Deep Learning-Enabled Automatic Detection of Bridges for Promoting Transportation Surveillance under Different Imaging Conditions

. The reliable and accurate detection of bridges plays an important role in imaging-driven transportation surveillance. It is capable of timely providing the traﬃc information, leading to safer and more convenient transportation. However, the visual quality of observed images is often inevitably reduced owing to the adverse weather conditions, e.g., haze and low lightness. It is still diﬃcult to adopt the existing powerful deep learning methods to reliably and accurately detect the bridges under diﬀerent imaging conditions. To achieve satisfactory bridge detection results, we ﬁrst propose to exploit the data augmentation strategy and physical imaging method to generate the natural-looking experimental dataset, which contains latent high-quality images and their hazy and low-light versions. We then investigate how to further promote the deep learning-based bridge detection methods through the manually generated dataset. It is obvious that the generalization abilities of these deep neural networks are signiﬁcantly improved using this data augmentation strategy. In this work, we constructed an original dataset consisting of 3500 images of size 900 × 600, collected under normal imaging condition. Extensive detection experiments will be performed based on the augmented dataset. Experimental results have demonstrated that our automatic bridge detection framework could generate more reliable and accurate results compared with existing detection methods.


Introduction
With the rapid developments of unmanned aerial vehicles (UAV) and remote sensing techniques [1,2], it has become flexible to capture the visual data of bridges in a wide range [3]. e synthetic aperture radar (SAR) images [3] are able to expand the imaging range containing the bridges. However, the SAR images inevitably suffer from the random noise and low-contrast signal. In addition, it is not flexible to timely and accurately capture the visual images related to the traffic flow on bridges. In addition, the visual data collected from UAVenabled imaging sensors are able to generate meaningful and important information in transportation surveillance. It is capable of timely providing the traffic information, leading to safer and more efficient transportation. Due to the poor weather conditions (e.g., haze [4] and low-light [5]), the visual quality of observed images is often inevitably impaired in practice. Under poor imaging conditions, it will be challenging to detect the bridges precisely. In recent years, considerable effort has been invested in the detection of bridges under normal imaging conditions. Traditional detection methods, deep learning-based methods, and combined detection methods comprise the three main categories of general detection techniques. Deep learning-based methods have significantly improved their detection precision and efficiency due to their powerful learning and representation capabilities. However, the degraded images provide few details in terms of geometrical structures and color appearance, making it difficult to detect bridges reliably and precisely under adverse imaging conditions. e deep learning-based detection methods are highly dependent on the training dataset. To achieve satisfactory bridge detection results, we will construct an original dataset with 3500 images of size 900×600, collected under normal imaging conditions. To promote the volume and diversity of the training dataset, we will exploit the data augmentation method based on physical imaging modeling [6][7][8] to enlarge the original image dataset. In particular, the enlarged dataset will contain the latent sharp images and their hazy and low-light versions. Furthermore, four typical deep learning-based detection methods (i.e., Faster R-CNN [9], YOLOv3 [10], YOLOv4 [11], and YOLOv5 [12]) will be introduced to implement the detection of bridges of interest. In this work, we will investigate how to further promote the deep learning-based bridge detection methods through the enlarged dataset. Existing studies have demonstrated that the generalization abilities of these deep neural networks could be significantly promoted according to the data augmentation strategy. e proposed deep learning-enabled automatic computational framework could significantly improve the reliability and accuracy of bridge detection. erefore, the main contributions of this work can be summarized as follows: (1) We proposed a deep learning-enabled automatic computational framework for detection of bridges in transportation surveillance. (2) We construct an original image dataset consisting of 3500 images of size 900×600 captured under standard imaging conditions. Owing to the physical imaging methods, the synthetically degraded images are generated to expand the image dataset and enhance the generalization capabilities of deep learning methods. e expanded dataset is used to train deep learning-based detection methods, resulting in more reliable and accurate detection of bridges under various imaging conditions. (3) Experiments have shown that our automatic framework can produce more robust and accurate detection results than existing methods. e satisfactory detection performance is primarily attributable to the robust learning ability of neural networks and the physical imaging-based data augmentation strategy. e remainder of this work is structured as follows. Section 2 briefly reviews the recent work related to bridge detection. Four typical deep learning-based bridge detection methods and data augmentation strategies are introduced in Section 3. Numerous experiments are implemented to illustrate the effectiveness of our method in Section 4. is work is finally ended by giving the main contributions in Section 5.

