Artificial Intelligence Tools in the Management of Reinforced Concrete Structures: Potential, Critical Issues, and Preliminary Results on Structural Degradation

Abstract

The durability and management of reinforced concrete structures and infrastructures are a central issue in contemporary civil engineering. Efficient structural maintenance has become strategically critical to sustainable land and community management due to aging infrastructure, increasing operational stress, and limited financial resources. This study focuses specifically on reinforced concrete bridge piers, whose fundamental structural role influences road infrastructure management strategies. The objective of this study is to develop and use a system based on convolutional neural networks to visually, rapidly, and automatically identify degraded portions of the reinforcement, based on images acquired on-site or from visual inspections, and classify their level of degradation. The topic addressed is highly innovative. The need to define and calibrate reliable degradation classification criteria, and the difficulty of obtaining images and classifying them correctly for database construction, have influenced the development of the study and make the results interesting and promising, but absolutely preliminary.

1. Introduction

The durability and management of reinforced concrete structures and infrastructures is a central issue in contemporary civil engineering, especially in relation to the existing built heritage [1]. With the progressive aging of the engineering works and the intensification of loads, usage and economic stresses, the efficient maintenance of the structures assumed a strategic role in the sustainable management of the territory and communities.

In particular, the degradation of existing structures and infrastructures, especially the older ones, represents one of the main causes of loss of performances and built heritage. This has the potential to result in repercussions on the safety and reliability of the system as a whole, with very high direct and indirect costs and losses.

In existing and older reinforced concrete structures, the degradation of the reinforcement represents an extremely important phenomenon [2,3,4,5,6,7,8,9].

The safety of bridges and viaducts in Italy has been a relevant issue for several years. It is strictly connected to the management and maintenance of existing structures. Much attention has been paid in the regulatory field to the assessment activities of networks and individual structures [10,11,12,13]. Currently, safety assessment activities are conducted following guidelines [14] that provide different levels of detail, based on six levels (ranging from zero to five). For the lowest levels (levels 0 and 1) the proposed study could be of great help not only by helping to provide information on bridges and viaducts but also by limiting the costs related to accurate surveys that are in any case based on visual inspections, design documentation and, finally, assessments that are in any case conditioned by the subjectivity and experience of the operator. In this context, the evaluation of the attention class (based on the ageing and degradation of the materials) can be supported by the proposed methodology which is also useful for reducing and making periodic inspections more efficient, providing a valuable tool to the management body to evaluate any further investigations (levels 2–5) by guiding and limiting the costs of in situ experimentation [15,16,17]. Furthermore, correct knowledge of the state of degradation can correctly guide intervention strategies, in a multi-criteria and multidisciplinary process [18,19].

In the context of traditional inspection techniques, which are primarily visual and encompass a range of accuracy levels, significant limitations emerge. These limitations are associated with the subjective nature of assessments and the substantial financial costs of investigation campaigns. These costs are frequently deemed to be unsustainable, even at low levels, given the substantial number of extant structures and the difficulty of extending the surveys to their significant portions [20,21,22,23,24,25].

In this scenario, the utilisation of advanced methodologies and instruments grounded in artificial intelligence (AI) could be a potentially efficacious solution to address the problem in a systematic and automated manner. The most recent scientific literature highlights the strong interest in the application of artificial intelligence, in particular deep learning and image processing, to improve the efficiency, accuracy and automation of crack detection, measurement and assessment in concrete structures, thus contributing significantly to the field of structural health monitoring [26,27].

Several researchers have investigated the use of convolutional neural networks (CNNs) and deep learning architectures, such as Fully Convolutional Networks (FCNs), for the analysis of damage to RC structures in terms of crack detection [28,29,30,31,32,33].

The use of various network architectures and the effectiveness of transfer learning to improve the detection accuracy is mainly oriented to the identification and segmentation of cracks (at appropriate levels of quality and quantity [34,35,36]), through the application of FCN-based methods and evaluating the performance of different pre-trained network architectures [37], also on civil infrastructures, reusing knowledge accumulated from other contexts [22]. Some studies focus on the use of image processing techniques to measure crack characteristics, such as length and width. This includes the use of smartphones and dedicated applications to facilitate non-contact measurements [38].

