A Rapid Bridge Crack Detection Method Based on Deep Learning

Abstract

The aim of this study is to enhance the efficiency and lower the expense of detecting cracks in large-scale concrete structures. A rapid crack detection method based on deep learning is proposed. A large number of artificial samples from existing concrete crack images were generated by a deep convolutional generative adversarial network (DCGAN), and the artificial samples were balanced and feature-rich. Then, the dataset was established by mixing the artificial samples with the original samples. You Only Look Once v5 (YOLOv5) was trained on this dataset to implement rapid detection of concrete bridge cracks, and the detection accuracy was compared with the results using only the original samples. The experiments show that DCGAN can mine the potential distribution of image data and extract crack features through the deep transposed convolution layer and down sampling operation. Moreover, the light-weight YOLOv5 increases channel capacity and reduces the dimensions of the input image without losing pixel information. This method maintains the generalization performance of the neural network and provides an alternative solution with a low cost of data acquisition while accomplishing the rapid detection of bridge cracks with high precision.

1. Introduction

Concrete is widely used in dams, bridges, and other large-scale engineering structures [1]. For these structures, maintenance, monitoring, and life assessments are very important tasks after the post-construction period [2,3,4]. Among the various disasters that can occur during the maintenance period of bridge engineering, cracks often appear first [5]. This is due to the uneven settlement of the bridge foundation in the vertical direction and displacement in the horizontal direction, leading to internal stresses in the concrete structure and resulting in cracks [6]. For foundations that are built in phases or subjected to the effects of frost in cold areas, deformation of the structure and cracks can also occur [7,8]. Additionally, bridge cracks can cause significant harm: (1) Cracks can result in leaks, causing water flow to into the interior of the bridge and destroying the concrete’s mechanical characteristics and physical properties. When flowing water freezes, it will cause the formation of deeper and wider cracks, which will cause instability of the main structure of the bridge, decreasing the safety level. The water in the bridge cracks will cause further development and expansion of cracks under the influence of gravity and pressure during construction [9]. (2) The cracks will lead to carbonation in the concrete structure of the bridge. The concrete material reacts with CO₂ from the environment in the presence of moisture to produce calcium carbonate. As a result, the safety and mechanical properties of the concrete structure are reduced [10]. (3) Bridge cracks will destroy the purification film of steel bars and metal components and corrode under the simultaneous penetration and action of external air and water. The rust generated after the corrosion of steel bars is more than ten times larger than the initial volume, which will reduce the stability of the reinforced concrete engineering [11]. In summary, it is very important to obtain the location, length, width, and extension condition of bridge cracks in time through detection technology.

However, with continuous increases in the number of completed bridges and bridge spans, crack detection work becomes increasingly demanding. At the same time, the economic costs also become higher [12]. With the accelerated growth of computing capabilities in recent years, advancements in deep learning technology have motivated us to develop a novel approach to address these problems [13,14,15]. Li et al. [16] presented a novel approach for bridge crack detection by enhancing the encoder–decoder framework and utilizing a mixed pooling module. The encoder employs dilated convolutions to extract the fundamental characteristics of crack images. This methodology ensures preservation of the feature image resolution and facilitates the acquisition of a wide receptive field. Notably, the experimental results demonstrated that this technique achieved significantly higher precision and recall metrics. Li et al. [17] proposed the utilization of deep learning techniques for bridge crack detection using unmanned aerial vehicles (UAVs). They adopted the faster region convolutional neural network (faster R-CNN) algorithm based on VGG16 transfer learning to detect bridge cracks effectively. The experimental results indicated that the automatic detection of bridge cracks using UAVs and the faster R-CNN algorithm could provide enhanced efficiency without compromising accuracy. By leveraging the advantages of depthwise separable convolution, atrous convolution, and the atrous spatial pyramid pooling module, Xu et al. [18] presented a convolutional neural network (CNN) end-to-end crack detection method. This study showed that the model had better capabilities compared with conventional classification models. Moreover, the model had the potential to be integrated into any convolutional network, serving as an efficient module for feature extraction. Based on the current research status, it is obvious that obtaining large amounts of high-quality data on cracks is still a time-consuming and costly task [19]. Additionally, it is worth discussing which form of neural network is most suitable for the rapid detection of bridge cracks.

