An integrated vision material intelligence framework for infrastructure condition assessment and asset valuation

Highlights

• Proposed an integrated framework for infrastructure condition assessment that combines visual defect detection, concrete strength prediction, and analytical condition assessment.
• Developed the Hierarchical Edge-Aware Residual Network (HEAR-Net), incorporating Edge-Gated Activation and Global Average Pooling to enhance boundary-sensitive crack detection while maintaining a lightweight architecture.
• Integrated a regression-based material strength prediction module for estimating concrete compressive strength from physicochemical composition variables, achieving an R 2 of 0.84.
• Formulated an illustrative analytical condition-assessment methodology that combines visual damage characteristics and predicted material strength to derive a normalized structural health indicator.
• Demonstrated a unified analytical workflow integrating computer vision, material strength prediction, and illustrative condition assessment for infrastructure evaluation.

1 Introduction

The accelerating deterioration of global civil infrastructure represents a significant techno-economic challenge of the 21st century. A large portion of reinforced concrete structures are approaching or exceeding their design service life, and the cumulative effects of environmental weathering, fatigue loading, and chemical carbonation contribute to progressive structural degradation that affects both public safety and economic stability (Cui et al., 2021; Jiang et al., 2017). Recent assessments by the American Society of Civil Engineers (ASCE) estimate that the infrastructure rehabilitation investment gap in the United States alone exceeds $2.5 trillion, with comparable deficits reported across Europe and Asia. This situation reflects a persistent mismatch between the physical condition of infrastructure assets and the administrative frameworks used to manage and value them (Adeniyi et al., 2024).

Current practice relies heavily on manual visual inspection. While widely adopted, this approach is inherently subjective, labor-intensive, and susceptible to human error. Studies indicate that fatigue-induced observer bias may result in defect miss-rates of up to 30%, particularly in complex environments such as high-rise façades and large bridge decks (Calzada, 2024). Consequently, maintenance interventions are often reactive rather than preventive, leading to increased repair costs once deterioration becomes visibly advanced (Koch et al., 2015).

Beyond immediate safety risks, reliance on subjective assessment introduces ambiguity in the engineering evaluation of infrastructure assets. Decisions regarding maintenance, rehabilitation, and lifecycle planning frequently depend on visual observations or chronological age rather than an integrated understanding of structural condition and material performance (Paez-Perez and Sanchez-Silva, 2016; Egbo et al., 2021). Consequently, infrastructure management may not accurately reflect the combined influence of visible deterioration and underlying material characteristics, limiting the ability to perform comprehensive infrastructure condition assessment.

Advances in Computational Intelligence (CI) have enabled the development of automated tools across construction and infrastructure domains (Geirhos et al., 2018; Yaghoubzadeh-Bavandpour et al., 2022). In computer vision, Convolutional Neural Networks (CNNs) have demonstrated promising capabilities in automated surface crack detection (Hoskere et al., 2020). However, accurately identifying fine crack boundaries under heterogeneous surface conditions remains challenging because variations in texture, illumination, and background characteristics can obscure crack morphology and reduce the robustness of learned feature representations. Motivated by these challenges, the proposed Hierarchical Edge-Aware Residual Network (HEAR-Net) incorporates an Edge-Gated Activation mechanism that enhances boundary-sensitive feature learning, enabling improved representation of subtle crack characteristics while maintaining a lightweight convolutional architecture. Similarly, material science research has extensively investigated regression-based prediction of concrete compressive strength from physicochemical composition variables (Altuncı, 2024; Shaaban et al., 2025). Although these approaches have demonstrated encouraging predictive capability, many conventional regression models rely on simplified assumptions that may not adequately capture the nonlinear hydration behaviour governing concrete strength development, as described by Abram’s Law and related cement hydration mechanisms (Naderpour and Mirrashid, 2020; Ramadan, 2022). Moreover, these visual and material analyses are generally developed and evaluated independently, providing limited interaction between complementary sources of engineering information.

Although recent Structural Health Monitoring (SHM) research has increasingly incorporated artificial intelligence, optimization techniques, and multimodal sensing technologies, many existing frameworks primarily focus on individual sources of structural evidence, such as vision-based inspection, vibration monitoring, or material characterization, with the resulting analyses frequently remaining independent rather than being jointly interpreted within a unified infrastructure assessment workflow. Recent reviews have highlighted the rapid integration of machine learning and deep learning methodologies across mechanical and civil engineering SHM while emphasizing the continued need for more comprehensive and integrated intelligent assessment frameworks (Torti et al., 2022; Khatir et al., 2025). Consequently, there remains a need for analytical frameworks capable of systematically combining complementary information derived from visual deterioration assessment and material property estimation to support a more comprehensive understanding of infrastructure condition. Rather than introducing a fundamentally new SHM methodology, the present work investigates the feasibility of organizing these complementary analyses within a unified infrastructure condition-assessment framework.

To address this objective, this study proposes an integrated infrastructure condition-assessment framework that combines automated visual inspection, material property estimation, and analytical condition assessment within a unified analytical workflow. Unlike a conventional sequential combination of independent predictive models, the proposed framework is designed to organize the outputs of these complementary components into a common assessment pipeline, enabling their joint interpretation for infrastructure condition evaluation. The framework is centered on the following key elements:

• HEAR-Net (Hierarchical Edge-Aware Residual Network): A lightweight convolutional architecture incorporating an Edge-Gated Activation mechanism to enhance boundary-sensitive feature representation for crack detection under heterogeneous concrete surface conditions.
• Integrated Infrastructure Assessment Framework: A unified analytical workflow that combines automated crack detection with material property estimation to provide complementary engineering evidence for illustrative infrastructure condition assessment.

The primary contribution of this work is the development of an integrated infrastructure assessment framework that demonstrates how complementary outputs from computer vision and material property estimation can be jointly interpreted within a unified analytical workflow. The proposed framework was evaluated using publicly available visual and physicochemical datasets to investigate its feasibility for infrastructure condition assessment. As the constituent components were developed and evaluated using independent datasets, the present study demonstrates conceptual integration rather than validation on a unified real-world infrastructure dataset. Furthermore, the analytical condition assessment serves as an illustrative demonstration of the integration strategy and is not intended to represent a validated decision-support or asset valuation methodology. The remainder of this paper presents the proposed methodology, experimental evaluation, discussion of the framework’s limitations, practical implications, and future research directions.

