Automatic Detection Method for Concrete Spalling and Exposed Steel Bars in Reinforced Concrete Structures Based on Machine Vision

Abstract

Reinforced concrete (RC), renowned for its amalgamation of strength and durability, stands as a cornerstone in modern engineering, extensively employed in various structures such as buildings, bridges, and pipe culverts. However, prevalent issues of concrete spalling and exposed steel bars within RC structures pose significant challenges. An automated identification methodology is proposed to detect concrete spalling and exposed steel bars, leveraging machine vision technology and deep learning algorithms. Initially, a classifier is utilized to discern concrete spalling areas within the image domain at the image level. Subsequently, a semantic segmentation algorithm is applied to precisely delineate the contours of both concrete spalling areas and exposed steel bars at the pixel level. The efficacy and feasibility of the proposed method are validated through training and testing on both a publicly available dataset and actual RC structure images. The results illustrate that the average detection precision, Intersection over Union (IOU), recall, and F1-score for concrete spalling areas are 0.924, 0.872, 0.937, and 0.925, respectively, while for exposed steel areas, the corresponding values are 0.905, 0.820, 0.899, and 0.855. This method demonstrates promising prospects for wide-ranging applications in defect detection within RC structures.

1. Introduction

Reinforced concrete (RC) structures are widely used in infrastructure, including buildings, bridges, and dams, due to its exceptional strength, durability, flexibility, and stability [1]. However, structural and material defects in RC structures, such as concrete spalling, exposed steel bars, and steel bar corrosion, are inevitably caused by natural disasters, environmental changes, and extended periods of use [2]. The safety of individuals and their property, as well as the overall integrity of the city’s infrastructure, may be threatened by these potential safety hazards. It is crucial that the early detection and troubleshooting of various forms of performance degradation in RC structures, along with the implementation of effective maintenance and improvement measures, are carried out to ensure social safety and stability [3,4]. Therefore, the timely and accurate detection of concrete spalling and exposed steel bars after concrete spalling in reinforced concrete structures has become a task of far-reaching significance.

In RC structures, a prevalent structural ailment is the delamination of the concrete protective layer, often triggered by factors such as material aging and environmental corrosion. In fact, the issue of concrete cover spalling can lead to prolonged exposure of the steel bars to the air. The strength of steel bars is significantly reduced by corrosion caused by oxygen, water, and other airborne agents, posing a significant threat to the integrity of the structure. Structural health detection is dedicated to ensuring that buildings and infrastructure can be used safely and reliably within their designed lifespan [5]. Focusing on the detection of RC structural issues, a series of research methods and technical approaches have gradually evolved, from contact detection to non-contact detection [6,7] and from human eye recognition to machine vision [8,9]. However, most of these methods focus on structural cracks or concrete spalling. There remains a gap in detailed and in-depth research concerning issues such as exposed steel bars and the corrosion of steel bars resulting from the spalling of the concrete protective layer.

With the rapid development of computer technology, methods that use machine vision to replace the human eye and intelligent algorithms to simulate expert experience have become increasingly popular for identifying apparent damage in engineering structures. These approaches are favored by both researchers and industry insiders [10,11]. So far, multifunctional structural damage detection and identification technologies, represented by target detection [12], image classification [13], and semantic segmentation [14,15], have been gradually developed using structural surface disease images as data sources and combined with deep learning algorithms. Object detection identifies and categorizes the types of diseases contained in images on a macro scale. Through the feature extraction and information integration of structural diseases, target objects are selected and marked in the form of rectangular frames and text labels. Tran et al. [16] proposed a novel deep learning method for detecting cracks on concrete bridge cracks. Image classification is used to filter and identify objects in the image domain at the image level so as to qualitatively determine whether the structure is damaged. Abubakr et al. [17] used deep learning algorithms to classify and detect the five most common defects in RC bridges (cracks, corrosion, weathering, spalling, and exposed steel bars). Relying on machine vision technology and deep learning algorithms, the above work respectively proposed research methods for detecting engineering structural diseases. However, the application potential of machine vision-based structural disease detection methods extends beyond simply determining whether damage or disease has occurred in a structure. It is fully capable of identifying the specific form of the disease.