Related Work
Recent efforts in the literature have focused on the automatic detection of bridges of interest. e current detection methods can be roughly classified into three categories: traditional, deep learning-based, and combined versions. In this section, we will review these detection methods briefly.

Traditional Bridge Detection Methods.
As early as 1989, Baker et al. [13,14] studied the detection of concrete bridges in close-up side-shot images. It is assumed that the sides of the bridge had an approximate rectangular geometry. erefore, there would be a pair of parallel lines on the upper and lower sides of the bridge. e rectangular side could be obtained by detecting the parallel lines. According to this assumption, Wang et al. [15] proposed a multistep bridge detection framework. It first extracted the river regions from the image and then selected the regions between two disparate water areas as candidate selections. e final detection results could be obtained by extracting parallel lines from these regions. To further enhance detection performance, more structural correlations between bridges and rivers have been exploited [16]. By considering both local radiometric and textural features, Loménie et al. [17] proposed a computational system for robustly detecting bridges from highresolution satellite images. According to the image prior knowledge, an integrated method was developed [18] to automatically detect the bridges over rivers from satelliteborne visual images. In particular, this method first extracts the river areas and then detects the brides of interest through the knowledge-driven strategy. To take full advantage of the spatial relation between bridge and water, Liang et al. [19] developed an automatic computational method to detect bridges using the HJ-1 satellite remote sensing images. e multispectral and morphological information in images were combined to guarantee bridge detection. e above computational methods mainly focus on the extraction and processing of line features rather than the position relationship between the bridge and river. However, methods only based on the parallel line features of the target will easily generate unsatisfactory detection results. To overcome this limitation, Sita [20] first proposed a method for template matching, moment feature matching, and transformation feature matching according to the geometric or regional features, moment features, and transformation features of the target. On this basis, Wang et al. [21] proposed a depict segmentation method for recognizing bridges, which used Hough transform to extract the longest straight lines from the coarse-resolution images. ough these methods perform well in detection efficiency and accuracy, the parameter selection of the Hough transform has a great influence on the result. It is not suitable for the situation where there are multiple similar targets in the image.
To effectively solve the above problems, Han et al. [18] first used the grey level cooccurrence matrix (GLCM) method to obtain the entropy, contrast, homogeneity, energy, and other characteristics from the remote sensing image. en, they used these features to divide the image into river and land areas and finally detected bridges by using prior knowledge. As a consequence, Fu et al. [22] first established knowledge models of the bridges of interest. ey utilized concurrently the segmentation of waters, the extraction of regions of interest (ROI), the connectivity signal, and the detection of candidate regions for the approximate positioning of bridges. During the testing procedure, the grey characteristics were used to identify the potential bridge. To take full advantage of spatial information, both structural information and topological relations in bridges regions were considered to enhance bridge detection performance [23]. In addition, Liu et al. [24] proposed a bridge information extraction method based on multisource data fusion, which is mainly based on the representation of bridges, land, and rivers, by fusing panchromatic and near-infrared images to recognize and extract bridge information automatically. e premise of these methods is that the road information in the image is available. For the positional relationship between roads, rivers, and bridges to be unknown, Shu-Kui and Nie [25] proposed a method for water bridge recognition which is object oriented, which used the region growth method to segment the image. According to the special reflection characteristics of the water body, the image objects generated after segmentation were used as the basic units to be classified. en, based on the characteristics of the bridge, they extracted the bridges from images by using the shape feature of image objects and the contextual relationship between bridges and water bodies. However, due to the limitation of spatial resolution, the detection of bridges on some small tributaries was not perfect. Subsequently, through the research on the morphological filter and regional growing, Gu et al. [26] proposed a method of bridge extracting based on the difference between the above filtering methods. By combining the characteristics of bridges, the bridge information can be extracted from the difference in filtered DEM. e proposed method was capable of detecting different bridge designs. Based on the high spatial images, Yuan et al. [27] developed a target-oriented computational method to automatically extract bridges from original images. Firstly, they selected the optimal segmentation scale through multiscale segmentation experiments and the underlying surface features; secondly, they established a rule set by using water body index, threshold function, and other methods. ey gradually obtained vectors of water bodies and potential areas of bridges. Finally, they successfully extracted the bridge through binarization, mathematical morphology processing, overlay analysis, and other methods.