Therefore, the DCGAN is employed to generate a large number of artificial crack samples to expand the dataset. As a representative artificial neural network (ANN), DCGAN can mining the potential distribution of image data and achieve image data fitting so that the neural network can produce high-quality realistic images based on learned image features [20,21]. In terms of application fields, generative adversarial networks, a popular means of dataset augmentation and generation, have been widely used in medical fields, safety inspection, agricultural production, and other fields [22]. An imbalanced fault diagnosis method based on the generative model of DCGAN was proposed by Luo et al. [23] to solve the problems of limited datasets. In order to expand fake fingerprint data, Choi et al. [24] proposed a method to investigate whether a fake fingerprint generated by DCGAN was similar to a fake fingerprint from the dataset. At present, image generation technology is still relatively rare in the field of roads, and the intelligent detection of road cracks is also in its exploratory stages. For example, the collected images of roads often have complex backgrounds and are heavily affected by light, and conditions such as cracks are often difficult to distinguish [25]. Therefore, it is of value to study bridge crack data augmentation methods based on DCGAN.

YOLOv5, which is considered to be an efficient and stable target detection neural network, is used to achieve the rapid detection of bridge cracks [26,27]. In recent years, the YOLOv5 object detection architecture has gained increasing attention in the field of computer vision due to its outstanding performance in detecting various objects in real-time scenarios. The utilization of YOLOv5 for bridge crack detection brings several advantages [28,29,30,31]. Firstly, YOLOv5 offers a high accuracy in detecting and localizing cracks on bridge surfaces. With its advanced anchor-based and anchor-free mechanisms, YOLOv5 can effectively identify and delineate fine cracks, regardless of their length, width, or orientation. This enables engineers and researchers to obtain precise information regarding crack locations and sizes. Secondly, the real-time capabilities of YOLOv5 enable rapid crack detection, aiding in efficient inspections. Its speed-optimized architecture allows inspectors to quickly assess the conditions of bridges. This reduces the time and resources required for manual inspections, making crack detection more time-efficient and cost-effective. Moreover, YOLOv5 can adapt to varying crack features and textures, ensuring consistent and reliable crack detection across different bridge types. This flexibility eliminates the need for multiple detection techniques, simplifying the crack detection process. This study shows that the combination of DCGAN and YOLOv5 can carry out the detection of engineering defects rapidly and accurately.

The sections of this study are structured as follows: Section 2 presents a detailed introduction of the establishment of the dataset and the structure and theory of DCGAN and YOLOv5. In Section 3, the training process, training environment, and predicted results are presented. Subsequently, Section 4 discusses the research motivation, limitations of this work, and future research directions. Finally, Section 5 provides the concluding remarks for this study.

2. Materials and Methods

2.1. Establishment of the Dataset

A total of 2068 sets of bridge crack datapoints with various morphologies were collected, as shown in Figure 1. The dataset of cracks is feature rich. Figure 1a–d shows vertical cracks, cross cracks, horizontal cracks, and wider cracks, respectively. These 2068 images were used as the training data for DCGAN.

2.2. The Theory of DCGAN

The GAN was first proposed by Goodfellow in 2014 [32], aiming to produce generated samples that are nearly consistent with the distribution of real data, which belongs to the category of unsupervised learning [33]. GANs have attracted a large number of researchers and have subsequently achieved many research results in computer vision fields such as image synthesis, style transfer, image repair, and object detection [34]. DCGAN combines CNN with GAN, which greatly improves the stability of GAN and the effect of the output results. DCGAN usually consists of a generator and discriminator.

