Automated Edge Detection for Cultural Heritage Conservation: Comparative Evaluation of Classical and Deep Learning Methods on Artworks Affected by Natural Disaster Damage

Abstract

Assessing the condition of artworks is a critical step in cultural heritage conservation that traditionally involves manual damage mapping, which is time-consuming and reliant on expert input. This study, conducted within the ChemiNova project, explores the automation of edge detection using both classical image processing techniques (Canny, Sobel, and Laplacian) and a deep learning model (DexiNed). The methodology integrates interdisciplinary collaboration between conservation professionals and computer scientists, applying these algorithms to artworks affected by environmental damage, including flooding. Preprocessing and post-processing techniques were used to enhance detection accuracy and reduce noise. The results show that while traditional methods often yield higher precision and recall scores, they are also sensitive to texture and contrast variations. These findings suggest that automated edge detection can support conservation efforts by streamlining condition assessments and improving documentation.

1. Introduction

UNESCO defines the conservation of cultural heritage as the measures taken to extend the life of cultural heritage items [1], and it adds that the aim is to maintain the physical and cultural characteristics of the object, thereby dictating the importance of analysis and knowledge of the artwork in this process, as outlined in restoration theory [2]. The conservation of cultural heritage is predicated on a preliminary analysis of the artwork, which enables the identification of all its aspects—physical, artistic, historical, cultural, and material—and its subsequent monitoring. The process of conservation and restoration of the artwork commences with the documentation of the artwork via photographic documentation of the cultural heritage, thus ensuring that a record of the form and structure is maintained prior to any intervention [3]. The documentation includes everything about the object, a formal description of the work, the technical aspects, the state of conservation, and any pathologies that are detected, past interventions, etc.

To facilitate a clearer and more concrete visualization of the state of conservation of the work, a graphical representation known as damage mapping is created at the time of the diagnosis and recording of the state of conservation. This approach is crucial and offers a visual framework that facilitates the identification of damage with greater precision and clarity.

A damage map [4] is defined as a scale drawing or plan that is typically created using a photograph of the site as a reference. The map’s design incorporates essential forms and characteristics, with greater levels of detail corresponding to increased precision. In this line drawing, the damaged areas are drawn on using different colors and textures in order to identify them. The presence of a legend is an inherent aspect of the process, serving the fundamental purpose of facilitating the identification of the various pathologies. It helps scientists to quantify the damage and to locate it on the surface of the construction site.

It is evident that, over the years and with technological advances, this process has become more sophisticated and somewhat easier. Initially, a printed photograph formed the base, and translucent paper was used to trace the lines of the shapes; once the contours of the artwork had been traced, lines were drawn to mark the damage. This process started to be performed with CAD technology, such as Adobe Illustrator^® or CorelDRAW^®, wherein the superimposition of layers can be used for the representation of the different levels of information. Instruments such as iPads or graphic tablets, digitally imitating the use of paper and pencil, are used too. In any event, all these steps have always been carried out manually, which means a great investment of time, especially if the item is a work of art with a great deal of detail. Obviously, the use of new technologies favors the preservation of cultural heritage; it is easy to find projects where a damage map was created on a 3D model [5], mixing technology and conservation, but the reference lines of the construction site were not drawn up in this project as the damage was superimposed onto the 3D model in a semi-transparent way.

Co-ordinated by the Universitat de València, the research project ChemiNova is funded by the Horizon Europe program. Its main objective is the development of new and advanced technologies for the analysis, conservation, and monitoring of cultural heritage that is mainly threatened by climate change and armed conflicts. Concern about climate change affecting both tangible and intangible cultural assets is growing, and this phenomenon needs to be researched in depth to establish its direct consequences on the cultural sector. In 2003, the European Commission became the first EU institution to launch a project linking climate change and cultural heritage [6]. In recent decades, there has been an increase in studies addressing the impact of climate change on cultural heritage, particularly in Europe, but also in other countries like China [7]. Research constantly shows that floods, heavy rains, and periods of drought are becoming more frequent. For example, Cacciotti et al. [8] emphasize the need for optimized management strategies to mitigate the effects of climate-induced disasters. Similarly, Bonazza et al. [9] focus on safeguarding cultural heritage from culture-related hazards by adopting integrated risk assessment frameworks and co-operation across regions. Nastou and Zerefos [10] provide a broader review of how climate change is affecting open-air heritage, where rising global temperatures and sea-level changes threaten archeological sites, historic landscapes, and vernacular architecture.

The idea of the ChemiNova project is to use existing technologies and create tools that have different uses within conservation and cultural heritage. The project is primarily people-centered, so any system will be intended for use by both heritage professionals and non-expert citizens. This system will combine 3D modeling, artificial intelligence, augmented reality, mixed reality, photogrammetry, and spectral data, among other elements. Currently, building damage maps are widely used after catastrophes such as earthquakes, and automated methodologies [11] and photogrammetry are used for the production of damage maps or the detection of damage, based on 3D point clouds [12]. In recent years, artificial intelligence has become a common tool in our lives. It is also finding its way into documentation processes in conservation and restoration. Recent research shows that thanks to machine learning, artificial intelligence is able to detect certain forms of damage on construction sites with high accuracy rates [13,14,15].