A new research idea for structural disease detection based on visual images is provided by end-to-end image segmentation algorithms. These models, based on images or videos, employ pixels as the fundamental units to categorize groups of pixels representing diseases in the image into distinct semantic categories, thereby achieving the refined identification and detection of structural damage [18]. Arafin et al. [19] studied the category classification of common concrete damage (e.g., cracks and spalling, etc.) and performed the pixel segmentation of concrete cracks and spalling areas based on a semantic segmentation algorithm. Jiang et al. [20] proposed an improved UNet network framework to detect corrosion-related damage inside dull steel box girders. The study described above utilized deep neural networks to extract disease features from structural damage images, classify them, and integrate the information, addressing the identification and detection of structural surface damage forms. The essence of the image segmentation algorithm is to perform a binary classification task at the pixel level. Whether it is the number of parameters of the model or the complexity of the network structure, the image segmentation model far exceeds the image classification model. The computational efficiency of the model is greatly impacted by the complex network structure combined with repeated iterative calculations, thus presenting new challenges to the practical engineering application of the segmentation algorithm.

Through an on-site investigation and literature collection, this study takes the two common forms of disease, namely exposed steel bars and concrete spalling in RC structures, as the research objects. It aims to use machine vision technology and deep learning algorithms to achieve an automated, efficient and high-precision identification of concrete spalling, exposed steel bars, and concrete spalling in RC structures. A cascaded network structure is proposed to detect exposed steel bars and concrete spalling areas on the surface of RC structures, combining the characteristics of the fast prediction speed of the image classification model and the high degree of refinement of the image segmentation model. The overall performance of the detection model in actual structural disease identification can be improved by constructing a suitable dataset and designing an effective deep neural network structure. In addition, this study analyzes the reliability of the model identification results and explores the application challenges of detection models in actual structural health inspection and maintenance activities. The contributions of this paper can be listed as follows:

(1)

A novel cascaded framework is proposed for detecting the co-occurring superficial defects of concrete spalling and exposed steel bars in RC structures. Preceding the segmentation process, the defect regions are preliminarily identified by a classification model, furnishing explicit prior information to guide subsequent refined semantic segmentation models and thus effectively enhancing the detection efficiency and accuracy of structural superficial defects.

(2)

The most suitable model combinations are selected for practical engineering applications through a lateral comparison of various feature encoders. Balancing the detection efficiency and predictive accuracy of the models comprehensively, it is determined that ResNet-50 and VGG-19 are respectively suitable for precise segmentation tasks of concrete spalling and exposed steel bars.

The framework of this paper is presented as follows. Section 1 introduces the research background. Section 2 elaborates the detection method proposed in this article. Section 3 reports the experimental verification. Section 4 analyzes the calculation results, and Section 5 gives the conclusions.

2. Methodology

2.1. Basic Framework

Concrete surface spalling represents a common form of disease in RC structures. This not only impacts the aesthetics and comfort of the engineering structure but also diminishes its load-bearing capacity and service life. Compounding the issue, the exposed steel bars resulting from the peeling concrete protective layer are susceptible to corrosion, leading to potential damage to the internal structure. Upon examination of a large number of structural disease images from inspection sites, it becomes evident that the two disease forms, namely the spalling of the concrete protective layer and the exposure of steel bars, frequently coexist in the same image. An automatic disease segmentation method that combines machine vision technology with deep learning algorithms is proposed to detect and separate the information of the two diseases from the same image. The basic process is shown in Figure 1.

The input of the detection algorithm is a surface image of the RC structure to be detected. In order to clearly record the surface damage status of a structure, inspectors usually take photos with devices such as smartphones, portable cameras, or drones. Considering the limited computing power of the hardware device, the high-resolution structural surface image is cropped by a fixed-size sliding window into a partial image with a side length of 224 pixels. The input image should be padded with background pixels into an image with a side length divisible by 224 to ensure the accuracy and completeness of image cropping. This detection method combines an image classifier and two semantic segmentation models to perform different degrees of feature extraction and feature expression on the disease pixel groups in the structural surface image at the image level and the pixel level, respectively. Specifically, the image classifier built by the VGG network determines the probability that each image patch in the input image domain contains peeling and outputs the corresponding probability matrix $P_{ij}$. The probability matrix for judging the peeling area in the image domain is processed into a binary matrix through taking 0.5 as the probability threshold. Image patches that do not contain peeling are filled with solid color pixels for saving computational resources. Image patches containing exfoliation are used as input to the segmentation model for pixel-level classification. Both the concrete spalling segmentation model and the exposed steel bar segmentation model are designed to learn and capture the pixel features corresponding to spalling and steel bars in the input image, respectively. The semantic information expressing the two objects is expressed in the form of mask images. Ultimately, image fusion technology is employed to amalgamate the two types of disease information, namely concrete spalling and exposed steel bars, detected from the structural surface images.