The evolution of artificial neural networks, particularly convolutional neural networks (CNNs), has enabled the automatic analysis of images, with potentially significant implications for the field of visual recognition of degradation [39,40,41]. Nevertheless, the efficacy of these tools is contingent upon the availability of correctly structured and significant data, which can reflect the complexity of the phenomena involved and the considerable variability of situations, details and structures [42]. The limited availability of representative image datasets, combined with the absence of standardized protocols for the acquisition and characterization of images and associated information (such as construction period, materials used and their mechanical characteristics, exposure environment, level of actions and stresses, element function, etc.), continues to represent a significant limitation to the optimal use of these resources. This limitation mainly affects the reliability of predictions.

In this context, the present study aims to contribute to the development of tools for the automatic recognition of the degradation of the reinforcements in reinforced concrete bridge and viaduct structures, focusing on the degradation manifested through the reduction in the diameter of the reinforcement bars and the degradation of the surrounding concrete. This degradation often represents both the initial cause of the start of the oxidation and then corrosion of the reinforcements and an effect (in terms of expulsion of significant portions of concrete) of the corrosion in an advanced stage.

The present study focuses in particular on the piers of reinforced concrete bridges, subjected to heterogeneous environmental conditions, including freeze–thaw cycles and alternating dry and wet periods. Such conditions result in the non-uniform diffusion of aggressive agents within the structure, leading to the formation of zones exhibiting varying levels of vulnerability to corrosion.

The objective of the present study is to develop and use a system based on convolutional neural networks to visually identify, in a rapid and automated manner, the degraded portions of the reinforcement, starting from images acquired on site or from visual inspections, and to classify the level of degradation.

2. Materials, Methods, and Analyses

The degradation process of reinforced concrete structures is affected by a multitude of physical-chemical, environmental, and structural variables, attributable to the design methodologies (with particular reference to the construction details), the construction itself, and the characteristics of the materials (particularly the concrete that should act as protection), as well as the management strategies and practices employed. Damage and consequent cracking are often due to corrosion. Such damage typically manifests itself with cracks in the concrete, which can lead to structures being out of service or disproportionate collapse, even in the absence of accidental or unforeseen events but only as a result of live loads (Figure 1).

Figure 1. Effects of degradation due to lack of maintenance: incipient collapse (on the (left)) and disastrous collapse (on the (right)).

However, it should be noted that, on the one hand, the concept of robust structures in structural design remains partially controversial. On the other hand, in more recent structural design codes and standards, the term “robustness” is used to indicate the ability of a system to resist damage under extreme loads and to prevent the progressive spread of (local) failure that would lead to disproportionate collapse. As Figure 1 shows, even the application of robustness concepts could not have avoided the disastrous collapse which, evidently, is of a progressive type due to lack of maintenance.

The classification and cataloguing of degradation on individual portions of structures, on individual structural elements, on entire structures and infrastructures directly affects the possibility of supporting a correct, effective and efficient decision-making process for infrastructure management and intervention planning.

The methodology employed in this study involves the utilisation of machine learning (ML) tools to analyse images, with the aim of achieving spatial recognition and assessing the severity of the damage. This is achieved through the implementation of object detection techniques. In Figure 2, the flowchart is reported.

Figure 2. The flowchart of the used methodology.

The study identifies and addresses the main issues that are useful for a broad application of artificial intelligence tools in monitoring the degradation of structures and infrastructures:

−

The construction of datasets representing real degradation images is achieved through collection, cataloguing and annotation based on specific skills and protocols.

−

Procedures and methods of expansion of datasets also through the implementation of data augmentation techniques which is a promising avenue for enhancing the diversity and reliability of images.

The used procedure and training of the deep learning model are based on the YOLO (You Only Look Once) algorithm. YOLO-type algorithms are often used to improve damage detection in civil infrastructure [43]. The network was trained using a manually constructed dataset, in which each image was annotated with bounding boxes identifying the degraded regions of the armatures. Subsequently, data augmentation techniques (rotations, brightness variations, geometric distortions) were applied to virtually increase the size and variety of the dataset, simulating different operating conditions.

In accordance with the prevailing conventions for the implementation of AI-based recognition techniques, the dataset was segmented into three distinct subsets, designated for utilisation in the three fundamental phases of the research process: training, validation and testing. The model training was performed on the training subset, while the validation set enabled the monitoring of performance during the learning process. The final test enabled the evaluation of the effectiveness of the model on a set of images not previously used, thus verifying the ability to generalise the results obtained in variable conditions.

The model uses a deep convolutional network (CNN) that automatically learns the visual characteristics associated with each degradation class, performing successive iterations on the data until it optimises the ability to discriminate between the different conditions.