Unlike other studies involving ARTDET, which uses high-resolution images annotated by experts [13], in ChemiNova, we are not only using high-resolution RGB images for machine learning but also introducing spectral data from a hyperspectral camera, which adds information for damage characterization. This will mean that the data and characterization obtained with AI will be much more accurate, as it will have information on damage recorded in both dimensions by various sensors, along with non-invasive image analysis such as multiband imaging and hyperspectral imaging. Going a step further, this research focuses on the work prior to automatic damage detection or AI training for assisting visual inspections of cultural heritage [14,16], such as the extraction of contours to provide a basis for damage mapping. As yet, the first tests have been carried out with easel-mounted works belonging to the Martinez Guerricabeitia contemporary art collection of the Universitat de València. These artworks were affected by the floods in the province of València at the end of October 2024. Climate change not only causes slow cumulative damage but also results in extreme phenomena, as can be seen in the study by Bonazza and Sardella [17]. These events show how climate change impacts directly and most clearly on the materials of which the work is made, i.e., on the tangible part [18].

Edge detection is a fundamental tool in the fields of image processing and computer vision. This tool is primarily used to identify points in a digital image where the brightness changes sharply. It is often the case that these points correspond to object boundaries. Thus, it is a crucial step in a variety of tasks, including object recognition, image segmentation, and scene analysis [19].

Recent efforts have also explored the optimization of classical edge detection techniques. As an example, Li and Zhang [20] present an optimized edge detection method that enhances the classical Canny algorithm by integrating integral image detection to accelerate pyramid construction and uses Gaussian separation to simplify the filtering process. Additionally, they introduce a feature point filtering strategy using Canny edge detection with morphological dilatation, reducing the number of features while maintaining the accuracy of the detection process and improving computational efficiency.

In terms of deep learning, Casillo et al. [21] present a framework that combines deep learning with Internet of Things (IOT) and digital twin methodologies to safeguard cultural heritage buildings, focusing on the Archeological Park in Pompeii. By gathering mainly meteorological data, along with sensor readings and thermal images, a generative adversarial network (GAN) is trained to predict changes in the structures. Sun et al. [22] explore how deep learning models can be used to support the preservation of intangible cultural heritage, with a special emphasis on the Li ethnic subgroups (Hainan Island).

This paper is organized as follows: Section 2 describes the materials and methods, including the selection of artwork, the data acquisition process, and the edge detection techniques applied. Section 3 presents the results of the experiments, offering both a visual and qualitative evaluation of classical and deep learning-based algorithms. Finally, Section 4 discusses the implications of the findings, concludes the study, and outlines future directions.

2. Materials and Methods

2.1. The Collection

In the ChemiNova project, we are dealing with four different pilot schemes, namely, in Palermo, Italy; Kyiv, Ukraine; Vienna, Austria; and Valencia, Spain. In this paper, we are focusing on collections at the pilot scheme in Valencia. Each pilot includes different case studies divided into two sets of scales: the building and the artifacts. Valencia’s pilot examines both working scales; first, the historical building, the “Centre Cultural La Nau”, belonging to the Universitat de València; second, the contemporary collection of Martínez Guerricabeitia, which is part of the university’s heritage collection, located in the Centre Cultural La Nau. Finally, there is Lluis Vives’ sculpture, which is exhibited outdoors in the main cloister of the historical building.

Additionally, the artworks investigated in the current analysis were affected by the flash flood that took place in the south area of València on 29 October 2024. These artworks are part of the Martínez Guerricabeitia collection; initially, they were not selected for the project as they were in a traveling exhibition. However, since ChemiNova aims to improve the monitoring and conservation of cultural heritage that is threatened by climate change, it was considered interesting to include an analysis of these artworks in some of the project processes.

The artworks that will be the focus of this article are displayed in Figure 1, and

Table 1. Information about the artworks.

Title	Author	Date	Materials	Dimensions
Zoosofías no. 100	Paloma Pájaro	2018	Acrylics and pencil on wooden board	180 cm diameter, including the frame
Abrazada al fantasma del amor (ES) Embracing the ghost of love (EN)	Juan Ripollés	1980	Acrylics on canvas	65 × 80.5 cm
Nobody will speak of us when we are dead	María Carbonell	2017	Oil and spray paint on linen	150 × 195 cm
El nuevo orden mundial (ES) The new world order (EN)	Martín Caballero	1992	Acrylics on canvas	180 × 150 cm

contains some information about them. We have chosen pieces where the contours and geometry are different, as well as works with flat colors where the shapes are clearly distinguishable and others where the technique is diffuse, which makes it difficult for the system to recognize the edges of the various elements.

2.2. Data Acquisition and Traditional Methods for Edge Tracing

Damage mapping has evolved from the most analogue to the most digital method possible. Images were taken with analogue cameras and then had to be developed, enlarged, and hand-drawn on a translucent support that allowed the photograph to be traced.