2 Related works

2.1 The fragmentation of Structural Health Monitoring

The field of Structural Health Monitoring (SHM) has, over the years, been expanding through three separate branches in an almost parallel manner: one branch being Computer Vision for defect detection, the second branch being Material Science for strength characterization, and the third branch being Asset Management for financial valuation (Faris et al., 2025; Jaafaru, 2025). Even though improvements in Computational Intelligence (CI) have significantly advanced individual components of infrastructure assessment, recent studies have increasingly explored optimization-enhanced machine learning, deep learning, vibration-based structural diagnostics, and integrated data-driven frameworks for civil infrastructure monitoring (Khatir et al., 2026; Brahim et al., 2026; Rahmani et al., 2025; Khatir et al., 2025; Khatir et al., 2023). Nevertheless, a substantial gap remains in combining visual inspection, material characterization, and valuation analysis within a unified decision-support workflow (Angiulli and Versaci, 2025). The approaches adopted by most researchers so far have been to regard these different fields as completely independent of each other, thereby producing a segmented data environment wherein visual diagnostics are not used for material prognostics, and both of these branches are not used for economic decision-making (Pan et al., 2025; Li et al., 2022). Table 1 provides a comparative overview of representative data-driven studies, empirically illustrating the methodological fragmentation and the absence of integrated physics-aware and valuation-oriented frameworks.

TABLE 1. Comparative summary of recent data-driven SHM frameworks.

2.2 Visual defect detection: from filters to deep learning

Early automated crack detection systems were largely based on digital image processing (DIP) techniques. Among these, edge detection algorithms like Canny, Sobel, and Prewitt filters were commonly used due to their computational simplicity (Golding et al., 2022). Nevertheless, these gradient-based methods are extremely sensitive to environmental noise; hence, they often mistake shadows, oil stains, and surface roughness for structural defects, resulting in high False Positive Rates (FPR) that are not acceptable in real-world applications. In order to overcome the deficiencies of hard-coded filters, the research community has moved on to Machine Learning (ML) classifiers in the later stages of research (Figure 1). In paper employing Support Vector Machines (SVM) and Random Forests (RF), the use of hand-crafted feature extractors such as Histogram of Oriented Gradients (HOG) and Local Binary Patterns (LBP) was recommended (Jiang, 2022; Ali, 2023; Yuan et al., 2024). These methods enhanced detection accuracy in environments that were artificially set up, however, they are still intrinsically limited by the feature engineering quality. As pointed out in, manual feature descriptors are incapable of representing the complex, multi-scale topology of cracks on heterogeneous surfaces, like weathered asphalt or plaster-coated walls.

The paradigm shift toward Deep Learning (DL) has been the major source of the most significant improvements in performance (Liu et al., 2025). As a result, Convolutional Neural Networks (CNNs) have become the state-of-the-art, with models such as AlexNet, VGG-16, and ResNet-50 achieving accuracies of more than 90% on benchmark datasets (Zaheer et al., 2025). Additionally, semantic segmentation models like U-Net and object detection frameworks like YOLO (You Only Look Once) have recently been implemented to enable pixel-level crack localization. Nevertheless, these advances notwithstanding, the authors identify three critical limitations that exist in the current body of work:

Texture Bias: The majority of DL models have been trained on curated datasets that feature high contrast. These standard CNNs have a “Texture Bias” when they are tested on “noisy” textures such as aggregate-heavy pavement or stained concrete, thus they fail to recognize semantic edges as opposed to textural gradients.

The Black Box Problem: End-to-end learning architectures often lack interpretability. In safety-critical civil engineering, a model that detects a crack without explainable morphological justification is often rejected by regulators.

Data Scarcity: A significant portion of existing literature relies on small-scale datasets (< 2,000 images), limiting the model’s ability to generalize.

2.3 Material strength prediction: the linearity trap

Alongside visual inspection, the prediction of Concrete Compressive Strength (CCS) has been a major challenge of material informatics. Conventional methods consider empirical relationships based on Abram’s Law (1918), which indicates an inverse relationship between the water-cement ratio ( w / c ) and strength (Sobuz et al., 2024; Khan et al., 2025). Abram’s Law is a good starting point but it drastically simplifies the hydration kinetics of modern High-Performance Concrete (HPC), which has complicating factors like Fly Ash and Blast Furnace Slag as pozzolanic additives. In an attempt to explain these complicated factors, the researchers have widely used Linear Regression (LR) and Multiple Linear Regression (MLR) techniques (Ahmad et al., 2021; Ahmad et al., 2024). Nonetheless, these models miss the point of non-linear physicochemical interactions of concrete curing which is the main cause the researchers referred to. As an example, it has been demonstrated that linear models systematically underfit data in the domain of high-strength, because they cannot represent the log-like hardening curves and the threshold effects of superplasticizers.

Recently, several researches have gone farther to the point of using Artificial Neural Networks (ANNs) and Support Vector Regression (SVR) in their works to solve such a problem (Nguyen Anh et al., 2025). In general, ANNs have better potential in representing non-linear functions; however, they are prone to overfitting if the training dataset is small and tabular, which is often the case of construction projects data ( N < 500 ) (Mukherjee and Borthakur, 2025). Besides, typical “black box” ANNs do not integrate physicochemical constraints; they can statistically link strength with irrelevant features thereby going against the rules of material science (Karniadakis et al., 2021; Balasbaneh et al., 2025; Peng et al., 2026; Kaveh and Alhajj, 2024; Alkannad et al., 2025). Recent studies have increasingly emphasized incorporating domain knowledge and physics-informed learning strategies into concrete material characterization to improve interpretability and predictive robustness. Motivated by these developments, this study incorporates a regression-based material property estimation component to predict concrete compressive strength from physicochemical composition variables. Rather than serving as a standalone methodological contribution, the regression model complements the visual inspection module by providing material-related information within the proposed integrated infrastructure assessment framework. Recent optimization-enhanced machine learning studies have also demonstrated that hybrid metaheuristic optimization techniques can substantially improve predictive performance for engineering regression problems, including structural response, strain estimation, and material-related prediction tasks (Rahmani et al., 2025; Khatir et al., 2024; Bo et al., 2025). These studies highlight the growing importance of combining advanced optimization strategies with data-driven learning models to improve prediction accuracy and robustness across civil engineering applications.

2.4 The economic void: missing the “price” link

Most of SHM research ends at the stage of Detection or Prediction. Only a handful of papers try to convert these technical parameters into monetary ones. Present Life-Cycle Cost Analysis (LCCA) models are largely dependent on fixed depreciation schedules (e.g., ISO 15686) which determine value on the basis of chronological age rather than structural health in real-time.

According to recent reviews, “Digital Twin” technologies are getting developed, but they do not have a single algorithmic core that could facilitate the conversion of “Defect Density” to “Depreciation Rate” (Wang et al., 2025; Li et al., 2024; Lai et al., 2024). The lack of a dynamic valuation model results in the “Phantom Equity” issue, that is, the case when infrastructure assets are improperly valued because of the absence of detailed health data (Figure 2) (Kaveh and Alhajj, 2024; Alkannad et al., 2025; Sohaib et al., 2024a; Sohaib et al., 2024b).

3 Methodology

3.1 Data acquisition and compilation strategy

The proposed framework utilizes two complementary datasets corresponding to its visual and material analysis tasks: (i) a visual image dataset for automated surface crack detection using HEAR-Net and (ii) a physicochemical concrete dataset for material property estimation from concrete composition variables. The visual dataset enables the identification of surface deterioration, whereas the physicochemical dataset supports the estimation of concrete compressive strength. These datasets collectively provide the inputs required to investigate an integrated workflow that combines automated visual inspection with complementary material property estimation for infrastructure condition assessment.