Deep Learning-Based Bridge Detection Methods.
e above-mentioned bridge detection methods are highly depended on traditional image processing techniques. e corresponding detection results often suffer from several limitations, such as unstable detection performance under complex imaging conditions. Besides, it becomes difficult to generate high accuracy and essentially fails to generate real-time detection. As the rapid development of artificial intelligence technique [28,29], the deep learning-based detection methods have gained great success in the fields of target detection and recognition, due to its powerful learning capacity.
In the literature, the existing deep learning-based target detection methods can be broadly divided into two classes, i.e., the two-stage and one-stage methods. e representative two-stage detection methods mainly include the R-CNN [30] and its extensions, such as Fast R-CNN [31], Faster R-CNN [9], and mask R-CNN [32]. ese methods first extract the candidate bounding boxes according to the position of the target of interest. e classification and regression are then implemented to produce the detection results. ese methods are capable of accurately detecting the bridges from the original images. However, they inevitably suffer from high computational load, leading to unreal-time detection of bridges in practice. Fortunately, the introduction of one-stage detection method could deal with this limitation. e one-stage detection method mainly include the single shot multibox detector (SSD) [33], as well as you only look once (YOLO) [34] and its extensions. e popular YOLO [34] directly estimated the category probability, bounding box, and class confidence. It is able to significantly improve the detection efficiency. YOLOv2 [35] and YOLOv3 [10] have enhanced the capacities of feature extraction by introducing the Darknet-19 and Darknet-53, respectively. e target detection results were improved accordingly. More recently, the newly developed YOLOv4 [11] and YOLOv5 [12] have attracted considerable attention and could significantly improve the detection accuracy, efficiency, and robustness.
With the rapid developments of airborne and satellitebone imaging systems, it becomes more flexible to detect the bridges to monitor traffic status. For example, based on the satellite-bone SAR images, Chen et al. [36] proposed an advanced bridge detection method, termed as single shot detection-adaptive effective feature fusion (i.e., SSD-AEFF), to perform accurate bridge detection from complex SAR images. e multiresolution attention and balance network (i.e., MABN) [3] were also proposed to identify bridges from SAR images. e proposed network is mainly composed of three parts, i.e., the attention and balanced feature pyramid module, the region proposal network (RPN), the regression, and classification. However, the SAR images often suffer from the random noise and low-contrast signal. It is thus difficult to detect small-scale bridges. In contrast, the optical aerial images become more popular and practical for bridge detection. By taking into consideration the auxiliary water body extraction task, an accurate bridge detection network [37] has been presented, which could enhance the feature representation through semantic context. However, the quality of optimal images is often degraded due to the poor weathers, such as hazy and low lightness, leading to unsatisfactory detection results. In this work, we will incorporate the data augmentation strategy into existing deep learning method to make bridge detection more robust and accurate under complex imaging conditions.