(1)

Generator

The primary objective of the generator is to understand and absorb the attributes present in the training data. It accomplishes this by aligning the random noise distribution with the actual distribution of the training data under the guidance of the discriminator. This process enables the generator to produce synthetic data that closely resemble the characteristics observed in the training dataset. The training objective function can be expressed as follows:

$$G^{\ast }=\arg \underset{G}{\min } Div(P_{data(x)},P_G)$$

(1)

In the equation, $P_{data}$ and $P_G$ are the real and generated data distributions, respectively, and $Div()$ is the difference between the distributions.

(2)

Discriminator

The primary role of the discriminator is to differentiate between real and generated data produced by the generator while providing feedback to the generator. Both networks undergo iterative training, where their capabilities grow simultaneously until the generated network is capable of producing data that can deceive the discriminator. This process focuses on refining the discerning abilities of the discriminator and enhancing the generative potential of the network. Finally, a certain balance will be reached in terms of the capabilities of the generator and discriminator. The training objective function can be expressed as follows:

$$D^{\ast }=\arg \underset{D}{\max } Div(P_G,P_{data})$$

(2)

The true likelihood of the input sample is manifested in the discriminator objective function value, reflecting a binary classification challenge. The formula for the training objective function can be articulated as follows:

$$V(G,D)=E_{X~P_{\mathrm{data}(x)}}[\log D(x)]+E_{Z~P_{Z(z)}}[\log (1-D(G(z))]$$

(3)

In the equation, $E_{X~P_{data(x)}}$ and $E_{Z~P_{Z(z)}}$ are the real data x and the noise z expectations, respectively; $D(x)$ is the output of the real data x; and $G(\mathrm{z})$ represents the noise z through the generator.

Suppose that in a continuous space, Equation (3) can be expressed as:

$$V(G,D)={\displaystyle {\int }_x[P_{data(x)}\log (D(x)+P_G(x)\log (1-D(x)))]dx}$$

(4)

For the integrand function $F(x)$,

$$F(x)=P_{data(x)}\log D(x)+P_G\log (1-D(x))$$

(5)

While $P_{data}$ and $P_{G(x)}$ are arbitrary non-zero real numbers, Equation (5) obtains the maximum value when Equation (6) is satisfied. Equation (6) can be expressed in the following form:

$$D^{\ast }=\frac{P_{data}(x)}{P_{data}(x)+P_G(x)}$$

(6)

Equation (7) can be obtained by taking Equation (6) into Equation (4):

$$V(G,D)=-2\log 2+2JSD(P_{data}||P_G)$$

(7)

In the equation, $JSD(P_{\mathrm{data}}||P_G)$ is the $JS$ divergence between $P_{data}$ and $P_G$. The more similar $P_{data}$ and $P_G$ are, the smaller the $JS$ divergence; conversely, the more different $P_{data}$ and $P_G$ are, the larger the $JS$ divergence.

Considering the generator and the discriminator together, the objective function of DCGAN can be expressed as:

$$\underset{G}{\min } \underset{D}{\max }V(G,D)=E_{X~P_{data(x)}}[\log D(x)]+E_{Z~P_{Z(z)}}[\log (1-D(G(z)))]$$

(8)

2.3. The Architecture of the DCGAN

Figure 2 shows the generator architecture used in this study. To commence, the generator is supplied with a random noise vector of 100-dimensions and undergoes an upward sampling procedure facilitated by dee- transposed convolution layers. Then the feature maps are obtained, which have different scales. Transposed convolution is a special kind of forward convolution which first expands the dimensions of the input image [35]. This process is also called up-sampling, which can convert images into higher resolutions. Every up-sampling operation is succeeded by a layer of batch normalization and a layer implementing an activation function. The last layer uses the Tanh function, and the other activation function layers use the Relu function. The initial random noise vector undergoes a series of seven up-sampling operations, resulting in the generation of an image with dimensions of 256 × 256 × 1. The complete network structure and parameters of each layer of the generator is shown in

Table 1. The complete network structure of the generator.