3.1.1 Visual surface datasets

A visual dataset comprising 40,000 publicly available concrete surface images was employed for training and evaluation of the proposed HEAR-Net architecture. The dataset was obtained from the publicly available Surface Crack Detection dataset, which contains 20,000 crack (positive) images and 20,000 non-crack (negative) images. The dataset was generated from 458 high-resolution concrete surface photographs (4032 × 3024 pixels) acquired under diverse illumination conditions and surface finishes, thereby capturing substantial variability in concrete texture, crack morphology, and imaging conditions. Each image has a spatial resolution of 227 × 227 pixels with three RGB channels. Since the original dataset does not include artificially augmented samples, all images represent naturally acquired concrete surfaces. Prior to model development, the images were subjected to the preprocessing pipeline described in Section 3.2.

Each image belongs to one of two classes, C ∈ { 0,1 }, where Class 0 denotes non-cracked concrete surfaces and Class 1 denotes cracked concrete surfaces. The dataset maintains a balanced class distribution comprising 20,000 images in each category, thereby minimizing class imbalance during supervised model training (Equation 1).where N negative = 20,000 and N positive = 20,000. The binary class annotations supplied with the original dataset were retained throughout preprocessing and verified during dataset curation to ensure annotation consistency and conformity with the binary classification protocol adopted in this study.

N total = N negative + N positive = 40,000 (1)

3.1.2 Physicochemical concrete dataset

Concrete compressive strength prediction was performed using a publicly available physicochemical dataset comprising 1,030 concrete mixtures. Each sample contains eight descriptive variables characterizing material composition and curing age, together with the corresponding experimentally measured compressive strength. This dataset complements the visual inspection component by enabling the estimation of structural material performance from mixture characteristics.

3.1.2.1 Feature space definition

The physicochemical dataset is represented as a feature matrix X mat ∈ R 1030×8, where each row corresponds to an individual concrete mixture. The input vector comprises seven material composition variables and one temporal variable representing curing age, while the target variable Y strength denotes the measured compressive strength expressed in MPa. A summary of the input variables is provided in Table 2.

TABLE 2. Data dictionary of the physicochemical input vector.

3.1.2.2 Characteristics of the target variable

Concrete compressive strength is governed by complex interactions among cementitious materials, aggregates, water content, chemical admixtures, and curing duration. These relationships are inherently nonlinear and therefore cannot be adequately represented by a single global linear model. Accordingly, a regression-based material strength prediction approach was adopted to estimate compressive strength from physicochemical composition variables while preserving the interpretability of tree-based learning.

3.1.2.3 Data partitioning

Both the visual and physicochemical datasets were partitioned using a stratified hold-out strategy comprising 70% training, 15% validation, and 15% testing subsets. The validation subset was used exclusively for hyperparameter optimization, whereas the independent testing subset remained completely unseen throughout model development and was reserved solely for final performance evaluation.

Dataset partitioning was performed prior to model training, data augmentation, and all model-specific preprocessing operations using a fixed random seed to ensure reproducibility. Data augmentation was applied exclusively to the training subset, while the validation and testing subsets remained unchanged throughout model development and evaluation. As the publicly available visual datasets do not provide consistent structure-level or parent-photograph identifiers, partitioning was performed at the image level. Consequently, explicit photograph-level separation between the training, validation, and testing subsets could not be verified, and the possibility of neighboring image patches originating from the same parent photograph cannot be entirely excluded. Nevertheless, validation and testing samples were strictly excluded from model training, data augmentation, and hyperparameter optimization. Accordingly, the reported performance should be interpreted within the characteristics of the benchmark dataset, while future work will investigate photograph-level and infrastructure-level partitioning using comprehensive field datasets to further assess model generalization.

3.2 Statistical preprocessing and feature validation

Prior to model development, exploratory statistical analyses were conducted to characterize both the visual and physicochemical datasets and to examine whether their intrinsic statistical properties supported the design of the proposed framework. For the visual dataset, hypothesis testing was performed to investigate whether image-derived characteristics were associated with crack labels. For the physicochemical dataset, Principal Component Analysis (PCA) was employed exclusively as an exploratory technique to examine feature relationships and variance distribution. These statistical analyses were used solely for dataset characterization and methodological justification and were not incorporated directly into model training.

3.2.1 Statistical analysis of visual features

The visual dataset was first analyzed to determine whether global pixel intensity alone could distinguish cracked and intact infrastructure surfaces. The null hypothesis ( H 0 ) assumed that the mean pixel intensity distributions of the two classes were statistically identical.

3.2.1.1 One-way analysis of variance (ANOVA)

A one-way ANOVA was performed to compare the between-class and within-class variance. The F-statistic was computed as Equation 2 where K denotes the number of classes and N is the total number of observations.

F = ∑ i n i Y ̄ i − Y ̄ 2 / K − 1 ∑ i , j Y i j − Y ̄ i 2 / N − K (2)

The analysis yielded an F-statistic of 0.6962 with a corresponding p-value of 0.4107. Since p > 0.05, the null hypothesis was not rejected, indicating that global pixel intensity alone does not significantly distinguish cracked and intact surfaces within the evaluated dataset.

The inability of global pixel intensity to discriminate between the two classes suggests that crack identification depends primarily on localized structural characteristics rather than overall image brightness. This finding motivated the incorporation of the proposed Edge-Gated Activation mechanism, which emphasizes local structural discontinuities during feature extraction rather than relying solely on global intensity information.

3.2.1.2 Chi-Square test of independence

To investigate whether localized edge information was associated with crack occurrence, a Chi-Square test of independence was performed using gradient magnitude responses and the corresponding crack labels. The test statistic is defined as Equation 3 where O i j and E i j denote the observed and expected frequencies, respectively.

χ 2 = ∑ O i j − E i j 2 E i j (3)

The resulting p-value was less than 0.001, indicating a statistically significant association between localized gradient responses and crack labels. Unlike global intensity information, gradient-based structural features exhibit strong discriminative characteristics for crack identification. This statistical evidence provides methodological support for incorporating the proposed Edge-Gated Activation mechanism within HEAR-Net to emphasize boundary-sensitive feature representations during feature extraction.

Table 3 summarizes the statistical analyses performed on the visual dataset.

TABLE 3. Summary of statistical analyses performed on the visual dataset.

3.2.2 Exploratory analysis of physicochemical features

The physicochemical dataset contains eight continuous variables describing concrete composition and curing age. Since several variables exhibit varying degrees of correlation, Principal Component Analysis (PCA) was performed as an exploratory statistical technique to investigate the variance structure of the feature space. PCA was used solely for exploratory analysis and visualization and did not influence model training, optimization, or feature selection.

3.2.2.1 Principal component analysis

Following feature standardization, the covariance matrix was computed and decomposed into principal components. The transformed feature representation is given bywhere X denotes the standardized feature matrix and W contains the principal component vectors.