The computer vision employed in this application enables the system to interpret images autonomously, simulating the human observation and recognition process. The system has been developed to utilise deep learning in order to visually correlate the presence of specific defects with the degree of reduction in the metal section.

The following schematic illustrates the multi-layer architecture implemented in the neural network model used for processing. The architecture comprises a series of sequential layers, each with a specific computational function within the data flow.

The images employed for the training, validation and testing of the model were processed and annotated using the Roboflow tool, which facilitated the efficient creation and management of the dataset through object detection techniques.

The images selected for analysis span a range of deterioration conditions, influenced by various factors including the exhibition environment, design parameters, and construction variables. These variables encompass structural geometry, materials quality, concrete mix design, transport methods, installation processes, and management of routine maintenance.

The management and preparation of the dataset was conducted using the Roboflow platform, a specialized tool for workflows in artificial vision projects. Roboflow (https://roboflow.com/) facilitates the uploading of data in a variety of formats (images, videos, PDF), in addition to the possibility of accessing and integrating already annotated public datasets. In this phase, a proprietary dataset consisting of approximately 234 images was uploaded, with each image representing a different scenario of pathological degradation of the reinforcement.

2.1. Classification and Augmentation

The classification of degradation was determined by the intensity of the corrosive phenomenon, the percentage of reduction in the section of the reinforcement, the spread of cracks and the extent of detachment of the concrete cover. The classification of these classes was determined by the following intervals of reduction in thickness of the reinforcement (Figure 3):

Figure 3. Example of the unique criterion for deterioration classification and bounding box for classifying the reduction in rebar diameter.

−

Class 0: cracks without any assessment of diameter reduction.

−

Class 1: reduction in diameter less than 10%.

−

Class 2: reduction in diameter ranging from 10% to 20%.

−

Class 3: reduction in diameter ranging from 20% and 30%.

Each image was then assigned to a specific degradation class, representing the state of conservation of the visible reinforcement. The decision to incorporate a critical class is intended to facilitate the identification of the most compromised structural conditions, necessitating urgent interventions.

In many cases, the same pile exhibits different exposure conditions on its various faces, thereby determining a non-homogeneous behaviour in the degradation processes. This, in turn, necessitates a localized analysis. This approach therefore facilitates the evaluation of corrosion distribution with greater precision, thus supporting the planning of targeted interventions.

The following data sets were utilised for the training of the YOLO model, following standard convention:

−

The training set, constituting 70% of the data, is utilised for the purpose of acquiring knowledge regarding the salient features of the data.

−

The validation set (20%) was utilised for the purpose of conducting a comparison between the neural network predictions and the reference images. This process was undertaken with the objective of evaluating the reliability of the neural network in accurately classifying the degradation levels.

−

The test set, constituting 10% of the dataset, is employed for the purpose of evaluating the generalisation capability of the model when confronted with unseen data.

In this particular instance, the restricted nature of the dataset necessitated meticulous attention during the training and validation phases, with the implementation of data augmentation techniques to enhance the diversity of the available sample sets.

This classified information formed the basis for the definition of the training set, acting as a reference tool for model training and ensuring that the network could learn to recognise, in a supervised manner, the visual patterns associated with the different stages of deterioration. Furthermore, some of the images from the aforementioned investigations were also utilised in the validation datasets. This was done in order to compare the model predictions with the diagnostic labels that had already been defined, and to evaluate the generalisation capacity on data that was known but not used in the training phase.

This approach enabled the integration of the diagnostic expertise acquired through engineering inspections with the capabilities of artificial intelligence technologies, thereby creating a system capable of automating the recognition of degradation phenomena. This system provides an objective and efficient means of supporting the evaluation process of the conservation conditions of reinforced concrete structures.

The input data underwent a preprocessing procedure to enhance the efficacy of model training and improve performance in the inference phase. This phase encompassed fundamental operations, including image normalization, resizing to standardized resolutions, and conversion to compatible formats.

Preprocessing was found to be of paramount importance in reducing the computational complexity of the model, thereby expediting the training process and enhancing the accuracy of predictions. This approach also contributed to enhancing the overall robustness of the neural network.

Subsequently, advanced data augmentation techniques were applied with the aim of artificially expanding the dataset and rendering the model more generalisable. The transformations performed include:

−

Random angular rotations to simulate changes in structure orientation;

−

Crops to focus on specific portions of the degradation zones;

−

Proportional resizing to handle objects of different scales;

−

Grayscale conversion, useful for testing the independence of the model from chromatic components;

−

Brightness and contrast variations, to increase resilience to uncontrolled lighting conditions;

−

Cuts and translations to simulate partial occlusions and camera movements.