Layer	Output Shape	Parameter
Reshape_1	(1 × 1 × 100)	0
Conv2D_transpose_1	(4 × 4 × 1024)	1,639,424
BN_1	(4 × 4 × 1024)	4096
Activation_1	(4 × 4 × 1024)	0
Conv2D_transpose_2	(8 × 8 × 512)	8,389,120
BN_2	(8 × 8 × 512)	2048
Activation_2	(8 × 8 × 512)	0
Conv2D_transpose_3	(16 × 16 × 256)	2,097,408
BN_3	(16 × 16 × 256)	1024
Activation_3	(16 × 16 × 256)	0
Conv2D_transpose_4	(32 × 32 × 128)	524,416
BN_4	(32 × 32 × 128)	512
Activation_4	(32 × 32 × 128)	0
Conv2D_transpose_5	(64 × 64 × 64)	131,136
BN_5	(64 × 64 × 64)	256
Activation_5	(64 × 64 × 64)	0
Conv2D_transpose_6	(128 × 128 × 32)	32,800
BN_6	(128 × 128 × 32)	128
Activation_6	(128 × 128 × 32)	0
Conv2D_transpose_7	(256 × 256 × 1)	513
Activation_7	(256 × 256 × 1)	0

.

Figure 2 also shows the discriminator architecture used in this study. The discriminator undertakes seven down-sampling operations when presented with both the real bridge crack image and the generated crack image. These operations efficiently diminish both the size of the images and the dimensions of their respective features [36]. Each convolution layer employs convolution kernels measuring 4 × 4. The count of convolution kernels employed in each respective layer is 64, 128, 256, 512, 1024, and 1. For each down-sampling operation, a batch normalization layer and a Leaky Relu function layer follow. The Leaky Relu function layer incorporates a negative slope coefficient of 0.2. Consequently, a comprehensive down-sampling process enables the extraction of the features of the images. The ultimate loss value is derived based on the features extracted from the down sampling process. When the image of the unauthentic bridge crack or the real bridge crack is received, the loss value closes to 1 or 0, respectively. The complete network structure and the parameters of each layer of the generator are shown in

Table 2. The complete network structure of discriminator.

Layer	Output Shape	Parameter
Conv2D_1	(128 × 128 × 64)	1088
BN_1	(128 × 128 × 64)	256
Leaky Relu_1	(128 × 128 × 64)	0
Conv2D_2	(64 × 64 × 128)	131,200
BN_2	(64 × 64 × 128)	512
Leaky Relu_2	(64 × 64 × 128)	0
Conv2D_3	(32 × 32 × 256)	524,544
BN_3	(32 × 32 × 256)	1024
Leaky Relu_3	(32 × 32 × 256)	0
Conv2D_4	(16 × 16 × 512)	2,097,664
BN_4	(16 × 16 × 512)	2048
Leaky Relu_4	(16 × 16 × 512)	0
Conv2D_5	(8 × 8 × 1024)	8,389,632
BN_5	(8 × 8 × 1024)	4096
Leaky Relu_5	(8 × 8 × 1024)	0
Conv2D_6	(4 × 4 × 1)	16,385
Flatten_1	16	0
Dense_1	1	17

.

2.4. The Architecture of YOLOv5

YOLOv5 is an efficient target detection algorithm [37]. Similar to previous generations of Yolo algorithms, YOLOv5 adopts the concept of grids, that is, the image is divided into multiple meshes, each of which is responsible for predicting one or more objects. Each grid can produce prediction boxes (i.e., anchor), and templates for three prediction boxes are generally pre-stored. The anchor has a preset width, height, coordinates, and confidence level. The confidence level indicates the probability that an object is present in the mesh. When the mesh in which anchor is located has objects, the confidence level is 1, and vice versa is 0. If we regard the difference between the width and height of anchor and the difference of coordinates as losses, and the binary cross entropy as a loss of confidence, then the problem of target detection will be greatly simplified into a simple regression prediction and classification problem.