T = X W ,

Figure 3 illustrates the distribution of the concrete mixtures in the reduced two-dimensional principal component space. Although several localized clusters are observed, considerable overlap remains among samples with different compressive strength values, indicating that the underlying relationships between the physicochemical variables and compressive strength are not adequately represented by simple low-dimensional linear projections. This observation further supports the use of nonlinear regression rather than relying solely on linear feature transformations.

The first three principal components explained 63.4% of the total variance (Figure 4), indicating that a substantial proportion of the dataset variability can be represented within a reduced-dimensional space. However, the cumulative explained variance increases gradually across the remaining principal components, reaching approximately 100% only after all eight variables are retained. This trend suggests that the physicochemical information is distributed across multiple features rather than concentrated in only a few dominant components. Therefore, all eight input variables were retained throughout model development to preserve the complete physicochemical feature space and its nonlinear relationships.

The corresponding principal component loadings are presented in Table 4.

TABLE 4. Principal component loadings for the first three principal components.

Although PCA provides useful insight into the statistical structure of the dataset, dimensionality reduction was intentionally not incorporated into model development. Instead, all eight physicochemical input variables were retained to preserve the complete feature space and its underlying nonlinear relationships for material strength prediction.

3.2.3 Image preprocessing and data augmentation

Prior to model training, all images were normalized using min-max normalization (Equation 4),

I norm = I − I min I max − I min , (4)

Which scales pixel intensities to the interval [0, 1]. Normalization ensures consistent numerical scaling across all input images and contributes to stable optimization during network training.

To improve robustness to moderate geometric variations encountered during practical infrastructure inspections, random affine transformations were applied exclusively to the training dataset after data partitioning. The augmentation strategy included random rotations within ± 15 ° together with translational perturbations (Equation 5),

x ′ y ′ = cos ⁡ θ − sin ⁡ θ sin ⁡ θ cos ⁡ θ x y + t x t y , (5)

There by improving model robustness while preserving the underlying crack morphology. The validation and testing datasets remained unchanged throughout model development and evaluation to prevent data leakage and ensure an unbiased assessment of model generalization.

3.3 Visual defect identification via HEAR-Net

HEAR-Net was developed for automated binary classification of cracked and intact infrastructure surfaces. Unlike conventional object recognition tasks, crack detection requires the identification of thin, irregular, and spatially localized discontinuities that frequently appear under heterogeneous surface textures, varying illumination conditions, shadows, and background noise. These characteristics make the preservation of fine boundary information particularly important during feature extraction. To address these challenges, the proposed HEAR-Net integrates a hierarchical convolutional backbone with a novel Edge-Gated Activation mechanism and Global Average Pooling (GAP), enabling efficient boundary-aware feature learning while maintaining a lightweight network architecture.

3.3.1 Network architecture

The input to HEAR-Net is an RGB image represented aswhere H and W denote the image height and width, respectively.

X ∈ R H × W × 3 ,

HEAR-Net consists of four sequential convolutional feature extraction blocks. Each block comprises a 3 × 3 convolutional layer, batch normalization, the proposed Edge-Gated Activation function, and a 2 × 2 max-pooling operation. The number of convolutional filters increases progressively from 32, 64, 128, and 256, enabling the network to learn hierarchical feature representations ranging from low-level edge responses to high-level semantic crack patterns.

Following hierarchical feature extraction, the final convolutional feature maps are aggregated using Global Average Pooling (GAP), followed by a dropout layer with a rate of 0.5 and a single sigmoid output neuron for binary classification. Replacing conventional fully connected layers with GAP substantially reduces the number of trainable parameters while preserving the discriminative information learned by the convolutional backbone, thereby improving computational efficiency and reducing the risk of overfitting.

The hierarchical organization of convolutional blocks enables HEAR-Net to progressively capture increasingly abstract structural features, while the Edge-Gated Activation mechanism preserves boundary-sensitive information throughout the feature extraction process. The resulting compact architecture combines efficient feature learning with robust crack discrimination across heterogeneous infrastructure surface conditions.

3.3.2 Edge-gated activation mechanism

Accurate crack detection depends heavily on preserving weak edge responses that often correspond to thin structural discontinuities. Conventional activation functions process feature responses uniformly and may suppress these subtle boundary characteristics during feature propagation. To address this limitation, HEAR-Net incorporates the proposed Edge-Gated Activation mechanism, which adaptively modulates convolutional responses according to the local gradient magnitude estimated using the Sobel operator.

The Edge-Gated Activation is defined as Equation 6 wheredenotes the convolution output at layer l, ⊙ represents element-wise multiplication, ∇ X ( l − 1 ) is the Sobel gradient magnitude computed from the preceding feature map, σ ( ⋅ ) denotes the sigmoid activation function, and λ is a learnable scaling parameter controlling the contribution of gradient information.

Φ edge Z l = R e L U Z l ⊙ σ λ ∇ X l − 1 2 , (6)

Z l = W l ∗ X l − 1 + b l ,

The resulting gating function increases the contribution of feature responses associated with strong structural boundaries while suppressing comparatively uninformative regions. Consequently, the convolutional backbone is encouraged to learn boundary-sensitive feature representations that are particularly suited to crack localization under heterogeneous infrastructure surface conditions.

3.3.3 Model training

HEAR-Net was formulated as a supervised binary classification problem and optimized using the Binary Cross-Entropy (BCE) loss function (Equation 7),where y i denotes the ground-truth class label and p i represents the predicted probability of the cracked class for sample i.

L θ = − 1 N ∑ i = 1 N y i ⁡ log p i + 1 − y i log 1 − p i , (7)

Model optimization was performed using the Adam optimizer with an initial learning rate of 0.001, momentum parameters β 1 = 0.9 and β 2 = 0.999, and a batch size of 32 samples. Training was conducted for a maximum of 50 epochs using early stopping with a patience of 10 epochs based on the validation loss. The network parameters corresponding to the minimum validation loss were retained for subsequent evaluation on the independent testing dataset. This training strategy promotes stable convergence while reducing the likelihood of overfitting and ensuring unbiased performance evaluation.

All deep learning experiments were conducted using the Google Colab platform with GPU acceleration provided by an NVIDIA Tesla T4 GPU (16 GB VRAM). Model training required approximately 2.5 h under the adopted experimental configuration. Standard Google Colab virtual CPU resources and runtime memory were used for data preprocessing and execution. The total energy consumption recorded during model training was approximately 0.17 kWh. The inference performance of the evaluated architectures, including per-batch inference time and throughput, is reported in Table 9.

3.4 Material strength prediction

Material property estimation constitutes a complementary component of the proposed infrastructure assessment framework. In this study, a regression-based learning approach was employed to estimate concrete compressive strength from physicochemical composition variables. Recursive partitioning was selected because of its ability to model nonlinear relationships commonly observed between concrete composition and compressive strength while maintaining model interpretability. The predicted strength values complement the visual inspection results within the proposed framework and are not intended to replace standardized laboratory compressive strength testing.