With the development of digital cameras and the use of computers, the most practical and most common method nowadays is the digitization of maps, either via CAD programs or with drawing tools.

2.2.1. Data Acquisition

Images in the visible spectrum were taken using a Sony a7R IV digital camera (Sony Corporation, Tokio, Japan) paired with an FE 24–105 mm F4 G OSS lens (Sony Corporation, Beijing, China). To achieve optimal conditions, the acquisition parameters were meticulously managed. The lighting setup consisted of the ARRILITE 750 Plus, which utilizes radiation from an HPL* 750 W lamp (Arnold & Richter Cine Technik GmbH & Co, Stephanskrichen, Germany). With a correlated color temperature of 3200 K, the lights were arranged in accordance with CIE standards for colorimetric assessments. Both lights were symmetrically positioned at a 45° angle relative to the artwork surfaces, maintaining a distance of approximately 2 m from the canvas.

The acquisition parameters of each artwork are described in

Table 2. Acquisition parameters and imaging resolution used for each artwork.

Artwork (In English)	Focal Length	F Number	Exposure Time	Resolution (Pixels)
Zoosofias no. 100	24 mm	f/9	8 sec	4491 × 4425
Embracing the ghost of love	24 mm	f/11	1/4	4496 × 3655
Nobody will speak of us when we are dead	24 mm	f/9	1/3	6246 × 4971
The new world order	24 mm	f/9	1/3	4981 × 5832

, as well as their resolution in pixels.

2.2.2. Traditional Methods for Edge Tracing

Currently, damage maps, in terms of both the extraction of shapes and the characterization or determination of damage, are made with vector programs such as Corel Draw^® or Illustrator^®, among others. However, there are experts who prefer working with Photoshop^® [23], in this case with bitmaps instead of vectors. In contrast, there are people who use drawing software on devices such as tablets or iPads, recreating the traditional manual form of image capture. In all cases, the process of shape extraction followed by damage mapping, all completed manually without assistance from any software, is a considerable time investment.

Figure 2 shows the manual edge detection process, completed manually by a conservation expert with an iPad, an Apple Pencil, and a vectorial drawing software called Curve, which is then superimposed on its corresponding image with some transparency. This process is performed by tracing the vectors directly onto the original image, so that the lines are drawn directly on the edges.

These hand-drawn contour drawings are what conservators and restorers use as a reference for the subsequent generation of a damage map, so these images are the ones that will be used as a reference for the subsequent verification of what was generated for this research by the AI.

No formal protocol checks are made on these drawings as they are carefully made by an expert. Therefore, contours were traced according to the expert’s subjective judgment, guided by their professional knowledge and a visual interpretation of the images.

It is observable that the human-created edges are not entirely accurate. In Figure 2d, specific details such as the pattern of the woman’s jacket on the left and the bird’s own feathers have not been considered as an edge.

2.3. Condition Assessment

By means of condition reports that provide us with information about the work of art we are studying, we analyze the cultural element completely at the moment when we compile the report. It includes photographic documentation of the damage, the damage map, and a detailed written analysis of the work’s condition, creating a compilation of the presence of any structural, physical, or chemical damage. It also contains technical and material information about the object, non-invasive investigations, and a historical summary, which should also describe any previous intervention. In the ChemiNova project, we have created two different condition report templates, which were produced for the two scales of monitoring “Object” and “Building” according to the guidelines of European Technical Standards CEN/TC 346 “Conservation of Cultural Heritage” EN 16085:2012, EN 15898:2019, EN 17135:2020, EN 16853:2023 [24,25,26,27].

2.4. Automatization of Processes

We use two libraries in Python 3.10.18 (OpenCV and Scikit-Image), along with different methods of contour detection. As their names suggest, OpenCV focuses on computer vision algorithms, while Scikit-Image focuses on image processing.

The methods used are traditional algorithms such as Canny, Sobel, and Laplacian.

2.4.1. Image Preprocessing

Before implementing edge detection algorithms, it is essential to conduct a preprocessing stage to enhance detection quality. In our approach, we applied a series of transformations to the original image to assess how various preprocessing techniques influenced the outcome.

Initially, the image is converted to grayscale using OpenCV’s cvtColor function, reducing it to a single intensity channel. Then, we apply histogram equalization with equalizeHist to enhance the contrast of the grayscale image, making edges more distinguishable in areas with poor lighting or low contrast.

The final step involves applying adaptive thresholding (adaptiveThreshold) with a Gaussian weighting method, which segments the image into different areas based on pixel intensity.

2.4.2. Edge Detection Algorithms

The Canny edge algorithm follows a multistep process that starts with a Gaussian filter to smooth the image, followed by finding the intensity gradient [28]. Next, non-maximum suppression is applied, which reduces the edge width to a single pixel; finally, hysteresis thresholding is used to track and join edge segments by using two thresholds.