2.2. VGG Classifier

VGG19, a classic classification network, derives its name from its architecture, which includes 19 convolutional layers and 3 fully connected layers. It has demonstrated outstanding performance on large-scale image datasets such as ImageNet [21]. The transfer learning method of porting the pre-trained VGG-19 model is adopted to train a classifier with high generalization ability and accuracy when working with a limited number of RC concrete pictures. As shown in Figure 2, the structure of the classifier consists of an input layer, convolution layers, pooling layers, fully connected layers, and activation functions. The input layer is an image patch with dimensions of 1 × 224 × 224 × 3. The convolution block, which is a folded combination of convolutional layers and pooling layers, plays the role of extracting image features. In blocks, the input patches generate the spatial resolution of the feature maps through convolution, ReLU, and the maximum pooling layer. At the back end of the last block, one fully connected layer with dimensions of 1 × 1 × 1024 is connected for feature fusion and mapping. The output layer is a fully connected layer containing only two neuron nodes. It provides feedback on the category of the input image, indicating whether it belongs to the background or the category of a diseased image. The classifier uses a cross-entropy loss function for error fitting.

$$Loss=-\frac{1}{N}{\sum }_{i=1}^N{\sum }_{k=1}^K{y}_{i,k}·\log \left({\widehat{y}}_{i,k}\right),$$

(1)

where $N$ is the number of samples in the training set, $y_{i,k}$ is the true label of the $i$-th sample belonging to the k-th category, and ${\widehat{y}}_{i,k}$ is the predicted probability of the $k$-th category by the model.

2.3. UNet-Based Segmentation Model

UNet is an end-to-end deep convolutional neural network model, originally proposed by Ronneberger et al. [22] in 2015, aiming to solve the task of the refined segmentation of medical images. The network structure garnered significant attention in the industry immediately upon its proposal and has rapidly found applications in various fields related to machine vision, including medicine, machinery, agriculture, transportation, and more. This paper introduces a disease detection method that employs the UNet network structure for segmenting concrete spalling and exposed steel bars in RC structures. The basic structure of UNet disease segmentation network is shown in Figure 3.

The overall architecture can be divided into two main parts: encoder and decoder. The encoder plays a crucial role in extracting high-level feature representations from the input image. It achieves this by utilizing a series of convolutional layers and pooling layers to systematically decrease the resolution of the input image while simultaneously increasing the number of channels in the feature map. This step-by-step reduction in resolution aids in capturing local features and contextual information, ultimately generating compact and informative feature representations. The decoder is responsible for mapping the feature maps generated by the encoder back to the same dimensions as the original input image and recovering the semantic information. It typically includes up-sampling operations, often achieved through deconvolutional layers, and incorporates skip connections corresponding to the encoder. These skip connections allow the decoder to recover lower-level feature information, enhancing segmentation performance and mitigating information loss. The output of the UNet model is a segmentation mask image that delineates the locations of target objects or regions in the input image.

In the UNet segmentation model, the loss function is usually defined as the sum of the cross-entropy loss for each pixel. For a binary classification problem (foreground and background), the pixel-level cross-entropy loss function can be expressed as follows:

$$Loss=-\frac{1}{N}{\sum }_{i=1}^N{\sum }_{j=1}^M\left(y_{i,j}·\log \left({\widehat{y}}_{i,j}\right)+\left(1-y_{i,j}\right)·\log \left(1-{\widehat{y}}_{i,j}\right)\right),$$

(2)

where $N$ is the number of samples, $M$ is the total number of pixels in each sample, ${\widehat{y}}_{i,j}$ is the true label of the $j$-th pixel in the $i$-th sample, and ${\widehat{y}}_{i,j}$ is the model’s response to the $i$-th the predicted probability of the $j$-th pixel in the sample.

Concrete spalling and exposed steel bars exhibit diverse pixel intensities and combinations within the image. Different feature extraction networks with distinct structural forms and depths are selected to address varying requirements. Widely used feature extraction networks for image segmentation tasks encompass VGG, ResNet [23], DenseNet [24], and EfficientNet [25]. To enhance the precision and efficiency of objects segmentation, this article incorporates the aforementioned feature extractors into the UNet network framework.