Combined Bridge Detection Methods.
It is sometimes difficult for a single detection method to generate satisfactory detection results. To enhance the detection results, Mirmehdi et al. [38] proposed to introduce the specific Journal of Advanced Transportation 3 control strategies to enhance the traditional methods. Based on the joint features and knowledge rules, Sang et al. [39] proposed a new method for quickly detecting bridges on water from visible light remote sensing images. e whole detection framework could be mainly divided into four steps, i.e., water area segmentation, preliminary bridge edge detection, matching, and verification to remove false bridges. It is well known that traditional methods and deep learning methods have their own advantages and disadvantages. ere is a potential to combine the advantages to further enhance bridge detection. erefore, a hybrid detection framework, proposed by Loménie et al. [17], has been proposed to implement robust detection. is hybrid version exploited a voting strategy to classify the image pixels into the corresponding classes and introduced the bottomup and top-down parts to accurately detect bridges. Tria-Sanz and Loménie [40] first introduced the radiation features to classify the pixels into different geographic types and then searched for bridges based on the above classification.
By introducing the prior-information auxiliary module (PAM) [41], which is able to receive and integrate priorinformation feature maps, four popular target detection networks, i.e., YOLOv3 [10], YOLOv4 [11], Faster R-CNN [9], and CenterNet [42], have been extended to obtain more meaningful features. By integrating the prior-information feature maps into the existing neural networks, the domainspecific knowledge will be more meaningful, beneficial for promoted bridge detection.

Deep Learning-Enabled Automatic Detection of Bridges
is paper mainly focuses on the automatic detection of bridges under hazy and low-light situations. To improve the generalization ability of our detection method, we first synthesize both hazy and low-light images to expand the original dataset. We then train the object detection methods and compare the computational accuracy and efficiency under the same imaging conditions. In particular, several typical networks, e.g., Faster R-CNN, YOLOv3, YOLOv4, and YOLOv5, will be exploited to detect the bridges of interest in this work. Experiments show that our automatic framework can produce more robust and accurate bridge detection results than existing methods.

Faster R-CNN.
Faster R-CNN has recently become a typical two-stage target detection method. In particular, it first estimates the target candidate regions. ese candidate regions are then classified and regressed, beneficial for accurate and robust target detection. is is due to the fact that Faster R-CNN exploits the RPN to generate candidate regions on the feature map. Compared with the selective search (SS) algorithm [43], the RPN reduces the number of candidate regions for each image to 300, which significantly reduces the calculations and greatly improves the efficiency during the network training and test. To make it easier to understand, the architecture of Faster R-CNN is visually shown in Figure 1. e widely used Faster R-CNN is mainly composed of four components, i.e., backbone network, feature pyramid network (FPN), RPN, and feature branches. e backbone network is capable of extracting the feature information from the original input image. e RPN is a fully convolutional network, which greatly reduces the calculations because it does not contain a fully connected layer. e input of RPN is a feature map of any size. In addition, its output is a batch of candidate frame position information and the probability of whether there is a target. Because the size of feature map of the fully connected layer is fixed, the ROI pooling layer is required to perform feature transformation on the candidate region.
is part is achieved by the local max pooling. During the network training, the input image is firstly scaled to prevent distortion. e shared feature map is then extracted through the backbone network for subsequent RPN and ROI pooling. e RPN is composed of two layers, i.e., the convolutional layer and the activation function layer. In particular, it exploits the anchor mechanism to generate anchor boxes for three different scales and three aspect ratios. Let the size of the feature map of the input RPN be w × h, and a total of w × h × 9 anchor boxes are generated accordingly. After screening the candidate frames, the feature extraction is performed through a convolution kernel of size 3×3. Two 1×1 convolutions are then exploited for target classification and frame regression. It becomes flexible to obtain the position information of the candidate frame and the probability of containing the targets.
To obtain the candidate area, the candidate frame obtained through the RPN is mapped to the last layer of the feature map. e feature size transformation is then performed through the ROI pooling layers. e object classification and bounding-box regression are finally performed in the detection network. During the training process, the multitask loss is utilized to jointly train the regression and classification networks in practical applications.