The objective of this component is to estimate concrete compressive strength from the eight physicochemical input variables described in Section 3.1.2. The resulting strength estimates provide complementary material property information within the proposed infrastructure assessment framework and are not intended to replace standardized laboratory compressive strength testing.

3.4.1 Regression formulation

Let the training dataset be represented aswhere x i ∈ R 8 denotes the physicochemical feature vector and y i represents the corresponding measured compressive strength.

D = x i , y i i = 1 N ,

The material strength prediction component employs a regression-tree-based learning strategy that recursively partitions the physicochemical feature space. At each decision node, candidate splits are evaluated by minimizing a weighted variance objective (Equation 8),where R L and R R denote the left and right partitions generated by a candidate split, y ̂ L and y ̂ R represent the mean target values within the respective partitions, Φ chem ( j ) denotes a feature-weighting term incorporated during split evaluation to guide recursive partitioning based on physicochemical variables, and λ controls the contribution of the weighting term.

L split j , s = N L N ∑ i ∈ R L y i − y ̂ L 2 + N R N ∑ i ∈ R R y i − y ̂ R 2 + λ Φ chem j , (8)

At each node, the candidate split yielding the minimum weighted impurity is selected, recursively partitioning the feature space into progressively more homogeneous subsets. This hierarchical partitioning enables the regression model to capture nonlinear relationships between concrete composition variables and compressive strength while preserving the interpretability of decision-tree-based learning. Tree construction continues until the predefined stopping criteria are satisfied, after which pruning is performed to improve generalization and reduce model complexity.

3.4.2 Model configuration

To improve generalization and reduce model complexity, cost-complexity pruning was employed after tree construction. Beginning with the fully expanded regression tree, progressively smaller subtrees were generated by removing branches that contributed least to reducing prediction error. The optimal subtree was selected using the cost-complexity objective (Equation 9)where R ( T ) denotes the regression error, | T | represents the number of terminal nodes, and α is the pruning parameter controlling the trade-off between predictive accuracy and model complexity.

R α T = R T + α | T | , (9)

The pruning parameter and the remaining hyperparameters were determined using 10-fold cross-validation on the training dataset. The final model configuration is summarized in Table 5.

TABLE 5. Final hyperparameter configuration of the material strength prediction model.

3.4.3 Baseline regression

To provide a reference for evaluating the material strength prediction component, a conventional Linear Regression model was employed as the baseline. Linear Regression was selected because it represents a widely used statistical approach for modeling continuous target variables under the assumption of a global linear relationship between the predictors and the response.

The regression model and the baseline Linear Regression model were trained and evaluated using the same data partitioning strategy and preprocessing protocol described in Sections 3.1 and 3.2. Performance was assessed using the coefficient of determination ( R 2 ) and the Root Mean Squared Error (RMSE), enabling a consistent comparison between the adopted regression model and the conventional linear baseline. The quantitative evaluation is presented in Section 4.

3.5 Illustrative analytical condition assessment

To demonstrate how complementary outputs from the visual inspection and material strength prediction components may be interpreted together, an illustrative analytical condition assessment framework is presented. The framework combines the crack probability estimated by HEAR-Net with the predicted concrete compressive strength to derive a normalized structural health indicator. The analytical formulation is intended solely to demonstrate one possible integration of the generated outputs and has not been calibrated, optimized, or experimentally validated for engineering decision-making, infrastructure valuation, or remaining-service-life estimation. Accordingly, the analytical coefficients introduced in the following formulation are illustrative and are used only to demonstrate the proposed analytical workflow.

3.5.1 Illustrative condition assessment model

Let C 0 denote a normalized reference condition index and C ( t ) represent the corresponding analytical condition index at time t. The illustrative condition index is expressed as Equation 10

C t = C 0 ⁡ exp − ∫ 0 t λ τ d τ , (10)

For illustrative purposes, the analytical deterioration rate is expressed as Equation 11

λ t = λ base + α Φ edge x + β σ design − σ pred σ design 2 . (11)

Here,

• C 0 denotes the normalized reference condition index;
• λ base represents an illustrative baseline analytical deterioration coefficient introduced solely to establish the analytical formulation and was not calibrated using field observations or engineering inspection records;
• Φ edge ( x ) denotes the crack probability predicted by HEAR-Net;
• σ design represents the design compressive strength;
• σ pred denotes the predicted concrete compressive strength; and
• α and β are illustrative weighting coefficients governing the relative contributions of visual deterioration and material strength. These coefficients were introduced to demonstrate the analytical integration strategy and were not obtained through calibration, optimization, literature-derived parameterization, or experimental estimation.

Within this illustrative formulation, larger crack probabilities and greater deviations between the predicted and design compressive strengths produce lower condition index values. The coefficients were intentionally left uncalibrated because the objective of this component is to demonstrate a possible analytical integration of complementary visual and material information rather than to develop a predictive deterioration model. Calibration of the analytical coefficients using longitudinal inspection records, maintenance histories, and infrastructure monitoring data remains an important direction for future research.

3.5.2 Structural health indicator

For ease of interpretation, the analytical condition index is normalized as (Equation 12)where

H = C t C 0 , (12)

0 ≤ H ≤ 1 .

Higher values of H correspond to comparatively better estimated condition, whereas lower values indicate progressively greater analytical deterioration according to the illustrative formulation. The Structural Health Indicator is intended solely as a normalized analytical measure for comparative evaluation and should not be interpreted as a direct measure of structural safety or serviceability.

3.5.3 Illustrative condition categories

The normalized Structural Health Indicator, together with the crack probability estimated by HEAR-Net and the predicted compressive strength, is used to illustrate representative condition categories, as summarized in Table 6. These categories demonstrate one possible interpretation of the combined outputs generated by the proposed framework and are not intended as engineering recommendations or operational decision thresholds.

TABLE 6. Illustrative condition categories derived from the integrated analytical framework.

The threshold values presented in Table 6, together with the analytical coefficients used in the illustrative condition assessment model, are provided solely to demonstrate the proposed analytical integration framework. They were not calibrated, optimized, or experimentally validated using longitudinal field observations, engineering inspection records, or asset management databases and therefore should not be interpreted as universally applicable engineering parameters or operational decision thresholds. Future work will investigate the calibration and validation of both the analytical coefficients and condition thresholds using application-specific infrastructure datasets.

4 Results and discussions

4.1 Performance analysis of the visual defect detection module (HEAR-Net)

4.1.1 Quantitative classification performance

The performance of HEAR-Net was evaluated on an independent testing dataset to assess its capability for binary crack classification. Model performance was quantified using Accuracy, Precision, Recall (Sensitivity), Specificity, F1-score, Balanced Accuracy, Receiver Operating Characteristic Area Under the Curve (ROC-AUC), Precision-Recall Area Under the Curve (PR-AUC), Matthews Correlation Coefficient (MCC), Cohen’s Kappa, Log Loss, and False Negative Rate (FNR). These complementary evaluation metrics provide a comprehensive assessment of both the overall predictive performance and the balance between positive and negative classifications.