In our implementation, we decided to use a dynamic thresholding strategy based on the median intensity of the image, adjusting the lower and upper thresholds proportionally using the sigma value. This approach allows the function to adapt to different lighting conditions and contrast levels in the input image.

The Sobel operator is a gradient-based method that approximates the first derivative of image intensity. It uses convolutional kernels, typically two 3 × 3 kernels (Equation (1)), to compute gradients in the vertical and horizontal directions [29]. This means that this method can detect changes in the intensity of each direction that combine to obtain the gradient magnitude (Equation (2)).

$$G_x=\left[\begin{array}{ccc}-1& 0& 1\\ -2& 0& 2\\ -1& 0& 1\end{array}\right] G_y=\left[\begin{array}{ccc}-1& -2& 1\\ 0& 0& 0\\ 1& 2& 1\end{array}\right]$$

(1)

$$\mathrm{Magnitude}=\sqrt{{\left(G_x\right)}^2+{\left(G_y\right)}^2}$$

(2)

The last step of this method is to apply a threshold that will isolate the significant edges from the ones that are not important.

The Laplacian of the Gaussian filter, also known as the Laplacian, uses the second derivative of the image to detect the zero-crossing of the image intensity [30].

2.4.3. Refinement of Results

Following edge detection, several post-processing steps are implemented to enhance the clarity and usability of the results. Figure 3 illustrates the progressive refinement of the edge map, transitioning from the raw detection image (a) to a clean, skeletonized representation.

Initially, the image output is normalized using OpenCV’s “normalize” function (b). After normalization, a binarization process is employed to isolate the prominent edges and suppress background noise. Otsu’s thresholding is applied to segment the image into binary regions (c).

For further refinement, particularly when comparing the detected edges with the ground truth, a skeletonization process is conducted (d). This involves inverting the binary image so that edges appear white on a black background and reducing the binary edge map to a one-pixel-wide representation. This approach preserves the structure of the edges while simplifying the image.

In addition to the post-processing steps, certain images with intricate details, such as the artwork under investigation in Figure 3, have been processed with a high-pass filter. This technique employs Fourier transform to transition from the spatial domain to the frequency domain.

3. Results

In this section, we present the experimental results obtained from an evaluation of our edge detection method. The evaluation was conducted on a variety of image patches that vary in complexity, thereby enabling a more comprehensive analysis of the method’s robustness and adaptability.

To facilitate a more intuitive understanding of the model’s behavior, we generated error visualization images in which different types of classification outcomes are color-coded: false negatives are shown in purple (RGB: 156, 39, 176), tolerated true positives are in green (RGB: 76, 175, 80), and real false positives are in orange (RGB: 255, 152, 0). Tolerated true positives refer to the predicted edges that fall within a dilated region (15 × 15 kernel) around the ground truth edges, allowing for minor spatial deviations. False negatives correspond to ground truth edges that were not detected by the model, indicating missed detections. In contrast, false positives represent predicted edges that do not correspond to any ground truth edge, highlighting over-detection or noise sensitivity.

In addition, a comparison was made between our results and those obtained using the DexiNed model [31], for which a series of experiments were carried out on both the original input images and those processed through various preprocessing techniques. To ensure a fair comparison, the same post-processing steps were applied to the outputs of both methods. These visualizations and comparative analyses provide insight into edge localization accuracy, noise resilience, and computational efficiency.

For a quantitative evaluation, the F1-score was used as the reference metric because of its balance between precision and recall. Furthermore, we calculated several structural similarity metrics, including the SSIM (structural similarity index measure) and FOM (figure of merit), to evaluate the perceptual and spatial accuracy. Standard classification metrics were incorporated as well. The following parameters must be taken under consideration: precision, recall, specificity, false positive rate (FPR), false negative rate (FNR), true positive rate (TPR), and true negative rate (TNR). These metrics offer a comprehensive perspective on the model’s performance across various evaluation dimensions.

3.1. Zoosofías No. 100

In this experiment, we evaluated the edge detection algorithm on a natural scene that features two animals in a forest.

Figure 4 presents the error maps generated by each edge detection algorithm, offering a visual comparison of their performance in terms of misclassification.

One of the first noticeable aspects is that all classical, non-deep learning algorithms are influenced by the physical characteristics of the artwork, such as the texture of its materials, which affect the edge detection performance.

The Canny method produces some clean edges but introduces several false positives (orange areas), particularly in the background and textured regions. It also misses some details, resulting in false negatives (purple lines) among the animals, especially of their limbs, hair, and ears.

The Sobel method produces thicker and fragmented edges, with an increase in false positives across the sky and ground. It also fails to capture some fine anatomical features, which leads to false negatives in some regions.

The Laplacian method exhibits the highest density of false positives, generating excessive responses in areas with subtle tonal transitions. Although it detects some structural contours, it lacks precision and introduces significant noise. False negatives are also present, particularly in the internal contours of the animals, indicating limited effectiveness at capturing fine details.

DexiNed, a deep learning-based method, offers the most coherent and continuous edge structures. It achieves a higher number of tolerated true positives, especially along the main outlines of the animals. However, it still presents false positives in terms of background textures and false negatives in low-contrast or thin-edged regions, such as the legs and ears.