2.4. Evaluation Index

When assessing the performance of the segmentation model for concrete spalling and exposed steel bars, commonly employed metrics include pixel precision, Intersection over Union (IOU), recall, and F1-score. Evaluation indices for the segmentation model are defined using a confusion matrix. True positive (TP) signifies the number of positive samples correctly predicted by the model. True negative (TN) is the count of negative samples correctly predicted by the model. False positive (FP) denotes the number of positive samples incorrectly predicted by the model. False negative (FN) represents the count of negative samples incorrectly predicted by the model.

Pixel precision is the ratio of the number of correctly classified pixels to the total number of pixels. Its calculation formula is given as follows:

$$Precision=\frac{TP}{TP+FP}$$

(3)

The Intersection over Union (IOU) ratio quantifies the relationship between the intersection and union of the predicted region and the true region. Its calculation formula is given as follows:

$$IoU=\frac{TP}{TP+FP+FN},$$

(4)

Recall is the ratio of true positive samples correctly predicted as positive samples. Its calculation formula is given as follows:

$$Recall=\frac{TP}{TP+FN},$$

(5)

F1-score is the harmonic average of precision and recall. Its calculation formula is given as follows:

$$F1=2\times \frac{Precision\times Recall}{Precision+Recall}$$

(6)

3. Experimental Verification

In this section, calculation models were constructed using the Python 3.7 language and the PyTorch 2.0 deep learning platform. The efficacy and reliability of the detection method are validated by combining images depicting concrete spalling and exposed steel bars in real engineering structures. To ensure accuracy and consistency, the training, testing, and prediction of the detection model are all performed within the same hardware environment. The computing device used in this study is an Intel(R) Xeon(R) W-2123 processor with RAM 64.0 GB, CPU @3.60 GHz, and GPU NVIDIA GeForce RTX 2080 Ti.

3.1. Datasets

This study collects disease images of concrete spalling and exposed steel bars inside RC structures from three methods, namely public datasets [26,27,28,29], Internet retrieval, and on-site photography. Whether public datasets or photos are collected by researchers, these RGB images require the manual production of label images for training classification and segmentation models. The mask label tool was used to create a disease mask map (sparing and exposed steel bars) of engineering structural images. Normally, the image resolution of engineering structures vary, depending on the shooting equipment. In addition, due to the computational bottleneck of hardware devices, the images input to the image segmentation model need to be divided into small-sized image patches in common semantic segmentation work. For the classification task, the small-sized image patches are artificially divided into background images and disease images based on the presence of disease information. The specific production process of the dataset is shown in Figure 4.

After image cropping and image renaming, three image sets were obtained, each with a size of 224 × 224 pixels, including 3000 sub-images for disease classification, 1500 sub-images for concrete spalling area segmentation, and 1500 sub-images for the segmentation of exposed steel bars. These images are divided into training sets and testing sets in a ratio of 9:1. The specific dataset information is shown in

Table 1. Datasets in different models.

Model	Train Image	Train Label	Test Image	Test Label
Classifier	2700	2700	300	300
Concrete spalling segmentation	1350	1350	150	150
Steel bars fell off segmentation	1350	1350	150	150

.

3.2. Training of the Classifier

Understanding structural surface disease information at the image level is a task that can be easily accomplished. A classification model based on the VGG-19 network is built to distinguish background and diseases (concrete spalling and exposed steel bars) in RC surface images. The training history curve of the classifier is shown in Figure 5. During training, the optimal model can be saved by monitoring the loss value on the validation set. Observing the training process of the model, it is not difficult to find that the image classifier begins to converge from the 20-th cycle, and the convergence value of the training set is slightly lower than the convergence value of the validation set, which are 0.091 and 0.100, respectively. In addition, the classification accuracy is used as an evaluation indicator of the quality of the training model. Likewise, the accuracy of image classification stabilizes after the 20-th training cycle. The classification accuracy of the training set remains at 0.975, which is slightly higher than the 0.950 of the validation set. It can be seen that the classifier can maintain high classification accuracy for the binary classification task at the image level. This provides a guarantee and support for the subsequent accurate segmentation of concrete spalling areas and exposed areas of steel bars.