Figure 6 presents the Receiver Operating Characteristic (ROC) curve of HEAR-Net evaluated on the independent testing dataset. The curve closely follows the upper-left boundary of the ROC space, yielding an area under the curve (AUC) approaching unity. This indicates excellent class separability, demonstrating that the proposed architecture consistently distinguishes cracked from intact surfaces across a wide range of decision thresholds. The high AUC further confirms that the strong classification performance observed in the confusion matrix is not dependent on a single operating threshold but reflects the intrinsic discriminative capability of the learned feature representations.

Figure 5 illustrates the confusion matrix obtained on the independent testing dataset. HEAR-Net correctly classified 2,997 negative and 2,997 positive samples, while producing only three false positive and three false negative predictions. The balanced distribution of classification errors demonstrates that the proposed model maintains comparable predictive performance across both classes while minimizing missed crack detections.

The quantitative results summarized in Table 7 indicate that HEAR-Net achieved an overall classification accuracy of 99.90%, with precision, recall, specificity, balanced accuracy, and F1-score all reaching 99.90%. The ROC-AUC and PR-AUC values of 0.999993 demonstrate excellent separability between cracked and intact surfaces over a wide range of decision thresholds, confirming the robustness of the proposed classifier.

TABLE 7. Performance metrics of HEAR-Net on the independent testing dataset.

Although HEAR-Net achieved exceptionally high quantitative performance on the benchmark dataset, these results should be interpreted within the characteristics of the evaluated dataset. The reported metrics reflect performance on a standardized binary crack classification benchmark under the adopted experimental protocol and should not be interpreted as evidence of comprehensive real-world generalization. As discussed in Section 3.1.2(C), photograph-level separation could not be explicitly verified because the publicly available dataset does not provide parent-photograph identifiers. Consequently, additional validation using independent field datasets with photograph-level or infrastructure-level partitioning is necessary to further assess the generalization capability of HEAR-Net under real-world inspection conditions.

The extremely low false negative rate of 0.10% indicates that only three cracked samples were incorrectly classified as non-cracked during independent evaluation. Since false negatives represent missed structural defects that may remain undetected during inspection, minimizing this error is particularly important for infrastructure monitoring applications. The observed confusion matrix therefore demonstrates that HEAR-Net provides highly reliable crack detection while maintaining a balanced trade-off between sensitivity and specificity under the evaluated experimental conditions.

The exceptionally low false positive and false negative counts demonstrate that HEAR-Net maintains balanced discrimination across both classes. This balance is particularly important for infrastructure inspection, where missed crack detections may have greater practical consequences than isolated false alarms. The observed performance therefore indicates that the proposed architectural design effectively captures crack-specific visual characteristics while maintaining robust generalization on previously unseen samples.

4.1.2 Training behaviour and model convergence

The convergence characteristics of HEAR-Net were examined using the training and validation accuracy and loss recorded throughout the optimization process. Monitoring these learning curves provides insight into the optimization stability, convergence behaviour, and generalization capability of the proposed architecture.

Figure 7 illustrates the training and validation accuracy obtained during optimization. Training accuracy increased steadily and converged near unity, while validation accuracy remained consistently high despite minor epoch-to-epoch fluctuations. The close agreement between the two curves throughout training indicates stable optimization with no sustained divergence between training and validation performance. The transient variations observed in the validation accuracy are likely attributable to the relatively small validation subset and do not indicate persistent overfitting, as both curves converge to comparable final performance. The rapid convergence and close agreement between the training and validation curves indicate that the proposed architecture learns stable feature representations without exhibiting significant optimization instability or overfitting. The absence of large oscillations throughout training suggests that the Edge-Gated feature extraction and Global Average Pooling classifier jointly promote efficient optimization while maintaining good generalization.

4.1.3 Explainability analysis

To improve the interpretability of the proposed deep learning framework, Gradient-weighted Class Activation Mapping (Grad-CAM) was employed to visualize the image regions contributing most strongly to the classification decisions. This analysis provides qualitative evidence that the learned representations are primarily driven by structural crack characteristics rather than irrelevant background information.

Figure 8 presents representative Grad-CAM visualizations generated by HEAR-Net for cracked and intact surface samples. The highest activation regions consistently coincide with the visible crack boundaries, while comparatively little activation is observed over homogeneous background textures. This localization behaviour indicates that the proposed architecture bases its predictions primarily on structurally meaningful defect patterns rather than unrelated surface characteristics. The observed attention distributions further provide qualitative evidence that the Edge-Gated Activation mechanism promotes boundary-sensitive feature learning, improving the interpretability of the classification decisions.

The Grad-CAM visualizations demonstrate that HEAR-Net consistently concentrates its attention on crack boundaries rather than surrounding concrete textures or background regions. This behaviour indicates that the proposed Edge-Gated Activation successfully guides feature extraction toward structurally meaningful regions, providing qualitative evidence that the network bases its decisions on physically relevant defect characteristics instead of spurious image features.

4.1.4 Ablation study

To quantify the individual contributions of the proposed architectural components, an ablation study was conducted using four network configurations evaluated under identical data partitioning, preprocessing, hyperparameter settings, and training protocol. The evaluated configurations comprised: (i) a conventional baseline CNN employing ReLU activation with a Flatten classifier, (ii) an EdgeGate-only variant replacing the ReLU activation with the proposed Edge-Gated Activation while retaining the Flatten classifier, (iii) a GAP-only variant replacing the Flatten classifier with Global Average Pooling while preserving the conventional ReLU activation, and (iv) the complete HEAR-Net architecture integrating both Edge-Gated Activation and Global Average Pooling.

The objective of this study was to isolate the contribution of each architectural component and determine whether the performance improvements arise from the individual modules or from their combined integration.

Table 8 demonstrates that each architectural modification contributes positively to classification performance. Replacing the conventional ReLU activation with the proposed Edge-Gated Activation consistently improves discrimination by emphasizing boundary-sensitive crack features during feature extraction. Similarly, substituting the Flatten classifier with Global Average Pooling reduces parameter redundancy while encouraging the learning of globally representative feature maps. Although each modification independently improves the baseline architecture, the complete HEAR-Net consistently achieves the highest performance across all evaluation metrics, demonstrating that the two components operate synergistically rather than providing isolated improvements.

TABLE 8. Ablation analysis of the proposed HEAR-Net architecture under identical training and evaluation settings.

Bold values indicate the best performance for the corresponding metric among the evaluated models.

To further investigate the optimization behaviour of the evaluated architectures, the corresponding training and validation convergence curves are presented in Figures 9–12. These curves provide complementary evidence regarding convergence stability, optimization efficiency, and generalization behaviour beyond the final quantitative metrics.