Table 3. Metrics and parameters for each experiment on the “Zoosofías no. 100” artwork.

Algorithm	Preprocess	Kernel Size	Sigma	F1 Score	SSIM	FOM	Runtime (s)
Canny	Equalized		0.3	0.4000	0.8833	0.2075	0.8080
Sobel	Equalized	1		0.3668	0.8143	0.1504	1.4573
Laplacian	Gray	5		0.3388	0.8049	0.1417	0.8611
DexiNed	Gray			0.2770	0.8963	0.1489	39.9514

provides complementary details for each experiment, including the parameters used and the quantitative metrics obtained, allowing for a more comprehensive evaluation of each method.

The Canny method was applied to an equalized image with a sigma of 0.3, meaning that the lower and upper thresholds were calculated based on that parameter and the median of the image. It achieves the highest F1 score (0.4000), indicating a good balance between precision and recall. It also maintains a high SSIM score (0.8833), suggesting strong structural similarity with the ground edges, and the best FOM (0.2075), reflecting accurate edge detection.

Sobel, when also applied to an equalized image, shows slightly lower performance across all metrics, with an F1 score of 0.3668 and a noticeable drop in SSIM (0.8143), indicating less structural similarity. The Laplacian algorithm, using a grayscale input and a kernel size of 5, results in slightly lower values for all metrics, which suggests an error in the edge detection, as mentioned in the visual comparison.

DexiNed, despite being a deep learning-based method, achieves the highest SSIM score (0.8963). However, its F1 score (0.2770) is the lowest of all methods, indicating that while its output is visually coherent, there might be more errors compared to the other methods.

In terms of runtime, classical algorithms demonstrated significantly faster execution compared to the deep learning-based DexiNed. Canny completed its processing in 0.808 s, while Sobel and Laplacian required 1.457 and 0.861 s, respectively. In contrast, DexiNed took 39.95 s to process the same image. This substantial difference highlights the computational efficiency of traditional methods, which may be advantageous in time-sensitive conservation workflows or when operating with limited hardware resources.

3.2. Embracing the Ghost of Love

In this experiment, we evaluated the edge detection algorithm on a scene where the edges are clearly visible because of the traces of different colors.

Figure 5 presents the error maps generated by each edge detection algorithm, offering a visual comparison of their performance in terms of misclassifications.

The Canny method shows a high number of tolerated true positives (green areas), indicating that much of the edge detection is correct, thanks to the tolerated 15 × 15 kernel. However, there are some false negatives (orange areas), especially around finer details such as the whole cupboard or some tiles, suggesting low-contrast edges. It also displays some missed edges as false negatives (purple) in the center of the artwork.

Despite having some similarities with the Canny result, Sobel shows a higher concentration of false negatives (purple), again in areas with low contrast. True positives are less prominent, indicating that Sobel struggles to capture the full extent of the edges.

The Laplacian method maintains a similar level of false negatives to Sobel, with purple areas distributed across the image. True positives (green) are more accurately aligned with the ground truth, showing better edge localization. False positives (orange) remain present but appear less intense, indicating a slight improvement in noise handling.

DexiNed demonstrates the most balanced and accurate performance. False negatives (purple) are minimal, showing that the model rarely misses ground truth edges. True positives (green) are densely and precisely distributed, closely matching the reference edges. False positives (orange) are significantly reduced, highlighting DexiNed’s robustness against noise and its ability to avoid spurious detections. Most of the errors produced by this model surround the ground truth edges, unlike all the other methods.

Table 4. Metrics and parameters for each experiment on the “Embracing the ghost of love” artwork.

Algorithm	Preprocess	Kernel Size	Sigma	F1 Score	SSIM	FOM	Runtime (s)
Sobel	Gray	3		0.5550	0.8418	0.4067	0.6769
Laplacian	Gray	3		0.5191	0.8197	0.3424	0.1674
DexiNed	Gray			0.4766	0.8174	0.2712	50.2866
Canny	Gray		0.1	0.4410	0.8297	0.3008	0.1770

provides complementary details for each experiment, including the parameters used and the quantitative metrics obtained, allowing for a more comprehensive evaluation of each method.

The Sobel method achieves the highest F1 score (0.5550) and FOM (0.4067), indicating a strong balance between detecting true edges and minimizing false positives. Its SSIM score (0.8418) is also the highest, suggesting that the detected edges closely resemble the structural layout of the ground truth. This implies that Sobel, while simple, is effective at capturing the relevant edges with relatively few false negatives and a controlled number of false positives.

The Laplacian method followed closely with an F1 score of 0.5191 and SSIM score of 0.8197. Its FOM (0.3424) is slightly lower than Sobel’s, indicating slightly less good spatial precision. The results suggest that Laplacian detects a reasonable number of true positives but may introduce more false positives or miss finer details, as reflected in the visual analysis.