Figure 3 shows the YOLOv5 architecture used in this study. YOLOv5 mainly consists of Backbone network, Neck network and Head network. The Backbone part is mainly used for feature extraction and continuous reduction of feature map. The fusion of shallow graphic features and deep semantic features is performed by Neck network and the Head network is the classifier and regressor of YOLOv5. The main modules of YOLOv5 include Focus, CBL, CSP1_X, CSP2_X and SPP. Among them, the Focus module operates as part of the initial processing before the Backbone network. This operation essentially retrieves a value from alternate pixels in an image, creating an effect akin to proximate down sampling. The original 640 × 640 × 3 image is inputted into the Focus module, and the feature map of 320 × 320 × 12 is first changed by slicing operation. It finally becomes a feature map with the size of 320 × 320 × 32 after a convolution operation; The convolutional layer, batch normalization layer and Leaky Relu function are the main parts of The CBL module; CSP1_X module uses the CSPNet structure, consisting of three convolutional layers and X Res units modules while CSP2_X module replaces Res unit with CBL; Multi-scale fusion is carried out by SPP module with Maxpool operation of 5 × 5, 9 × 9 and 13 × 13 convolution kernel sizes.

3. Results

3.1. The Training Process and Results of DCGAN

The experimental environment is presented in

Table 3. Parameters of the computer environment.

Name	Parameter
System	Windows 10
CPU	Inter Core i7-11800H CPU @ 2.3 GHz
Memory	8 GB
Graphics card	NIVIDA GeForce RTX3060
Environment	Python 3.6, Tensorfolw 2.8.0, Keras 2.8.0, Numpy 1.22.2
Configuration	CUDA 11.2

. The training process of the neural network involves 2000 epochs. A batch size of 64 is used, and the weight parameters of both the discriminator and generator are automatically saved every 50 epochs. The loss function adopts the binary cross entropy, which can be expressed by Equation (9) [38]. The loss curves of the generator and discriminator are presented in Figure 4. The training process is mainly divided into two phases (an unstable phase and a stable phase). The loss values fluctuate greatly in the unstable phase, and the loss curves convergence gradually. Real-time visualization allows for visual inspection of the samples. As shown in Figure 5, the training results of four representative epochs are selected. In the early phase of training (approximately 0~1200 epochs), the generated cracks have obvious defects, and the morphology of the cracks is only slightly learned by the neural network. In approximately 1200~1600 epochs, the morphology of the cracks gradually appears, but the generated results are still unstable. The quality of the produced cracks notably improves as the training progresses into its later stages.

$$\mathrm{Loss}=-\frac{1}{N}{\displaystyle \sum _{i=1}^N{y}_i\cdot \log {\widehat{y}}_i}+(1-y_i)\cdot \log (1-{\widehat{y}}_i)$$

(9)

In this equation, N is the output size of the predicted results of neural network, $y_i$ is the true value of the i label, and ${\widehat{y}}_i$ is the predicted value of the i label.

For computational efficiency, the best training results are chosen from the images generated at the 1950th training epoch. The current weight parameters are saved for future use in generating bridge cracks efficiently. Under our computer hardware conditions, the time cost of generating 1000 artificial crack samples is on the order of 10 s. Figure 6 shows part of original crack images and generated crack images. The diversity in shapes of the generated cracks is evident, including vertical cracks, cross cracks, horizontal cracks, and X-type cracks, which are consistent with the real dataset. From a subjective perspective, the bridge cracks produced by the trained DCGAN effectively encapsulate the primary features of the original cracks. It can be challenging to differentiate the generated samples from the original cracks. Finally, 90 generated bridge cracks are used as the dataset in the following experiments and analysis.

3.2. The Training Process and Results of YOLOv5

The experimental environment is presented in

Table 4. Parameters of the computer environment.