These operations enabled the definition of an ordered sequence of transformations to be applied to the images, and to visualise their effects in real time. It is evident that the implementation of this structured procedure has enabled the creation of a more extensive, diverse and representative dataset, encompassing the various conditions observed within the structural field. This development has consequently led to an enhancement in the effectiveness of the training process and the model’s capacity to accurately identify the diverse levels of reinforcement degradation.

Finally, a total of 588 output images were generated, encompassing both the original images and those obtained through data augmentation. This constituted a more robust and balanced dataset for training the neural network (Figure 4).

Figure 4. Sample of segmentation and augmentation of images of dataset for (a) Bridge Piers foundation, (b) Bridge Piers, (c) Pier Cap.

2.2. Training and Validation

Necessity for this operation arose from the requirement to automate and simplify the workflow between the two environments, whilst ensuring the protection of data. The YOLO model was trained for 100 epochs, at the conclusion of which graphical visualisations were produced that document the performance of the evaluation metrics, such as the loss function, precision, recall value and accuracy, monitored during the training process (see Figure 5 and Figure 6). During the training phase of the object detection model, several performance metrics calculated on the validation set were monitored, in order to evaluate the generalisation capacity of the network. Figure 4 shows the evolution of the box loss, i.e., the error between the bounding boxes predicted by the model and the actual ones, during the training epochs. A progressive decrease in the loss is observed on the training data, a sign that the model is learning the relationships present in the data. However, the loss on the validation data remains almost constant throughout the training period, suggesting overfitting.

Figure 5. The horizontal axis shows the epochs (from 0 to 100), while the vertical axis represents the loss value. The curves indicate the loss trend on the training set (in blue) and the validation set (in orange), respectively.

Figure 6. Comparison of results of the final test between final test (on the (left)) and direct investigations (on the (right)).

The x-axis shows the number of epochs, each of which represents a complete cycle of forward propagation and back propagation on the data of the entire dataset training. The y-axis shows the values of the accuracy metrics which constitute quantitative indicators of the model performance. The examination of the metrics reveals that the model’s performance in terms of the classification and localisation of degradations in the reinforcements of reinforced concrete structures is not yet satisfactory.

Based on these considerations, the balance, quality and quantity of the dataset have been improved. Several training runs of the neural network were performed, varying the subdivisions between the training, validation and test sets, with the aim of improving the model’s ability to generalize to previously unseen scenarios.

3. Results

The results of the study appear to be highly encouraging. The model demonstrated a high degree of accuracy in cases that were analogous to the images employed during the training process. However, it also exhibited a certain degree of difficulty in recognizing scenarios that diverged from those utilized during the learning process. The findings confirmed the necessity for datasets of greater size, as well as the importance of integrating visual data with contextual engineering information, such as the age of the structure, the structural typology, the materials used, and the environmental conditions.

Among the various iterations, the most significant learning curves were selected, which describe the trend of accuracy and loss during training and validation. The results show:

−

A reduction in the distance between the training and validation curves, a sign that the model is tending to a better generalization;

−

An improvement in the stability in the localization of the bounding boxes (box loss);

−

A decrease in the classification error (class loss);

−

An increase in the ability of the model to correctly detect objects (objectness loss), with less marked oscillations compared to the previous version.

A comparison of the curves reveals the presence of residual overfitting, as evidenced by discrepancies between the training and validation metrics. This finding underscores the necessity to augment the variability of the dataset through the incorporation of additional images, thereby enhancing the robustness of the model.

During the system’s test phase, it demonstrated the capability to identify the degradation classes of the reinforcements, thereby achieving results that are in compliance with direct methodologies. Figure 6 depict some examples of processing produced by the model, with the predictions superimposed on the original images and direct comparison with the experimental data. In each case, the coordinates of the bounding boxes detected by the network and the degradation areas actually found on site will be compared. This will allow for the evaluation of spatial accuracy, the completeness of the detection and any discrepancies. This evaluation will guide subsequent activities of dataset extension and model calibration.

In order to enhance the efficacy of the inference phase, the threshold value of the confidence index in the predictions was appropriately reduced, thus increasing the sensitivity of the detection even in the presence of partial or small defects. This choice resulted in an enhancement of the capacity to discern potential instances of degradation, despite a marginal increase in the number of false positives, a phenomenon that can be effectively addressed through subsequent analyses. However, it should be noted that not all images in the dataset meet the necessary quality standards. In some cases, low resolution, the presence of visual noise or non-optimal lighting conditions have limited the reliability of the recognition process.