DexiNed, despite its strong visual performance, scored lower in all three metrics: F1 score (0.4766), SSIM score (0.8174), and FOM (0.2712). This discrepancy highlights a key insight: while DexiNed excels at producing visually clean and perceptually accurate edge maps with minimal false positives (as seen in the qualitative analysis), its strict alignment with the ground truth may be less precise, possibly due to its tolerance for spatial deviations. This results in a lower F1 and FOM score, despite its high visual quality.

The Canny method had the lowest F1 score (0.4410) and a moderate SSIM score (0.8297). Its FOM score (0.3008) was slightly better than DexiNed’s, suggesting that while it misses more true edges (higher false negatives), it maintains a reasonable spatial alignment when it does detect edges. The results reflect Canny’s sensitivity to parameter tuning and its tendency to either under-detect or over-detect, depending on the image complexity.

The runtime analysis for this artwork revealed a similar pattern to the last artwork. Classical methods again performed efficiently, with Laplacian and Canny completing in approximately 0.17 s, and with Sobel slightly faster at 0.6769 s. DexiNed, however, required 50.2866 s, the highest requirement among all the methods tested.

3.3. Nobody Will Speak of Us When We Are Dead

In this experiment, the edges of this image are not defined by explicit outlines, but through nuanced variations in color and texture. Utilizing oil and spray paint on linen, the composition challenges conventional edge definition by allowing contours to emerge organically from chromatic shifts and surface irregularities.

Figure 6 presents the error maps generated by each edge detection algorithm, offering a visual comparison of their performance in terms of misclassifications.

The visual inspection of the non-deep learning algorithms shows, again, how the edge detection is affected by the texture of the artwork. The different color strokes generate some noise in the detection image and result in a great deal of false positives.

The Canny method captures many of the prominent edges, particularly tree outlines. However, it introduces some noise and misses finer details, especially in terms of the background foliage and the contours of the people. This results in a moderate presence of false negatives and false positives. True positives are present but are not dominant, indicating that while Canny detects the key structures, it struggles with subtle or low-contrast edges.

The Sobel output appears more fragmented. It detects some strong vertical and horizontal edges but fails to capture the full structure of any curved or diagonal elements, such as the tent’s roof or the inflatable pool. This leads to a higher number of false negatives and fewer true positives compared to Canny. False positives are also present, particularly in the textured areas, indicating limited precision.

The Laplacian method produces a denser edge map, with more continuous lines than Sobel, but also introduces more noise. It captures some of the finer details missed by Sobel, improving the number of true positives. However, the increase in false positives, especially in textured regions of the image like the trees and grass, suggests a trade-off between detail and noise resilience.

DexiNed demonstrates a nuanced performance in edge detection. False negatives (purple) are sparsely distributed but tend to be concentrated around finer details, such as the edges of the tent and tree trunks, indicating occasional missed detections in more complex regions. True positives (green) are visible along major structural edges, including the outlines of people, the tent, and larger tree trunks, reflecting a strong alignment with the ground truth. False positives (orange) are scattered across the image and, again, around the ground truth edges, suggesting a level of noise sensitivity and over-detection in less defined areas.

Table 5. Metrics and parameters for each experiment on the “Nobody will speak of us when we are dead” artwork.

Algorithm	Preprocess	Kernel Size	Sigma	F1 Score	SSIM	FOM	Runtime (s)
Sobel	Gray	3		0.3721	0.8704	0.2003	2.2227
Canny	Equalized		0.3	0.3265	0.8785	0.1814	0.9691
Laplacian	Gray	1		0.3217	0.8224	0.1459	1.4182
DexiNed	Gray			0.2137	0.8773	0.1063	25.6445

provides complementary details for each experiment, including the parameters used and the quantitative metrics obtained, allowing for a more comprehensive evaluation of each method.

The Sobel method achieved the highest F1 score (0.3721) and FOM (0.2003), indicating a strong balance between precision and recall. However, visual analysis reveals that Sobel struggles in low-contrast regions, where it produces a high number of false negatives.

The Canny method, on the other hand, demonstrates the highest SSIM score (0.8785), reflecting its excellent structural similarity with the ground truth. The use of histogram equalization and a large kernel size enhances edge continuity and reduces noise. While Canny effectively captures prominent edges, it tends to miss the finer details, particularly in low-contrast areas such as tiles and cupboards. These missed detections result in a lower F1 score (0.3265) and FOM score (0.1814) compared to Sobel, but its overall structural fidelity makes it a reliable choice for general edge detection tasks.

The Laplacian method offered a middle ground between Sobel and Canny. With an F1 score of 0.3217 and the lowest SSIM score (0.8224), it showed improved edge localization, aligning more accurately with the ground truth.

DexiNed, a deep learning-based method, stood out in terms of visual accuracy. Despite having the lowest F1 score (0.2137) and FOM score (0.1063), it produced the most faithful edge representations. The SSIM score (0.8773) was nearly as high as Canny’s, reinforcing its effectiveness in preserving structural integrity.

The runtime for classical methods remained low, with Canny completing the task in 0.9691 s, Laplacian in 1.4182 s, and Sobel in 2.2227 s. DexiNed, although visually effective, required 25.6445 s to complete the task.