Name	Parameter
System	Windows 10
CPU	Inter Core i7-11800H CPU @ 2.3 GHz
Memory	8 GB
Graphics card	NIVIDA GeForce RTX3060
Environment	Python 3.8.5, Pytorch 1.8.0, NUMPY 1.21.5
Configuration	CUDA 11.2

. With a total of 300 training epochs, the batch size is set to 4. The binary cross entropy, box loss, and object loss are employed as the loss functions of the neural network. The classification loss is evaluated by the binary cross entropy; that is, calculating whether the anchor and the corresponding calibration classification are correct. The effect of the box loss function is to evaluate the location loss; that is, calculating the error between the anchor and the calibration box, which can be expressed by Equation (10). The object loss is calculated by IOU [39], which is the intersection and union ratio between the real and predicted boxes.

$${\mathrm{L}}_{\mathrm{CIOU}}=1-\mathrm{IOU}+\frac{{\rho }^2(b,b^{gt})}{c^2}+\alpha v$$

(10)

$$\alpha =\frac{v}{(1-\mathrm{IOU})+v}$$

(11)

$$v=\frac{4}{{\pi }^2}{(\arctan \frac{w^{gt}}{h^{gt}}-\arctan \frac{w}{h})}^2$$

(12)

In Equations (10)–(12), $w$ and $w^{gt}$ are the weight of the predicted and real boxes, respectively; $h$ and $h^{gt}$ are the height of predicted and real boxes, respectively; $b$ and $b^{gt}$ are the central points of the predicted and real boxes, respectively; $\rho$ represents the Euclidean distance between two central points; $c$ is the diagonal distance across the smallest enclosed area that is able to encompass both the predicted and real boxes.

The learning rate is a very important hyperparameter in the neural network, which will influence the accuracy and speed of the training process [40]. The learning rate curves of the training process of YOLOv5 are shown in Figure 7. The warm-up algorithm is used in the initial phase of training (phase 1 in Figure 7). The purpose of the warm-up algorithm is as follows [41]: At the beginning of training, the weights of the model are randomly initialized, and at this time, a larger learning rate may bring instability (oscillation) to the model. However, the warm-up algorithm can involve several epochs with a small learning rate at the beginning of training. Therefore, the model can slowly tend to stabilize, then select the pre-set learning rate for follow-up training, which makes the model convergence faster and the model effect better. The cosine annealing algorithm is used throughout the remainder of the training (phase 2 in Figure 7) [42]. The idea of the cosine annealing algorithm is as follows: The cosine function is used by the cosine annealing to reduce the learning rate. The cosine value first decreases slowly with an increase in x in the cosine function, then it decreases more rapidly followed by decreasing slowly again, which satisfies the requirements of the learning rate of the gradient descent algorithm. In this study, the initial learning rate and the pre-set learning rate are set to 0.001 and 0.01, respectively.

The original dataset (including 180 sets of training samples and 10 sets of validation samples) and the extended dataset (including 180 sets of training samples and 10 sets of validation samples) were established. The samples in the original dataset are all real images, while half of the samples in the training dataset are generated. The detection accuracy and model performance were evaluated using Precision, Recall, MAP-0.5, and MAP-0.5:0.95. The precision and recall can be expressed by Equations (13) and (14), respectively. AP is the average accuracy; that is, the area under the PR curve (recall on the horizontal axis and precision on the vertical axis) of a specific classification in all images. MAP is the average of all the classifications of AP in all images. Therefore, MAP-0.5 represents the value of MAP when IOU is equal to 0. When IOU is equal to 0.5, 0.55, 0.6, 0.65, 0.7, 0.75, 0.8, 0.85, 0.9, and 0.95, MAP-0.5:0.95 are the average values of MAP, respectively.

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

(13)

In Equation (13), TP is the number of positive sample (IOU ≥ threshold) with the correct classification, and FP is the number of negative samples (IOU < threshold) with the wrong classification.

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

(14)

In Equation (14), FN is the number of positive samples with the wrong classification.

