artificial intelligence (AI) emerges as an innovative approach, capable of reducing analysis time, automating the detection process, and identifying subtle losses that might go unnoticed in manual inspections.

In 2018, Meeus et al. [38] used the Ghent Altarpiece as a case study, applying 2D convolutional neural networks (CNNs) for the automatic detection of paint losses. The model demonstrated performance superior to traditional crack segmentation and manual detection methods, capturing subtle patterns of pigment loss even in complex areas or those with irregular textures. This study proved the applicability of deep learning in the conservation and documentation of artworks. Further development of this work was published in 2020.

In addition to confirming the effectiveness of AI in paint loss detection, the study integrated automatic inpainting for digital reconstruction of loss areas and implemented shared pretraining to reduce the need for large annotated datasets. This advancement broadened the applicability of the algorithm to different artworks.

Following these initial advances, more diverse approaches emerged. Li et al. [39] mapped and identified paint loss areas in ancient murals of the Qutan Temple, focusing on zones with severe degradation. A major challenge was the colour similarity between paint loss areas and calcified white patterns. To address this, they developed a 3D Residual Network (3D ResNet) model, capable of simultaneously analysing spatial and spectral features. The network employed filters of varying dimensions, capturing details at multiple scales and integrating information across image dimensions, effectively distinguishing paint losses from white patterns.

In 2025, Chen et al. [40] proposed the PLDS-YOLO model, based on the YOLOv8s-seg architecture, to overcome the limitations of existing methods, such as incomplete segmentation, high image noise, and failure to detect small areas. The improvements included the following:

• PA-FPN with residual connections, enhancing the fusion of high-resolution features with deep semantic features;
• Dual-Backbone (CSPDarkNet + ShuffleNet V2), increasing feature extraction at multiple scales and improving discrimination of degraded areas against complex backgrounds;
• SPD-Conv module (Space-to-Depth Convolution), replacing traditional pooling layers, enabling better detection of small areas and perception of degraded regions of different sizes.

The PLDS-YOLO model achieved a segmentation accuracy of 86.2%, outperforming existing models and demonstrating strong potential for practical application in the conservation of murals and polychrome paintings.

Considering what has been discussed regarding paint losses in paintings and their detection using artificial intelligence models, it is imperative to state that this approach substantially reduces the subjectivity associated with observation and diagnosis through traditional methods, while also contributing to the automation and acceleration of paint loss identification and mapping processes. Moreover, the use of AI as a tool enables the detection of subtle losses or losses with irregular boundaries that are often difficult to delineate manually, highlighting the importance of accurate delimitation for conservation planning and long-term degradation monitoring.

From a temporal perspective, early applications focused on two-dimensional convolutional neural networks (2D CNNs) and on specific case studies, such as the Ghent Altarpiece, later incorporating shared pretraining strategies to reduce dependence on large annotated datasets. The development of more complex models, including ResNet architectures, three-dimensional networks, and YOLO-based segmentation models, has made it possible to overcome previous limitations, particularly in handling chromatic variations, distinguishing paint losses from other forms of heterogeneous degradation, and adapting to different artwork typologies, including easel paintings, murals, and polychrome surfaces.

Nevertheless, it is equally important to acknowledge the limitations associated with these approaches. These include a strong dependence on the material, technical, and historical context of the artwork, as well as the need for expert validation to distinguish paint losses from other pathologies, such as surface abrasions or previous interventions. In addition, limitations in model generalisation across different painting typologies and image acquisition conditions persist. Consequently, artificial intelligence should be regarded as a decision-support tool that complements—but does not replace—the critical judgement of the conservator-restorer.

Despite the promising results of AI-based paint loss detection models, it is necessary to critically examine the datasets used. Most studies rely on specific case studies, with manually annotated datasets and high-quality images acquired under controlled conditions. Although these data allow for effective model training and validation, they are often limited in diversity and do not fully represent the variability of artworks in conservation contexts, as already mentioned above. In real-world scenarios, there are significant differences in image acquisition conditions, material heterogeneity, and the history of past restorations and interventions, all of which should be taken into account in model development and evaluation.

In this context, a critical issue of class imbalance also arises, since in paint loss detection the affected areas tend to be small compared to the dominant background in the image. This imbalance may lead models to prioritise learning background features rather than the regions of interest, resulting in reduced sensitivity to small or subtle losses. Consequently, the generalisation capability of these models may be limited, with variations in performance across different artworks and imaging conditions. In summary, generalisation remains an open challenge in the application of AI-based approaches in conservation and restoration, despite the positive performance observed on controlled datasets.