3.4. The New World Order

In this experiment, we evaluated the edge detection algorithm on a highly textured acrylic-on-canvas painting. This artwork is characterized by dense material layering and brushwork, resulting in a richly varied topography. Unlike artwork with clearly defined boundaries, this composition presents edges that emerge through physical texture and subtle tonal shifts.

Figure 7 presents the error maps generated by each edge detection algorithm, offering a visual comparison of their performance in terms of misclassifications.

The Canny method shows a substantial amount of tolerated true positives (green), indicating effective edge detection within the tolerance region. However, false negatives (purple) are visible, particularly around finer structural areas, such as the inner contours of objects. False positives (orange) are present, suggesting moderate noise sensitivity and issues with its capability to detect edges whenever there is a different texture or change.

The Sobel method exhibits a similar performance to Canny, showing details such as the jacket’s pattern or the feathers that are not considered edges on the ground truth. This results in small details being shown as false positives.

The Laplacian method detects a high number of true positives in the whole artwork. However, on a few occasions, like the interior of the window or the olive branch that the bird is holding, some edges are not detected, generating some false negatives. The number of false positives increases considerably, in contrast to the other methods.

DexiNed produces false positives around the true edges of the images while missing some of the real edges.

Table 6. Metrics and parameters for each experiment on the artwork “The new world order”.

Algorithm	Preprocess	Kernel Size	Sigma	F1 Score	SSIM	FOM	Runtime (s)
Sobel	Equalized	5		0.4190	0.8249	0.2681	24.9849
Canny	Equalized		0.5	0.4127	0.8240	0.2519	23.7141
Laplacian	Gray	5		0.4091	0.6515	0.1761	23.7148
DexiNed	Gray			0.2534	0.8243	0.1383	51.9780

provides complementary details for each experiment, including the parameters used and the quantitative metrics obtained, allowing for a more comprehensive evaluation of each method.

The Sobel method achieves the highest F1 score (0.4190) and FOM (0.2681), indicating a strong balance between detecting true edges and minimizing false detections. It also has the highest SSIM value (0.8249), suggesting that the detected edges closely resemble the structural layout of the ground truth. This performance highlights Sobel’s effectiveness when combined with histogram equalization preprocessing and a 5 × 5 kernel.

The Canny method, also using an equalized image, follows closely, with an F1 score of 0.4127 and an SSIM of 0.8240. Although slightly lower than Sobel, its performance remains competitive. However, its FOM (0.2519) was marginally lower, indicating a slightly reduced spatial precision in edge localization.

The Laplacian method, applied to a grayscale image with a 5 × 5 kernel, shows a moderate drop in performance. Its F1 score (0.4091) was comparable to Canny and Sobel, but its SSIM value (0.6515) was significantly lower, reflecting a weaker structural match with the ground truth. The FOM (0.1761) also indicated reduced spatial accuracy, likely due to increased sensitivity to noise and less effective edge continuity.

Surprisingly, DexiNed, despite its superior visual performance, recorded the lowest scores in some metrics, namely, F1 score (0.2534) and FOM (0.1383). This discrepancy suggests that while DexiNed excels in perceptual quality and robustness to noise, its outputs may not align as precisely with the binary ground truth used for metric computation.

This highly textured artwork posed a greater computational challenge, reflected in longer runtimes across all methods. Canny, Sobel, and Laplacian each required approximately 23 to 25 s, a notable increase compared to the previous artworks. DexiNed, however, still exhibited the longest runtime at 51.98 s. The elevated processing times for all methods in this case are partly due to the application of a high-pass filter using Fourier transform techniques, which was necessary to enhance the edge clarity in the presence of dense material textures.

4. Discussion

The comparative evaluation of classical and deep learning-based edge detection algorithms in this study offers valuable insights into their respective strengths and limitations within the realm of cultural heritage conservation. Classical methods, such as Sobel and Canny, consistently achieved higher F1 scores, especially in artworks with distinct contours and high contrast. These algorithms are well established in image processing and excel at detecting sharp intensity transitions. However, their performance heavily relies on the visual clarity of the input image. In artworks featuring diffused brushwork, layered textures, or gradients, these methods often result in fragmented or noisy edge maps. Their dependence on fixed thresholds and local gradients limits their adaptability to the complex visuals found in contemporary and modern art.

Conversely, the deep learning-based DexiNed model produced more visually coherent and accurate edge maps, although its F1 score was consistently lower than that of the classical methods. This discrepancy can be attributed to several factors. Firstly, DexiNed was trained using the BIPED dataset (Barcelona Images for Perceptual Edge Detection), which primarily consists of scenes with well-defined objects. Consequently, the model may struggle with the ambiguities and textured edges present in cultural heritage artworks. Secondly, the evaluation metrics, particularly the F1 score, are sensitive to pixel-level alignment with the ground truth. Since DexiNed tends to produce smoother edges that may not precisely match the annotated contour, it is penalized despite exhibiting high structural similarity, as reflected in its SSIM values.