The convergence curves indicate that all four architectures successfully learned the binary crack classification task; however, the complete HEAR-Net exhibits faster convergence, consistently higher validation accuracy, and lower validation loss throughout training. The Edge-Gated Activation facilitates the learning of discriminative boundary-aware representations, whereas Global Average Pooling promotes stable optimization by reducing the dependence on high-dimensional fully connected representations. When combined, these components produce smoother optimization trajectories and superior generalization performance compared with the remaining variants. Collectively, the ablation study demonstrates that the performance gains achieved by HEAR-Net result from the complementary interaction between Edge-Gated Activation and Global Average Pooling rather than from either component alone.

4.1.5 Computational efficiency analysis

In addition to predictive performance, practical infrastructure inspection systems require models that achieve high accuracy while maintaining computational efficiency. Therefore, the computational characteristics of the evaluated architectures were assessed in terms of parameter count, trainable parameters, model size, inference latency, and throughput under an identical experimental environment.

Table 9 summarizes the computational characteristics of the evaluated architectures. Replacing the conventional Flatten classifier with Global Average Pooling reduces the number of trainable parameters from over 13.2 million to approximately 0.46 million, representing a reduction of approximately 96.6%, while simultaneously decreasing the model size from approximately 151.5 MB–5.3 MB. These reductions are achieved without compromising predictive performance, demonstrating the effectiveness of Global Average Pooling as a compact classification strategy.

TABLE 9. Computational efficiency comparison of the evaluated network architectures.

Bold values indicate the best performance for the corresponding metric among the evaluated models.

The proposed Edge-Gated Activation introduces only four additional trainable parameters compared with the GAP-only variant, indicating that the enhanced boundary-aware feature extraction is achieved with negligible architectural overhead. Although the complete HEAR-Net exhibits a higher inference latency than the GAP-only configuration due to the additional edge-aware computations, it consistently provides the highest classification performance while maintaining the same compact model size and parameter count.

Overall, the computational analysis demonstrates that HEAR-Net achieves an effective balance between predictive accuracy and computational efficiency. Compared with the conventional CNN architectures, the proposed network substantially reduces model complexity while delivering superior classification performance, supporting its suitability for practical automated infrastructure inspection applications where compact model size and reliable predictive capability are essential.

4.1.6 Benchmark comparison with standard CNN architectures

To further validate the proposed architecture, HEAR-Net was benchmarked against representative convolutional neural network models, including a conventional baseline CNN, MobileNetV2, and ResNet50. All models were trained and evaluated using identical data partitions and optimization settings to ensure a consistent and unbiased comparison.

The benchmark comparison demonstrates that HEAR-Net achieves the highest overall classification performance among the evaluated architectures while simultaneously maintaining a comparatively lightweight model representation. Although MobileNetV2 and ResNet50 are highly effective general-purpose image classification networks, their architectural designs are intended for broad visual recognition tasks rather than the localization of fine-scale structural defects.

In contrast, HEAR-Net explicitly incorporates Edge-Gated Activation to strengthen crack-boundary representations and employs Global Average Pooling to improve feature aggregation while reducing parameter redundancy. These task-oriented architectural modifications enable the proposed network to capture subtle crack morphology more effectively than generic CNN backbones, thereby improving classification accuracy without substantially increasing computational complexity.

When considered together with the ablation study and computational efficiency analysis, the benchmark results indicate that the superior performance of HEAR-Net originates from its architectural design rather than from increased network depth or parameter count. The proposed framework therefore achieves an effective balance between predictive accuracy, optimization stability, computational efficiency, and model interpretability, making it well suited for automated infrastructure condition assessment in practical engineering environments.

4.2 Validation of material strength prediction

The predictive capability of the material strength prediction component was evaluated using the Concrete Compressive Strength dataset comprising 1,030 concrete mixtures. A Linear Regression model was employed as the baseline because it provides a simple linear approximation against which the adopted regression model can be assessed. Model performance was evaluated using the coefficient of determination ( R 2 ) and the Root Mean Squared Error (RMSE), which together quantify the goodness of fit and prediction error.

4.2.1 Regression performance

Figure 13 illustrates the relationship between the measured and predicted compressive strengths obtained using the Linear Regression baseline.

The Linear Regression baseline achieved an R 2 value of 0.61, indicating moderate agreement between the predicted and measured compressive strengths. As shown in Figure 13, the predictions exhibit increasing dispersion from the identity line as compressive strength increases, suggesting that a simple linear model is insufficient to represent the nonlinear relationships present within the evaluated concrete mixtures.

The adopted regression model was subsequently evaluated using the same training and testing protocol. Table 10 summarizes the quantitative comparison between the baseline model and the adopted regression approach.

TABLE 10. Performance comparison of regression models.

The adopted regression model achieved an R 2 value of 0.84 with an RMSE of 4.20 MPa, demonstrating a substantial improvement over the Linear Regression baseline. The results indicate that the nonlinear regression strategy more effectively models the complex interactions among the physicochemical properties governing concrete compressive strength. These predictions provide complementary material property information within the proposed infrastructure assessment framework while maintaining model interpretability.

4.3 General discussion

This study presents an integrated framework for infrastructure condition assessment that combines deep learning-based visual defect detection, material strength prediction, and an illustrative analytical condition assessment within a unified workflow. The proposed framework demonstrates how complementary sources of structural information can be systematically organized and jointly interpreted to support infrastructure condition assessment while allowing each component to operate independently. It should be emphasized, however, that the visual inspection, material strength prediction, and analytical condition assessment components were developed and evaluated using independent publicly available datasets. Consequently, the present study demonstrates the feasibility of conceptual integration rather than validation of a unified end-to-end Structural Health Monitoring (SHM) system on a single real-world infrastructure dataset.

The experimental results demonstrate that HEAR-Net provides accurate and computationally efficient crack detection across multiple publicly available benchmark datasets. Consistently high classification performance, together with convergence behaviour, Grad-CAM visualizations, ablation studies, computational efficiency analysis, and benchmark comparisons, indicates that the observed improvements arise primarily from the proposed architectural design rather than increased model complexity. In particular, the Edge-Gated Activation module improves boundary-sensitive feature extraction by emphasizing crack edges during feature learning, while the incorporation of Global Average Pooling reduces parameter redundancy and computational cost without compromising discriminative capability. These architectural choices enable HEAR-Net to achieve high predictive performance while maintaining a lightweight design that is suitable for practical deployment in infrastructure inspection scenarios.

The material strength prediction component complements the visual inspection module by estimating concrete compressive strength from physicochemical material properties. Compared with the Linear Regression baseline, the adopted regression model achieved improved predictive performance, indicating that nonlinear regression provides a more suitable representation of the complex relationships governing concrete strength. In addition to improved predictive accuracy, the regression tree offers an interpretable representation of the learned decision process, allowing engineers to examine the variables contributing to individual predictions. Within the proposed framework, these strength estimates provide complementary material information that enhances structural assessment but are not intended to replace standardized laboratory compressive strength testing.