Similar to the other chapters addressing degradation detection, this chapter also includes a summary and comparative table,

Table 3. Synthesis of the studies addressed regarding the detection of paint loss models.

Study	Meeus et al. 2018 [38]	Meeus et al. 2021 [41]	Li et al. 2023 [39]	Chen et al. 2025 [40]
Objective/focus	Automatic detection of paint loss	Expansion of previous study; integration of reconstruction of loss areas	Mapping and identification of paint loss in ancient murals	Overcome limitations of previous methods and improve segmentation
Method/model	2D convolutional neural network (CNN)	CNN + descriptor-based Inpainting + Shared Pretraining	3D Residual Network (3D ResNet)	PLDS-YOLO (YOLOv8s-seg + PA-FPN + Dual-Backbone + SPD-Conv)
Data/Case-study	Ghent Altarpiece, visible and multispectral images	Ghent Altarpiece	Ancient murals of Qutan Temple	Ancient murals and polychrome paintings
Contributions	Demonstrated the applicability of 2D convolutional neural networks for the automatic detection of paint losses in paintings, showing superior performance over traditional and manual methods and enabling the identification of subtle pigment losses in complex pictorial areas.	Integrates automatic inpainting and shared pretraining strategies, thereby reducing the need for large annotated datasets and broadening the applicability of the method to different artworks beyond a single case study.	Simultaneous analysis of spatial and spectral features; effective distinction between paint loss and calcified white patterns; multiscale detail extraction.	Segmentation accuracy of 86.2%; better detection of small and complex degraded areas; optimised for rapid and precise paint loss segmentation

, aiming to provide an overview of the studies analysed regarding paint loss detection.

4. Contemporary Paintings: Investigating Material Diversity and Degradation Using AI

It is important to highlight the need to critically examine the applicability of AI methodologies to real-world conservation and restoration practice. Firstly, the limitations of generic datasets must be acknowledged. Synthetic data or annotated collections often fail to capture the full variability of materials, techniques, and degradation phenomena present in actual artworks [42]. For instance, craquelure patterns vary significantly between individual paintings [43] and are not represented in standardised datasets. Consequently, models trained exclusively on controlled datasets may underperform when applied to real artworks, where multiple variables and interferences coexist [42,44]. Moreover, testing AI in real contexts allows for the identification of segmentation errors or model limitations that would remain undetected in simulations [29,44]. In sum, while generic datasets demonstrate the theoretical capabilities of AI, only validation in real-world contexts ensures its practical effectiveness under the complex conditions of conservation.

Within this framework, the study of contemporary paintings emerges as a particularly suitable scenario, as it allows for the exploration of degradation patterns specific to this artistic and material context [45,46]. Contemporary painting departs from traditional practices in several aspects, most notably in material heterogeneity [46,47,48]. It is common to observe the combination of multiple materials, such as collages or mixed techniques with different binders (e.g., oil and acrylic), as well as a wide variety of supports [46,47,48]. The paintings of Paula Rego (1935–2022) from the 1960s are a good example of this diversity (Figure 3 and Figure 4) [49].

Material heterogeneity plays a critical role in the performance of AI-based models for damage detection. As discussed in Section 3 of this review, many of these models rely on the extraction of visual and structural features; however, in the case of contemporary paintings, the coexistence of multiple materials, techniques, and surface properties may significantly alter these features during data acquisition. As a result, surface textures or material variations may be misinterpreted as degradation phenomena, leading to false positives, while actual damage may remain undetected.

Furthermore, models trained on more homogeneous datasets—such as traditional easel paintings—may face difficulties in generalising to more complex material contexts, resulting in reduced detection accuracy when applied to contemporary artworks. This highlights the need for models capable of integrating material variability, as well as for validation procedures specifically designed for heterogeneous artistic systems.

The 1960–1970 paintings of the artist Paula Rego are particularly relevant in this context, as they provide valuable cases for evaluating existing AI-based damage detection models, given the material and technical heterogeneity of these paintings when compared to traditional easel paintings. The materials and techniques used in her practice offer a useful testing ground for assessing the robustness of current models. As discussed in Section 3, many AI approaches—particularly those based on crack segmentation or pattern recognition—rely on the assumption of relatively consistent surface characteristics. However, the presence of collage elements, adhesive layers, and non-conventional materials may mislead these models, causing them to interpret inherent material features as degradation or fail in detecting actual damage. This highlights a critical gap in current research, as the performance of existing models under such complex material conditions remains largely untested.