3.3. Training and Testing of the Segmentation Model

In order to reduce the risk of misjudgment and reduce computational costs, the initially classified background image blocks do not participate in subsequent disease information segmentation. The Efficientnet-B7 network is used as the feature encoder to extract features, and we combine it with the UNet framework to build a basic semantic segmentation model. As the main detection objects, concrete spalling and exposed steel bars are trained into independent semantic segmentation models. During the model training and testing process, the precision, IOU, recall, F1-score, and the convergence value of the loss function are used to evaluate the performance of the segmentation model.

The training curve of the concrete spalling segmentation model is shown in Figure 6. Judging from the trend of the loss function curve, the model gradually begins to converge from the 90-th training cycle and finally stabilizes within a specific value range. The convergence value of the loss function on the training set finally approaches 0.052, which is lower than 0.083 on the validation set. The intersection ratio reflects the segmentation performance of the model to a certain extent. When the model converges, the intersection ratios of the training set and validation set of the concrete spalling segmentation model reach 0.912 and 0.865, respectively. The prediction results of some disease images that have not participated in model training are shown in Figure 7. It can be seen from the prediction result graph and each evaluation index value that the segmentation effect of concrete spalling is very good in RC structures. This also shows that the pixel characteristics of the concrete spalling area on the image are more obvious.

The training curve of the exposed steel bar segmentation model is shown in Figure 8. The damage functions of the training set and validation set start to converge from the 60-th training epoch. Finally, the convergence values of the loss function are 0.092 and 0.131, respectively. As one of the evaluation indicators for the performance of the segmentation model, the intersection-to-union ratio curve shows a trend of increasing first and then stabilizing, and the convergences of the intersection-to-union ratio of the training set and the verification set are stable at 0.835 and 0.816. The prediction results and evaluation index values of some sub-images are shown in Figure 9. Exposed steel bars will undergo chemical changes such as corrosion and rust due to long-term contact with air, thus changing the original external characteristics of the steel bars. Compared to concrete spalling, the precise segmentation of exposed areas of steel bars in reinforced concrete structures is more challenging.

4. Results and Analysis

The performance of submodules in the reinforced concrete disease detection method proposed in this paper was discussed in the previous section. This section verifies the overall detection performance of this method based on disease samples existing in actual projects. Before this, an ablation experiment with contrast function should first be studied.

4.1. Ablation Experiment

The detection method proposed in this article is implemented based on the UNet network architecture. In fact, the feature extraction capability of the encoder is a key factor affecting the accuracy of reinforced concrete disease segmentation. For the two segmentation models of concrete spalling and exposed steel bars, four feature encoders, namely VGG-19, ResNet-50, DenseNet-121, and EfficientNet-B7, were used to build the segmentation models to explore the disease information (concrete spalling and exposed steel bars) with the participation of different encoder segmentation effects. In addition to the four evaluation indicators mentioned above, the size and calculation time of each model were also considered to provide a timeliness analysis for the actual application of the detection model.

Moreover, 250 image patches with a size of 224 × 224 that were not involved in training were used to test the segmentation performance of the concrete spalling area. The segmentation results are shown in

Table 2. The concrete spalling segmentation model for different encoders based on the UNet framework.

Encoder	Precision	IoU	Recall	F1-Score	Val-Loss	Parameters	Time Costs (Seconds/Image Patches)
VGG-19	0.909	0.855	0.932	0.911	0.103	29,057,937	0.040
ResNet-50	0.915	0.858	0.931	0.915	0.109	32,521,105	0.050
DenseNet-121	0.911	0.857	0.931	0.938	0.104	13,607,633	0.057
EfficientNet-B7	0.924	0.872	0.937	0.925	0.083	67,095,329	0.099

. Each parameter in the table is the average test result of multiple images. Without considering the computational efficiency of the model, Efficientnet-B7, which has the most weight parameters, shows the best segmentation effect. If calculation efficiency is taken into account, the detection model based on the Efficientnet-B7 encoder takes an average of 0.099 s to predict an RGB image with a size of 224 × 224, which is approximately twice that of the other three models. Comparing the three encoders horizontally, VGG-19, ResNet-50, and DenseNet-121, ResNet-50 can maintain a high level of overall segmentation effect in concrete spalling areas while maintaining high computational efficiency.

The test results of the exposed steel area segmentation model under different encoders are shown in

Table 3. The exposed steel bar segmentation model for different encoders based on the UNet framework.