4. Discussion

The analysis of the model performance demonstrated a discrepancy between the accuracy achieved on the training set and that observed on the test set, thus indicating the presence of overfitting. This behaviour can be attributed to the limited size of the dataset, which does not allow the model to effectively generalise on conditions not seen during the training phase. Consequently, the model acquired overly specific characteristics of the training data, failing to develop sufficiently robust and transferable representations to different or more complex conditions.

The neural network exhibited a commendable capacity to adapt and identify the degradation classes that were most prevalent in the dataset, which included bridge piers. However, the performance is less satisfactory in terms of the recognition of degradation on structural components with more complex geometries, inhomogeneous surfaces and variable shooting conditions. In such cases, the model is unable to accurately identify discontinuities, indicating a limited capacity to generalise beyond the predominant scenarios present in the training dataset.

In any case, the preliminary assessment of the severity of the damage related to the reduction in the resistant sections of the reinforcements is satisfactory. The results obtained confirm the potential of the method in supporting more efficient and adaptable evaluation and maintenance processes. The results obtained demonstrated a high degree of accuracy in cases that were most similar to the images used for training. However, the model exhibited difficulties in recognizing scenarios that differed from those present in the training dataset, indicating a limited capacity for generalization.

In such contexts, the effective organisation of data plays a fundamental role, since the dataset is the basis for the training and performance of the model. In the present study, this approach allowed the provision of structured information to the system on both the position of the degraded elements and the degree of degradation, significantly improving the ability of the object detection model to automatically recognise the different forms and levels of degradation.

In order to achieve this objective, an effort was made to define and apply a unique criterion for the restitution of the frames of the images used in the datasets and for the classification of the various types of degradation of the reinforcements. This was done to ensure coherence and uniformity in the management and representation of the data. This systematic methodology has been proven to be effective in supporting the model’s ability to accurately identify deterioration phenomena in the analysed structures.

Existing structures frequently exhibit considerable heterogeneity in terms of structural and construction typology, maintenance status, conservation and operating conditions. This poses a significant challenge to the unification of methodologies and the direct utilisation of generalized models that lack sufficient validation. This variability is indicative of the inherent complexity of infrastructure projects yet concomitantly introduces inhomogeneities that impede the efficacy of artificial intelligence models. In this context, the employment of computer vision techniques, bolstered by deep learning models, has the potential to deliver solutions that are both faster and more efficient. Nevertheless, the full effectiveness of these measures is contingent upon the availability of a comprehensive database.

Another critical aspect pertains to the assessment of the state of degradation, which is often not directly and simply visible, particularly in the case of bridges and viaducts. While direct inspections remain a fundamental component of infrastructure management, they are inadequate to guarantee extensive and continuous control of the entire infrastructure heritage, especially considering the high number and morphological complexity of the elements to be monitored.

Despite significant advances in the field of automatic recognition of structural damage through the utilisation of deep learning-based models, the efficacy of these methodologies remains substantially constrained by the accessibility of training data. These data are frequently not only deficient in terms of quality but also in absolute terms. This constraint is identified as a key challenge in the development of robust and generalizable predictive models.

The scarcity of available datasets has the effect of diminishing the capacity of models to recognise damage scenarios in a range of material and environmental conditions. It is evident that there is an absence of a univocal criterion for the restitution of the frames of the images utilised in the datasets. The heterogeneity of the visual acquisition conditions renders the classification of the various types of degradation of the reinforcement difficult.

The evaluation of the model’s performance on unseen scenarios entails an examination of its generalisation capacity and its ability to avoid overfitting. The interpretation of the predictions made by artificial intelligence must be considered within a broader engineering context. This context encompasses the structural and environmental data of the analysed work. Improvement can be achieved through several factors, including:

−

Increased dataset size (fewer parameterized images) and its potential imbalance compared to current degradation classes.

−

Improved model architecture with better parameterization for the complexity of the civil engineering problem being investigated.

−

Improved spatial and chromatic diversity of images in the dataset, increasing the model’s ability to generalize results to extremely variable real conditions.