The characteristics of the Martínez Guerricabeitia collection played a central role in shaping the outcomes of this study. The 4 selected artworks were chosen from a larger set of 24 easel paintings available to the conservators of the University of València. Selection criteria included the presence of documented flood-related damage, the availability of high-resolution imaging, and the relevance of the works to ongoing conservation efforts. Additionally, the selected artworks present varying degrees of visual complexity and edge definition, which allow us to test the algorithms under different levels of difficulty from a human perception standpoint. While the sample size is limited, the chosen pieces represent a range of visual styles and technical challenges, from flat, high-contrast compositions to layered, textured surfaces. This diversity provided an initial test for evaluating the algorithms’ performance.

The integration of preprocessing and post-processing techniques—such as grayscale conversion, histogram equalization, adaptive thresholds, binarization, and skeletonization—proved essential for enhancing edge clarity and reducing noise. These steps were particularly beneficial in artworks with layered or textured surfaces, where raw edges detection often produces excessive noise or missed contours. The use of a 15 × 15 tolerance kernel in the evaluation also helped account for minor spatial deviations, offering a realistic assessment of algorithm performance.

The runtime analysis highlights a fundamental trade-off between visual accuracy and computational efficiency. Although DexiNed generates edge maps that are more visually coherent and structurally faithful, its high computational demand restricts its use in scenarios where processing speed is crucial. This is especially pertinent in cultural heritage conservation, where technical resources might be limited and rapid decision-making is often necessary, particularly in post-disaster situations.

In contrast, classical methods, while less advanced in terms of visual representation, offer faster and more lightweight implementation. This makes them valuable for preliminary documentation tasks or for integration into hybrid workflows, where an initial quick detection process can subsequently be refined with more advanced techniques.

Thus, runtime should not be seen merely as a technical metric but as a strategic factor in choosing the most suitable algorithm for a specific case. Future research could investigate hybrid approaches that merge the speed of classical methods with the perceptual accuracy of deep learning models, thereby optimizing both quality and efficiency in edge detection for conservation purposes.

From a conservation perspective, these findings suggest that automated edge detection can significantly streamline the process of damage mapping, reducing the time and effort required for manual tracing. However, the choice of algorithm should be made according to the specific visual and material characteristics of the artwork. While classical methods may offer better metric performance in some cases, deep learning models provide more flexible and visually interpretable outputs, especially when trained on domain-specific data.

Although this study focused on easel paintings, the methodology explored here could be extended to other art forms such as murals and sculptures. In the case of murals, edge detection could assist in mapping surface deterioration or graffiti over large wall surfaces, especially when combined with photogrammetry or panoramic imaging. For sculptures and other three-dimensional objects, edge detection could be applied to surface textures and cracks, particularly when integrated with 3D scanning or structured light data. However, these applications would require adaptations to account for lighting variability, surface curvature, and material properties.

Future research should explore the development of deep learning models trained specifically on annotated cultural heritage datasets, including spectral and multiband imaging. Additionally, integrating edge detection with 3D modeling could enhance the spatial understanding of surface deformation and material textures. Combining the strengths of classical and AI methods may also offer a more robust solution for diverse conservation scenarios. Finally, the creation of user-friendly tools that support expert conservation and non-specialist users would align with the ChemiNova project’s people-centered mission.

While this study emphasizes the potential of automation, it is important to acknowledge the practical challenges associated with deploying such methods in real-world conservation workflows. One of the main obstacles is the computational cost of certain steps, particularly during preprocessing. Techniques such as histogram equalization, adaptive thresholds, and Fourier-based filtering can be resource-intensive, especially with the high-resolution images typical in heritage documentation. These steps may require significant processing time and memory, which could limit their use in field conditions or on devices with limited hardware capabilities.

5. Conclusions

This study provides a comprehensive evaluation of traditional and deep learning-based edge detection algorithms applied to the Martínez Guerricabeitia collection. The use of quantitative metrics such as F1 Score, SSIM, and FOM revealed that Sobel consistently achieved the highest scores in terms of precision and recall, while Canny, most of the time, demonstrated a strong structural index. Although scoring lower in some metrics, DexiNed produced the most visually coherent, accurate edge detection performance.

The integration of preprocessing techniques (grayscale, histogram equalization, and adaptive thresholding) and post-processing (binarization and skeletonization) proved necessary for enhancing the edge clarity and detection accuracy.

However, the results obtained here underscore the importance of tailoring edge detection strategies to the specific characteristics of each artwork, especially in the context of cultural heritage conservation. Automated edge detection offers valuable support for documenting and monitoring damage.

One of the primary directions for future research involves the training of deep learning models that are specifically tailored to cultural heritage using the data obtained from all the pilot schemes. However, this requires the availability of large, high-quality annotated datasets, which are currently quite limited.

Another promising path is the integration of 3D modeling techniques into the edge and damage analysis pipeline. By incorporating three-dimensional data, it would be possible to enhance the comprehension of surface deformations, material textures, and damage.