Encoder	Precision	IoU	Recall	F1-Score	Val-Loss	Parameters	Time Costs (Seconds/Image Patches)
VGG-19	0.898	0.812	0.896	0.847	0.152	29,057,937	0.040
ResNet-50	0.903	0.802	0.876	0.835	0.129	32,521,105	0.050
DenseNet-121	0.893	0.793	0.876	0.828	0.133	13,607,633	0.058
EfficientNet-B7	0.905	0.820	0.899	0.855	0.112	67,095,329	0.097

. The data in the table are the average test results of 250 samples. Overall, the overall segmentation effect of the exposed steel bar area is slightly lower than the segmentation effect of the concrete spalling area. The reasons for this phenomenon were analyzed in detail in Section 3.3. The model based on the EfficientNet-B7 encoder also shows the best segmentation effect, but its computational efficiency is low. While maintaining optimal computational efficiency, the model based on the VGG-19 encoder can also achieve excellent segmentation accuracy.

4.2. The Detection Cases in Engineering

The accurate segmentation of concrete spalling and even exposed areas of steel bars is of great significance in actual safety inspections of reinforced concrete structures. This section takes actual engineering structural disease photos as the research object and uses the RC structural disease detection method proposed in this article to predict the specific forms of concrete spalling and exposed steel bars that may exist in the images. Among them, the UNet models based on ResNet-50 and VGG-19 are used to segment concrete spalling and the exposed steel bars. Some test results are shown in Figure 10. Concrete spalling and exposed steel bars exist in the same image domain, and there is an obvious position overlapping relationship. The segmentation model comprehensively learns and understands the distinctions in pixel group distribution for spalling concrete, steel bars, and background areas. Therefore, judging from the prediction result map, the predicted mask image of the concrete spalling area can basically reflect the diseased area of concrete in the actual structure. Compared with the overall image of the concrete spalling area, the exposed steel bars are more complex in form and texture. Fortunately, the specific shape of exposed steel can also be predicted very well. An image fusion technique is used to merge the mask images of these two disease forms. In addition, the original structure image and the disease mask image are once again fused to examine the segmentation results in an intuitive way.

During actual engineering structure inspections, drones, mobile phones, or portable cameras are used to capture photos of the structure’s surface on-site. These photos are then input into a pre-trained RC structural disease detection model for disease prediction. Due to the cascaded structure of the detection method, the image classifier initially identifies the approximate region of the disease within the image domain, followed by the image segmentation model performing detailed pixel-level segmentation. The on-site detection of a high-resolution, complete photo takes approximately 0.5 to 2 s. From the perspective of computational efficiency, this method meets the real-time performance demands of the actual detection process.

5. Conclusions

Based on the UNet framework, this paper proposes a segmentation method suitable for two common structural issues—concrete spalling and exposed steel bars—in RC structures. The accuracy and practicability of the detection method are verified by training and testing public datasets and actual structural surface disease images. It provides novel methods and strategies for the safety detection of reinforced concrete structures. The conclusions of this article are drawn as follows:

• A cascade method is introduced for the detection of RC structural diseases. Initially, the image domain is initially divided into background areas and disease areas at the image level. Subsequently, the pixel-level semantic segmentation method is used to accurately identify concrete spalling and exposed steel bars within the image. This cascade detection method not only enhances the efficiency of detecting RC surface images but also eliminates potential interference from objects in the image background.
• The concrete spalling segmentation model and the exposed steel bar segmentation model proposed in this article both show excellent segmentation results in the test images. The average precision, IOU, recall, and F1-score of the two models predicting the spalling concrete and exposed steel bars reached 0.924, 0.872, 0.937, and 0.925 and 0.905, 0.820, 0.899, and 0.855, respectively.
• Four encoders, VGG-19, ResNet-50, DenseNet-121, and EfficientNet-B7, are used to build segmentation models for verifying the computational efficiency of the detection method. The results show that the EfficientNet-B7 encoder can achieve the highest detection accuracy, but the computational efficiency is the lowest. Taking calculation efficiency into consideration, ResNet-50 and VGG-19 are suitable for precise segmentation tasks of concrete spalling and exposed steel bars, respectively.

In future work, we aim to develop a method for the quantitative detection of concrete spalling and exposed areas of steel bars. Additionally, we plan to propose a lightweight detection model suitable for installation on mobile devices (e.g., smartphones) to enhance practical engineering applications. To solve the problem of difficulty in producing image segmentation datasets, the large-scale segmentation algorithm SAM can be used to assist in the production of image labels. The method is also a good way to save prerequisite costs.