Each loss curve and index of the training process is shown in Figure 8. It can be seen that the loss and accuracy curves of original and extended datasets almost coincide. The final loss and accuracy convergence values are also approximate. Therefore, the trained YOLOv5 based on the original dataset and extended dataset have similar detection accuracies. Moreover, two representative YOLOv5 models (YOLOv5m and YOLOv5l) were trained on the extended dataset and compared with the presented model [43]. Figure 9 shows the approximate training losses and detection accuracy of the three models. However, the model sizes of YOLOv5m and YOLOv5l are larger, leading to longer training times and greater GPU memory requirements. Figure 10 shows the detection results based on the original dataset (Figure 10a–c) and extended dataset (Figure 10d–f) of the trained YOLOv5. Under our computer hardware conditions, the time cost of detecting one crack image is on the order of 10 ms. Zheng et al. [44] used an improved YOLOv5 to carry out automatic concrete pavement crack detection, and the time cost of detecting one crack image was on the order of 5.1 ms. Pei et al. [45] also employed a deep learning method to extend the dataset and the faster R-CNN to detect the cracks. The average precision was approximately 0.86, but they used a large dataset containing 1000 real images and 3000 generated cracks. The results indicate that the proposed method is effective.

4. Discussion

Large-scale engineering projects play a pivotal role in modern society, ranging from infrastructure development to environmental protection. However, the success and sustainability of these projects heavily rely on their post-construction maintenance and monitoring. Effective post-construction management is important for ensuring the long-term functionality and safety of such projects [46]. One primary reason for the significance of post-construction maintenance and monitoring is the dynamic properties of engineering structures and systems. As these structures are exposed to changing environmental conditions, they undergo various forms of deterioration over time. Without proper maintenance, this deterioration can escalate, resulting in impaired functionality, decreased performance, and compromised safety. Regular inspections, repairs, and replacements are essential to rectify damage, avoid catastrophic failures, and extend the lifespan of large-scale engineering projects [47]. Additionally, post-construction monitoring provides crucial data for understanding the behavior and performance of these projects in real-world conditions. By collecting and analyzing information on structural stresses, vibrations, deflections, environmental impacts, and defects, it is possible to identify potential issues and optimize the design and operation of similar future projects. Such monitoring also facilitates prompt interventions, allowing for early identification of problems and effective maintenance strategies [48].

Bridge crack detection is an important task in large-scale engineering project monitoring, as cracks on bridges may pose significant hazards [49]. It is important to recognize that the severity and potential hazards associated with cracks depend on various factors, including the type, size, and location of the crack, as well as the overall condition of the bridge. The damage to the bridge caused by cracks can also be diverse [50,51]:

(1)

Cracks can weaken the structure, making it susceptible to sudden failure, particularly under heavy traffic loads or extreme weather conditions.

(2)

Cracks can accelerate the degradation and deterioration of bridge materials. Moisture penetration through cracks can promote corrosion in reinforced concrete or steel elements, further compromising their structural integrity.

(3)

Cracks can affect the dynamic behavior of bridges, leading to decreased stability and increased vulnerability to external forces, such as earthquakes or strong winds. Moreover, fatigue cracks caused by repetitive loading and stress cycles introduce a gradual deterioration process that can eventually lead to catastrophic failure.

(4)

A key concern related to cracks on bridges is their impact on the safety of transportation users.

(5)

Settlement cracks, arising from the differential settlement of bridge foundations, can cause misalignments and deformations that affect the overall stability and functionality of the structure.

To address the hazards posed by cracks on bridges, effective maintenance and repair strategies are essential. Routine inspections and monitoring programs play a crucial role in detecting cracks in early stages and evaluating their severity. Depending on the size and extent of a crack, appropriate repair techniques, such as crack injection or grouting, need to be implemented to restore the structural integrity. In conclusion, cracks pose significant hazards to bridges, jeopardizing their structural integrity and safety. Understanding the causes, types, detection methods, and consequences of cracks in bridges is crucial for designing appropriate maintenance and repair strategies [12].