YOLOv3.
e popular one-stage target detection network YOLOv3 essentially considers the target detection task as a regression problem. Compared with the Faster R-CNN, it avoids the generation of candidate regions. It directly performs classification and regression on the feature map, which greatly improves the detection efficiency and reduces the computational cost compared with traditional two-stage detection networks.
As shown in Figure 2, YOLOv3 can predict targets of interest at three different scales. By dividing the image into grid cells of the same size (i.e., 13 × 13, 26 × 26 or 52 × 52), each grid point is responsible for the prediction of a region. e final prediction result is represented using a 3-d tensor, which contains the position information of the predicted bounding box, the confidence score, and the class predictions. In particular, YOLOv3 utilizes the Darknet53 as a feature extractor, which applies the convolution kernels of size 3 × 3 and 1 × 1, alternately. It also employs the skip connections to enhance the ability of the network to extract meaningful features. In the literature [10], YOLOv3 can extract feature maps with three different scales, which are By strengthening the capacity of feature extraction, the prediction heads could exploit the feature maps by FPN to robustly and accurately predict the bounding-box coordinates and class probabilities. e YOLOv3 network firstly resizes the input image to a fixed size and then divides the image into grids of three scales. e k-means clustering is introduced to obtain a priori boxes of each different scale. e important feature extraction strategy is performed on the image through the backbone network.
rough an existing multiscale prediction method, the feature maps with three different sizes extracted from the model are utilized to predict the final results.
e feature fusion is finally performed on the prediction results of the three scales. e network filters out the bounding boxes whose confidence is lower than the predefined threshold. e final bounding boxes can be obtained through the widely used nonmaximum suppression (NMS).

YOLOv4.
YOLOv4 is an extension of traditional YOLOv3, which is able to generate more robust and accurate detection results. In particular, YOLOv4 exploits the mosaic data augmentation method to enrich the existing dataset.
is method randomly scales four training images and then randomly stitches them into a new image. is generation process of image dataset could make the detection network more robust and general under complex imaging conditions. As shown in Figure 3, YOLOv4 directly exploits the CSPDarknet53 as the backbone network to extract meaningful image features. Unlike the Darknet-53 by YOLOv3, YOLOv4 further modifies the residual block structure on Darknet-53. It also tends to replace the Leaky ReLU with the Mish activation function in the backbone network. Different from the FPN utilized as the parameter aggregation method in YOLOv3, the more powerful path aggregation network (PANet) is introduced in YOLOv4. e neck is exploited to fuse image features and transfer these features to the prediction layer. In the neck part, the spatial pyramid pooling (SPP) [44] is introduced to increase the receptive field of the network. In addition, YOLOv4 employs the PANet to fuse and combine image features to promote the capacity of target detection with different scales, which is obviously different from the FPN utilized in YOLOv3. e head of YOLOv4 network is responsible for robustly and accurately predicting the bounding boxes and object classification.

Journal of Advanced Transportation
In addition, YOLOv4 commonly exploits the complete intersection over union (CIoU) as the loss function, which considers not only the overlap area of the bounding box, but also the distance from the center point and the aspect ratio. As a consequence, the regression speed and accuracy, related to the bounding-box prediction, are improved accordingly.

YOLOv5.
YOLOv5 also exploits the mosaic data augmentation method to enrich the image dataset. However, unlike the YOLOv4, it employs the mechanism of auto learning bounding-box anchors to learn the anchor boxes based on the enriched training data. e architecture of YOLOv5 network is visually illustrated in Figure 4. It can be observed that the backbone of YOLOv5 directly combines the focus structure and CSP structure. e focus layer slices the input image and warps H × W × 3 into H/2 × W/2 × 12.
is method allows more detailed information to be retained during the downsampling process, which can prevent the information loss in practical applications. e neck part contains FPN and PAN structures. In particular, the FPN transfers and integrates the high-level feature information through the upsampling from top to bottom. It is able to convey the robust semantic features accordingly. In addition, the PAN is a bottom-up feature pyramid which conveys the robust positioning features. Both FPN and PAN are simultaneously utilized to strengthen the network feature fusion capabilities. In the literature [12], YOLOv5 employs the generalized intersection over union (GIoU) as the loss function to constrain the prediction of bounding boxes. is strategy could effectively improve the performance for target classification and bounding-box regression. e nonmaximum suppression (NMS) is then exploited to eliminate the redundant bounding boxes and generate the optimal detection results.

Data Augmentation Method.
To evaluate the detection performance of our learning method, it is necessary to enlarge the original image dataset under different imaging conditions. Both hazy and low-light imaging conditions will be considered to generate the degraded images.