While Paula Rego is discussed here due to her relevance to our future research focus, it is important to note that, within both Portuguese and international contemporary painting, there are numerous cases characterised by material heterogeneity, which constitute valuable examples for the application of artificial intelligence models in degradation detection. In particular, the use of collage became prominent with the emergence of the avant-garde movements, when artists moved away from traditional canons and began exploring alternative artistic approaches [50]. This led to the development of mixed media practices, combining materials of different origins to achieve the final artistic outcome [51]. Among Portuguese artists, Pedro Cabrita Reis (b. 1956) [52] is a relevant example, while in the international context Antoni Tàpies (1923–2012) [53] provides a comparable case.

De Winter et al. [54] analyse paintings containing fluorescent layers and show that materials such as foams, polyester, and fluorescent pigments are frequently used in modern artworks, creating specific challenges for conservation practices that are not encountered in the conservation of traditional artworks. Consequently, degradation in contemporary paintings differs from that in traditional works. Firstly, the relatively recent production of these artworks limits long-term knowledge of their ageing processes, constituting a significant temporal constraint [55]. Furthermore, the diversity of materials enhances degradation potential due to novel chemical interactions arising from the combination of often less stable, industrially manufactured materials [55]. This creates a realistic and challenging context for AI applicability.

This study hypothesises that the capability of AI to handle the material and visual diversity of contemporary painting stems from its ability to process images and data of various origins, including multiple types of reflectance, texture, and colour, thereby recognising degradation patterns even in heterogeneous layers [10,29]. AI models trained in real contexts can identify anomalies regardless of the support, a clear advantage for contemporary artworks. Additionally, AI is able to extract relevant features from complex images, reducing interpretation errors [25,54], which can be particularly valuable given the variation in colour, finishes, and surface appearance present in contemporary painting [55,56,57,58].

AI also proves to be a critical tool for the early detection of subtle degradation, such as initial-stage fissures or discolorations, and for continuous monitoring by comparing images of the same painting over time [28,29]. Early detection facilitates preventive intervention planning, reducing both costs and the risk of advanced deterioration [28,59]. Furthermore, AI enables the integration of multimodal data and the combination of analytical techniques—such as hyperspectral imaging, XRF, Raman, and FTIR—allowing for the correlation of visual alterations with chemical composition, thereby assisting in the identification of materials and degradation reactions [28,29]. The extraction and correlation of multiple features across these analytical datasets, without AI, would be extremely time-consuming and less precise. This can be particularly useful in contemporary paintings given the complexity of their compositions.

Although artificial intelligence has been successfully applied to the detection of degradation in various types of paintings, using architectures such as CNNs and U-Nets to identify craquelure, paint loss, and other forms of deterioration [30,31,32,33,34,35,36,37,38,39,40,41,42], the applicability of these methods to contemporary painting remains unexplored. At present, there are no studies addressing the capabilities of AI to detect damage specifically and exclusively in contemporary paintings. Van Vijle et al. [10] provide a comprehensive review of international applications of AI in painting conservation, highlighting successful implementations in classical and historical artworks. However, the existing literature does not specifically address contemporary paintings, revealing a gap in current research on the application of artificial intelligence to this area. On the other hand, García-Moreno et al. [32] present a concrete example of a machine learning technique (ARTDET) for detecting deterioration in easel paintings, illustrating that, while such tools exist, their application to contemporary artworks remains limited. Without this specific research, it is difficult to validate the effectiveness of AI models given the high variability of materials, techniques, and supports present in contemporary artworks [35].

This lack of studies highlights the need to explore contemporary case studies, allowing for the testing and adaptation of AI models to real-world contexts. The present review and planned future work aim to fill this gap by focusing on complex and representative works, with the artist Paula Rego serving as a particularly pertinent case study due to her relevance and material diversity [60,61]. Additionally, it would be valuable to explore the application of existing models to this new and distinct context of contemporary art, particularly focusing on artists and works that are already supported by a well-established body of research. This would allow for the validation of existing knowledge and help assess whether these approaches can be effectively adapted to this new scenario.