Finally, to broaden the research and its applications regarding the life cycle of structures, it is worth emphasizing that the study deals with existing reinforced concrete infrastructures, aged between approximately 40 and 60 years, for which the design did not consider the life cycle. For these types of structures, deterioration is undoubtedly increased by outdated design and construction procedures. Therefore, the use of appropriately validated artificial intelligence tools can be a key factor in defining management strategies and even in completely redefining the life cycle of the civil engineering works being studied. For newly designed and built structures, the life cycle is undoubtedly a fundamental element of design and management. For these types of civil engineering works, the use of artificial intelligence tools can provide important management support throughout the entire life cycle but must be specifically validated. New damage criteria and inspection methodologies will likely need to be defined.

5. Conclusions

In recent years, there has been growing interest in integrating AI-based tools into civil engineering, aiming to streamline the design, monitoring, maintenance, and management of structures and infrastructures. In particular, the automation of processes traditionally performed through manual activities (design, management, inspection, etc.) represents a concrete opportunity to reduce operating times and costs, while simultaneously improving the consistency, accuracy, and repeatability of technical assessments and the effectiveness and efficiency of procedures.

The adoption of AI is relatively easy and in its advanced stages for the design, construction, and management of new structures and infrastructures. This is thanks to the availability of consolidated procedures and experience, relatively clear regulations and guidelines, comprehensive design data, knowledge of the materials used, and the pursuit of structural and geometric regularity of structural elements. These factors favor the integration of AI models, coupled with the possibility of installing sensor devices from the initial stages for monitoring purposes.

The primary objective of this study is to develop a system based on intelligent algorithms to support the analysis of structural degradation, with reference to the reduction in the thickness of metal reinforcement due to corrosion in existing reinforced concrete structures. The focus is on road infrastructures, such as bridges and viaducts, which are strategic elements for mobility. Given the large number of existing structures, it is essential to have objective, reliable, and replicable tools for rapid and/or more in-depth assessments.

The results of the study are to be considered absolutely preliminary but of considerable interest and very promising. The results obtained highlight the importance of having large, representative, and balanced datasets in terms of image distribution across the different types of degradation (carbonation, cracking, washout, etc.) and structural features, as well as with respect to environmental conditions, the geometric characteristics of the elements, and the image acquisition methods. Ensuring an adequate and uniform number of examples for each category allows the model to learn comprehensively and correctly generalize to different deterioration scenarios.

The need to integrate visual information with more complex and comprehensive models is clearly emerging, capable of adapting to different structural typologies, environmental conditions (humidity, temperature, exposure to atmospheric agents), the presence of carbonation, the mechanical function it performs within the construction system, the age of the structure, and the materials used. These are key factors that influence the progression of degradation and should be appropriately considered in the training phase of artificial neural networks to improve their predictive effectiveness and model robustness in real and complex scenarios.

Furthermore, a significant contribution could be made using synthetic data generation techniques, facilitated by three-dimensional graphic engines. The employment of three-dimensional graphic engines would indeed facilitate the regulated generation of realistic images, distinguished by elevated visual coherence and a substantial array of degradation scenarios, in conjunction with meticulously modelled environmental conditions. This approach would facilitate the large-scale production of synthetic datasets, which would in turn be capable of more accurately representing the complexity of existing infrastructure heritage.

The real problem, however, is acquiring a large enough number of images to make the results reliable. Currently, data is still scarce and perhaps insufficient for optimal training. To achieve effective improvements, it is necessary to increase the number of images. In this regard, it should be noted that several freely available datasets are available online (for example, https://universe.roboflow.com/sanacon/reinforcement-400-images/dataset/2, accessed on 1 September 2025, https://universe.roboflow.com/project-ida/structural-defects-cmies/, accessed on 1 September 2025); on the other hand, it should be emphasized that these datasets cannot be used for the objectives of this study. In fact, it is not possible to label the images since it is not possible to determine the physical quantities expected from the damage classes. Some images may be suitable for the purposes of this study (for example, for classes 0 and 1) but would not be usable to build a new training, which would be heavily unbalanced and strongly conditioned on the lower classes. Effective improvement is possible only through specifically detected images associated with the physical parameters needed to characterize the damage. Further improvements are possible through specific image processing from which some of the aforementioned information can be obtained (element dimension, diameter, etc.) but in this case the characterization and labelling process is still more lengthy and not free from possible and recurring evaluation errors. The complexity of the problem can be reduced by:

−

Defining criteria for classifying degradation;

−

Defining methods, procedures, and tools for assessing the damage caused by degradation;

−

Defining image acquisition and classification procedures;

−

Defining and classifying structures typologically, also based on their environmental context.