Synthetic Generation of Hazy Images.
According to the popular atmospheric scattering theory [6], the scattering of light by particulate impurities suspended in the air is the main reason for the degradation of hazy images. e incident light attenuation model refers to the attenuation of the reflected light on the surface of the object due to the scattering effect, which reduces the intensity of the light reaching the imaging system. eoretically, as the propagation distance increases, the intensity of reflected light on the target surface will decay exponentially.
It is well known that the atmospheric scattering model has been widely exploited to describe the formation process of haze images, which can be expressed as follows: (x, y) denotes the pixel location, I is the observed hazy image, J denotes the latent sharp image, d(x, y) represents the depth at the pixel (x, y), β is a scattering coefficient which denotes the medium density, and A is the global airlight.
By manually setting different scattering coefficients and global airlights, we can synthesize different kinds of hazy images, contributing to an enlarged dataset which contains both original and hazy images. In our numerical experiments, we propose to randomly select different values of A and β, i.e., A ∈ [0.7, 1.0] and β ∈ [0.4, 0.8].

Synthetic Generation of Low-Light Images.
e Retinex theory has been widely used to synthetically generate the low-light images. In particular, the Retinex theory could be exploited to decompose an image into two independent components, i.e., reflectance and illumination components. It is well known that the color of a target observed by the human eyes is only related to the reflection component on the target surface. e Retinex-based image decomposition is defined as follows: where S(x, y) is the observed image, L(x, y) represents the light intensity (i.e., illumination component), and R(x, y) represents the reflection component of the object image.   Based on the Retinex theory, for each clean illumination image R, a low-light image can be synthesized by randomly setting a random illumination value L.

Experimental Results and Analysis
According to the representative target detection methods introduced in Section 3, we will implement extensive experiments using Faster R-CNN, YOLOv3, YOLOv4, and YOLOv5. In this study, all detection experiments are performed with Ubuntu 18.04 system and run in Intel ® Core i9-9900X CPU @ 3.50 GHz, 128 GB ram, and a single NVIDIA GeForce RTX 2080 Ti computing environment. It is worth noting that the YOLOv5 mainly includes 5 different types, i.e., YOLOv5s, YOLOv5m, YOLOv5l, and YOLOv5x. We propose to select the YOLOv5s in this work, which is the lightweight version to efficiently and robustly detect the targets of interest.

Evaluation Criteria.
In our experiments, two main evaluation metrics, i.e., frames per second (FPS) and average precision (AP), will be introduced to evaluate the detection performance. In particular, the FPS represents the execution efficiency for bridge detection. It is the number of FPS that the detection network is able to perform bridge detection. In addition, the AP denotes the area surrounded by the PR (precision-recall) curve. Instead of AP, we propose to use the average version (i.e., mAP) to evaluate the detection accuracy. e mAP represents the mean of the AP value of all classes. It is pointed out that there is only one class of bridge considered in our experiments. e value of mAP is thus equivalent to AP. Both FPS and mAP have become the most widely used standards for measuring target detection performance. e mAP (or AP) is defined as follows:

Experimental Dataset and Settings.
e original image dataset consists of 3500 images of size 900×600 collected under normal imaging condition. In our experiment, we take 67% of the bridge images for network training and the others for testing. Each image is annotated with correct bridge labels and ground truth boxes. In order to explore the influence of data augmentation on bridge detection, we enlarge the original image dataset according to the generation of synthetically degraded images shown in Section 3.5. In Figures 5 and 6, both hazy and low-light images include three different levels. We propose to exploit the TensorFlow 1.13.1 tool package as the framework and ResNet101 as the backbone of Faster R-CNN. We train YOLOv3 and YOLOv5s networks with the Pytorch1.8.1 package and the YOLOv4 network with the Darknet framework. During network training and testing, all detection networks are implemented with the optimal parameters.