5. Findings and Discussion

Overall, the reviewed studies demonstrate that artificial intelligence enables the detection of degradation phenomena in paintings, with the advantage of identifying subtle changes that are not yet visible to the naked eye. This capability supports decision-making processes in conservation and restoration, promoting more informed preventive conservation strategies, which are fundamental in the field of conservation and restoration. Additionally, the use of artificial intelligence models is shown to consistently contribute to reducing subjectivity in the assessment of degradation in pictorial works.

Despite the differences observed in terms of model architectures and applications, the analysed studies reveal several convergent points. In particular, they highlight the ability to detect small alterations such as cracks, paint losses, and imperceptible stains, as well as to interpret these changes within the overall context of the image. These methods include different artificial intelligence approaches, such as convolutional neural networks (CNNs), spatiotemporal models, Mask R-CNN, and YOLO, demonstrating that despite methodological differences, the results are generally consistent and promising.

However, the literature also highlights important limitations, namely the reliance on high-resolution images, which are often unavailable in real conservation contexts, the need for large datasets, and the requirement for expert annotation. Another relevant limitation is the limited generalisability of the models across the diversity of artistic materials and techniques, which constrains their application to contemporary painting. This limitation becomes particularly evident when moving from controlled datasets to real-world museum environments.

Regarding the observed results, the application of AI allows for the detection of subtle alterations in paintings before they become visible to the naked eye, enhancing preventive conservation and damage mitigation. Moreover, machine learning and deep learning tools provide greater objectivity and consistency in degradation assessment, reducing human subjectivity.

In generalist degradation detection models, spatiotemporal approaches [62] showed high sensitivity to subtle damages and visual reconstruction of affected areas. Autoencoders applied to X-ray images [24] effectively located degraded regions, while lightweight models [29] improved real-time performance, albeit with lower precision for subtle damages. Mask R-CNN [31] segmented visible losses and restorations with high accuracy, but was limited to visually observable damages.

For craquelure detection, approaches such as those by Yuan et al., Sizyakin et al. [35], Zabari et al. [36], Sindel et al. [37], and Chirosca et al. [33] outperformed traditional methods in precision and robustness, supporting conservation assessment, structural analysis, and artwork authentication. Common limitations include dependence on high-resolution images, expert validation, and limited applicability in museum contexts.

Paint loss detection models enabled identification of subtle or irregular losses, complementing traditional methods. Examples include 2D CNNs [38], CNN + inpainting [41], 3D ResNet [39], and YOLOv8s-seg [40], showing accurate segmentation, distinction of complex patterns, and reduced need for large annotated datasets, albeit dependent on the material and historical context and requiring expert validation.

The application of AI to contemporary painting presents challenges due to material and technical diversity. Paula Rego’s works are a relevant case study, combining collages, mixed techniques, and industrial materials. Observed issues, such as detachment of collage elements, adhesive oxidation, and pigment binder degradation, make these works a pertinent example for AI-based degradation detection studies.

6. Conclusions

The application of artificial intelligence in the analysis of painting degradation represents a drastic and structural shift in how conservation is approached. Moving from essentially descriptive approaches to more quantitative, objective, and data-driven methodologies, AI is transforming this field, despite its current limitations. Beyond improving degradation detection, AI enables the interpretation of subtle material changes as measurable patterns, opening the way for earlier and more systematic conservation and intervention strategies based on informed decision-making.

Despite the diversity of approaches observed in this review, a common trend emerges in which the most effective models are those that combine fine-grained spatial feature detection with a high level of contextual information integration. However, the performance of these models strongly depends on data quality, image resolution, and the availability of annotated datasets, which remains a key limitation for their application in real-world contexts.

The application of these models to contemporary artworks clearly exposes their limitations due to the high heterogeneity of materials and techniques, highlighting the need for systems that are more adaptive and sensitive to material and historical context. In this sense, contemporary case studies are particularly relevant not only as application examples but also as critical tests of model robustness and generalisation.

Nevertheless, artificial intelligence should not be understood as a replacement for expert conservation analysis, but rather as a complementary tool that supports interpretation and the synthesis of results, facilitating decision-making, which is essential in conservation and restoration practice. The greatest potential of AI lies in hybrid approaches that integrate computational analysis with material, technical, and historical knowledge, as well as expert validation.

Future development should therefore focus not only on improving algorithms but, more importantly, on their integration into real conservation contexts, bridging the gap between research and professional practice.