However, traditional methods are costly for massive bridges spanning large distances. Further research and technological advancements in the field of rapid crack detection technologies are instrumental for ensuring the long-term sustainability of bridge infrastructure and ensuring the safety of the public. Deep learning provides a new method for solving this problem with the rapid development of computer technology [52]. Firstly, drone technology can be used to sample the entire bridge [53]. This technique has the following advantages: (1) It is safe and reliable. The use of drones for inspecting bridges eliminates the need for manual inspection, avoids casualties, improves operational safety, and saves inspection costs. (2) The unmanned detection accuracy of bridges is high. The drone itself can carry high-definition cameras to take images of cracks. (3) It is more efficient to use drones for bridge inspection. Drone sampling technology is relatively mature. The proposal of collision-tolerant UAVs along with a two-stage inspection method for bridge coating was put forward by Jiang et al. [54]. Junwon et al. [55] extensively described the principles of drone-facilitated inspections as well as key factors to consider for optimal data collection.

The images of the entire bridge captured by the UAV can be divided into many small-sized pictures and entered into a computer for crack detection. Even though rapid detection of cracks can be achieved using deep learning techniques, it still requires a large dataset to train the neural network. Additionally, it should be pointed out that the existing small sample training strategy is not sufficient to solve the issue of data scarcity [56]. In order to further reduce the cost of data acquisition and manual detection, the DCGAN can be employed to generate artificial crack samples from existing crack images. Peng et al. [57] used the DCGAN to reconstruct the oil reservoir fracture model. The results showed that the reconstructed model could predict the pressure distribution accurately.

Finally, YOLOv5 is used to carry out the detection process. YOLOv5 applies the adaptive anchor. Therefore, YOLOv5 learns the best anchor boxes in the dataset during the training process without having to run the K-means clustering algorithm offline to obtain the K anchor boxes and modify the head network parameters. Overall, the training process of YOLOv5 is simple and automated. Moreover, the Focus module is used by YOLOv5, which augments the channel count (the channel count has a minimal influence on the computation quantity) and reduces the dimensions of the input image without losing pixel information. As a result, the model calculation amount is greatly reduced [58].

This study has several limitations. The quality of the samples generated by the DCGAN needs to be improved, which is influenced by many factors like the size of the training dataset, the computing power of the computer, and hyperparameter optimization. During the training of the DCGAN, problems such as mode collapse and image collapse may also be encountered [38]. In future work, we will investigate more efficient DCGAN training strategies and achieve a higher-quality reconstruction of the cracks.

5. Conclusions

A method combining DCGAN and YOLOv5 which can detect bridge cracks rapidly and accurately was proposed. Moreover, we described the theory, training environment, parameter setting, and neural architecture of the DCGAN and YOLOv5. This work investigated the performance of the proposed model and compared the training results of the extended dataset with the original dataset. The main findings are presented as follows:

(1)

The trained DCGAN can learn the characteristics of cracks and quickly generate a large number of artificial bridge crack images which are used to extend the real dataset. The time cost of generating 1000 artificial crack samples was on the order of 10 s. The generated images were balanced and feature-rich.

(2)

The YOLOv5 target detection neural network can perform crack identification and rapid detection. The time cost of detecting one crack image is on the order of 10 ms.

(3)

The results indicate that when YOLOv5 was trained on extended dataset, it had a similar detection accuracy compared with when it was trained on the original dataset (real dataset), which provides a new idea for the cost control of maintenance and monitoring of large-scale concrete structures.

The proposed method of combining DCGAN and YOLOv5 has been proven to be acceptable, especially in terms of cost-effectiveness. However, the quality of the images generated by the DCGAN and the detection accuracy of YOLOv5 need to be improved. These factors are influenced by factors such as the computing power of the computer, hyperparameter optimization, and training strategy. In future work, advanced training strategies and optimization algorithms for neural networks will be the focus of our research.