Comparisons with Other Detection Methods.
To evaluate the detection performance, we have implemented several experiments on bridge detection under normal imaging conditions. e bridge detection results can be visually illustrated in Figures 7-10. In particular, four typical methods, i.e., Faster R-CNN, YOLOv3, YOLOv4, and YOLOv5s, are simultaneously introduced to detect bridges of interest in our experiments. It can be found that it is easy to perform robust and accurate detection at the presence of large-scale bridges, as shown in Figure 7.
ese high-quality bridge detection results are beneficial for safer and more convenient transportation. From a theoretical point of view, the cooccurrences of water and land in Figure 8, single backgrounds in Figure 9, and complex backgrounds in Figure 10 could make bridge detection more difficult in practical applications. Due to the more powerful learning capacity, YOLOv5s is able to implement more robust and accurate detection results. e suboptimal detection results, yielded by Faster R-CNN, YOLOv3, and YOLOv4, could lead to poorer transportation surveillance.    Journal of Advanced Transportation As shown in Table 1, the models are trained in original images and then used to test the bridge detection model in original, low-light, and hazy images, respectively. e mAP results from high to low is YOLOv4, YOLOv5s, Faster R-CNN, and YOLOv3. e FPS performance of the model from high to low is YOLOv5s, YOLOv4, YOLOv3, and Faster R-CNN. e models can meet the needs of real-time bridge detection, except for Faster R-CNN, of which FPS is only 8. In Table 1    conditions. Figures 15 and 16 depict the detection results under low-light and hazy imaging conditions, respectively. e Faster R-CNN method has been found to have excellent detection accuracy. However, this method typically has a lengthy computational time. It is thus impossible to detect the bridges in real time. e detection capability of YOLOv3 and YOLOv4 may be impaired in low-light and hazy environments, whereas YOLOv5s have the best accuracy and speed performance.

Discussion.
In practical applications, the purpose of data augmentation is to reduce the negative effects of bridge detection. We can conclude from the imaging experiments that hazy and low-light images have different effects on the detection of bridges. According to exhaustive experiments, both hazy and low-light images can easily have a negative impact on the representation of bridge features. In practice, the effect of low-light conditions is weaker than that of hazy conditions. e improvement of low-light conditions after the implementation of a data augmentation strategy is superior to that of haze. Not only can data augmentation improve the detection accuracy, but it can also increase the robustness of bridge detection.

Conclusions
In this work, we propose a deep learning-enabled automatic bridge detection method for promoting transportation surveillance under different imaging conditions. It contributes to reliable and robust video surveillance to guarantee more reliable and effective intelligent transportation. e major contributions of this work are as follows: first, a deep learning-enabled automatic bridge detection framework is proposed for transportation surveillance. Second, we construct an original dataset consisting of 3500 images of size 900×600, collected under normal imaging conditions. Moreover, the synthetically degraded images obtained using physical imaging methods are generated to enlarge the imaging dataset to improve the generalization abilities of the deep learning methods. Last, the deep learning methods are effectively trained using the enlarged image dataset, contributing to more reliable and accurate bridge detection results under different conditions. Although our proposed method has already provided a solution for high-accuracy bridge detection, to further improve the accuracy of bridge detection for better application in tasks such as traffic management, we will conduct the related work from the following three aspects.
(1) To improve the generalization ability for the detection of morphologically inconsistent bridges, we will further expand our dataset through field photography, web browsing, etc. In addition, we will perform the collection of bridges in really harsh imaging environments to further improve the robustness of bridge detection. (2) To facilitate the subsequent work of bridge detection, we will introduce a rotated detection frame to match the length and width of the bridge as much as possible. erefore, we intend to redesign the network to further refine the bridge features under highaltitude photography. It will be able to improve the accuracy of small-scale bridge detection.
(3) In the real-world harsh imaging environment, not only fog and low-light are included, but rainy days and sandstorms are also key factors affecting the imaging quality. To be able to adapt to bridge detection in various scenarios, we will develop a bridge detection model suitable for various harsh imaging scenarios in practical applications.

Data Availability
e image data used to support the findings of this study are available from the corresponding author upon request.

Conflicts of Interest
e authors declare that they have no conflicts of interest.