The analytical condition assessment further illustrates how outputs from visual inspection and material strength estimation may be combined within a unified analytical workflow. Rather than serving as a validated asset valuation or maintenance decision model, this component demonstrates an illustrative approach for integrating heterogeneous structural indicators into a single analytical representation of infrastructure condition. Consequently, the analytical formulation should be interpreted as a proof of concept that highlights the feasibility of combining multiple assessment modalities, while future studies can investigate calibration using field inspection records, lifecycle information, and operational asset management data.

Overall, the proposed methodology demonstrates the feasibility of integrating computer vision, complementary material strength estimation, and analytical condition assessment within a unified infrastructure evaluation framework. The modular architecture provides flexibility for incorporating improved regression models, additional sensing modalities, and domain-specific analytical components without modifying the visual inspection pipeline. Nevertheless, the framework has not yet been validated using a single infrastructure dataset containing synchronized visual inspection images, material properties, and inspection or maintenance records. Future work will therefore focus on validating the proposed framework using integrated field datasets collected from operational infrastructure assets, while extending the methodology to multiclass defect classification, damage severity estimation, sensor fusion, temporal monitoring, and broader infrastructure categories.

4.4 Limitations

Despite the encouraging results, several limitations should be considered when interpreting the findings of this study. First, the visual defect detection component was trained and evaluated using publicly available crack image datasets. Although these datasets provide substantial diversity in terms of surface types and imaging conditions, they may not fully represent the variability encountered during real-world infrastructure inspections, including severe illumination changes, environmental degradation, occlusions, weather effects, and uncommon surface textures. Consequently, further validation using operational inspection datasets collected under diverse environmental and structural conditions is required.

Second, the material strength prediction component was evaluated using a standard concrete compressive strength dataset, with Linear Regression serving as the reference baseline. Although the adopted regression model demonstrated improved predictive performance, additional benchmarking against a broader range of established regression methods and validation using larger field-derived material datasets would provide a more comprehensive assessment of its robustness and practical applicability.

Third, the analytical condition assessment presented in this study is intended as an illustrative framework demonstrating how visual inspection results and estimated material properties may be integrated within a unified analytical workflow. The analytical formulation and associated coefficients have not been calibrated using longitudinal inspection records, maintenance histories, lifecycle cost information, or real-world asset management databases. Consequently, this component should be interpreted as a conceptual demonstration rather than a validated infrastructure valuation or maintenance decision model.

Fourth, although the proposed framework integrates visual defect detection, material strength prediction, and analytical condition assessment within a unified workflow, these components were developed and evaluated using independent publicly available datasets rather than a single infrastructure dataset containing synchronized visual inspection images, material properties, and inspection records. Consequently, the present study demonstrates the feasibility of conceptual integration rather than validation of a fully integrated end-to-end infrastructure assessment system. Future research should therefore validate the proposed framework using comprehensive field datasets acquired from the same infrastructure assets to establish its effectiveness under real-world inspection and asset management scenarios. Furthermore, the publicly available visual dataset does not provide parent-photograph or structure-level identifiers. Consequently, although independent training, validation, and testing subsets were maintained throughout experimentation, photograph-level separation could not be explicitly guaranteed, and similarity among neighboring image patches cannot be entirely excluded. Future studies should therefore evaluate HEAR-Net using photograph-level or infrastructure-level partitioning on comprehensive field datasets to further assess its generalization capability.

Fifth, the present study focuses on binary crack classification and does not explicitly address crack morphology, severity estimation, temporal crack progression, or multiple concurrent defect types. Future work will therefore focus on extending the proposed framework to multiclass structural defect classification, quantitative severity assessment, additional infrastructure asset categories, sensor fusion, temporal monitoring, and comprehensive validation using large-scale real-world inspection datasets.

5 Conclusion and future scope

This study presented an integrated framework for infrastructure condition assessment that combines deep learning-based visual defect detection, material strength prediction, and an illustrative analytical condition assessment within a unified workflow. Rather than proposing a fundamentally new Structural Health Monitoring (SHM) methodology, the proposed framework demonstrates how complementary sources of structural information derived from computer vision and material property estimation can be systematically organized and jointly interpreted within a unified infrastructure evaluation workflow while preserving modularity and interpretability.

The experimental results demonstrate the effectiveness of the proposed framework across its constituent components. The HEAR-Net module achieved highly accurate crack detection while maintaining strong generalization across multiple publicly available benchmark datasets. Comprehensive quantitative evaluation, convergence analysis, Grad-CAM visualizations, ablation studies, computational efficiency analysis, and benchmark comparisons collectively indicate that the proposed Edge-Gated Activation mechanism and Global Average Pooling contribute to improved crack localization, optimization stability, computational efficiency, and classification performance without relying on increased model complexity. These characteristics make HEAR-Net a lightweight and practical solution for automated structural defect detection.

The material strength prediction component complements the visual inspection module by providing estimates of concrete compressive strength from physicochemical material properties. Compared with the Linear Regression baseline, the adopted regression model achieved improved predictive performance, demonstrating that nonlinear regression provides a more suitable representation of the complex relationships governing concrete strength. The interpretable regression tree further improves transparency by allowing the learned decision process to be examined alongside the predicted strength estimates.

The analytical condition assessment illustrates how outputs from visual inspection and material strength estimation may be combined within a unified analytical workflow. Although this analytical formulation is intended solely as an illustrative demonstration and requires further validation using real-world inspection records and asset management data, it highlights the feasibility of integrating multiple engineering indicators into a single infrastructure assessment framework.

Overall, this work demonstrates the feasibility of integrating computer vision, complementary material strength estimation, and analytical condition assessment within a unified infrastructure evaluation framework. The proposed modular architecture provides flexibility for incorporating improved predictive models, additional sensing modalities, and domain-specific analytical components while maintaining transparency, computational efficiency, and practical applicability across diverse infrastructure inspection scenarios.

Future research will focus on validating the complete framework using integrated real-world infrastructure datasets containing synchronized visual inspection images, material properties, inspection records, and maintenance histories collected from the same physical assets. Such validation will enable assessment of the proposed framework as a fully integrated end-to-end infrastructure condition assessment system. Additional research will extend the visual inspection module to multiclass defect detection and quantitative severity estimation, evaluate the material strength prediction component using broader regression benchmarks and larger engineering datasets, and calibrate the analytical condition assessment using longitudinal inspection records, lifecycle information, and real-world asset management databases. Further investigations will also explore integration with conventional Structural Health Monitoring technologies, including vibration-based sensing, Internet of Things (IoT) platforms, and multimodal sensor fusion, to develop a more comprehensive intelligent infrastructure condition assessment framework.

Generative AI statement

The author(s) declared that generative AI was used in the creation of this manuscript. The authors used generative artificial intelligence during the preparation of this manuscript for language refinement, grammar correction and readability improvement. All the contents were verified, and approved by the authors.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Wei Lang, Sun Yat-sen University, China

Reviewed by: Mario Versaci, Mediterranea University of Reggio Calabria, Italy

Abdelwahhab Khatir, Oran University of Science and Technology - Mohamed Boudiaf, Algeria