Deep learning based water leakage detection for shield tunnel lining

Abstract

Shield tunnel lining is prone to water leakage, which may further bring about corrosion and structural damage to the walls, potentially leading to dangerous accidents. To avoid tedious and inefficient manual inspection, many projects use artificial intelligence (AI) to detect cracks and water leakage. A novel method for water leakage inspection in shield tunnel lining that utilizes deep learning is introduced in this paper. Our proposal includes a ConvNeXt-S backbone, deconvolutional-feature pyramid network (D-FPN), spatial attention module (SPAM). and a detection head. It can extract representative features of leaking areas to aid inspection processes. To further improve the model’s robustness, we innovatively use an inversed low-light enhancement method to convert normally illuminated images to low light ones and introduce them into the training samples. Validation experiments are performed, achieving the average precision (AP) score of 56.8%, which outperforms previous work by a margin of 5.7%. Visualization illustrations also support our method’s practical effectiveness.

Graphical abstract

Keywords

water leakage detection / deep learning / deconvolutional-feature pyramid / spatial attention

1 Introduction

Shield tunnels are critical infrastructures that require ongoing maintenance and inspection to ensure their safety and functionality. One of the major concerns is the formation and growth of water leakage, which can lead to structural deterioration of the tunnel wall and serious safety problems, such as the corrosion of metal structures. Researchers recommended that tunnels should be maintained regularly [1]. For most maintenance and overhaul work, manual inspection methods are tedious, time-consuming, costly, and may not provide consistent and accurate results. Since the 1980s, engineers have exploited artificial intelligence (AI) and expert systems to automatically detect structural defects.

In recent years, automated defect detection in underground infrastructures has been made possible by the emergence of powerful computer vision technologies, particularly those based on deep learning [2–5]. A deep neural network (DNN) can learn complex patterns and features from sufficient image samples, so it is well-suited for water leakage detection. As is well known, DNN is a data-driven architecture that is commonly hindered in performance by the limited quantity and quality of available samples [6–8]. Fig.1 shows some real images of water leakage taken during tunnel lining maintenance, from which we can observe the extent and location of the leakage spots, including cracks, stains, and erosion on the concrete surface. Such flaws relate three general challenges in applying computer vision techniques for tunnel problems. 1) The images are from low-light conditions and are not conducive to model optimization. 2) Some water leakage spots are similar in character to normal concrete wall joints. 3) Fast running speed of algorithms should be achieved. Our proposal aims to automatically identify and locate water leakage spots from images of shield tunnel linings, taking into account the challenging conditions listed above.

A tunnel defects detector (TDD) is proposed in this paper. The TDD starts with a computationally small baseline model, and finally achieves the state-of-the-art detection performance. In its backbone, TDD adopts CovNeXt-S, which has a satisfactory feature extraction capability while using a small number of parameters. the baseline model is followed by a detection neck consisting of deconvolutions of the backbone’s multi-stage features, by means of a deconvolutional-feature pyramid network (D-FPN). To improve the ability to represent features, each level of D-FPN concatenates with the output feature of the backbone. In the D-FPN design, the top-level feature does not contain the feature from the bottom level. This is unlike practice in using other FPNs. Each neck feature corresponds to one prediction layer. Considering the complex background of the tunnel lining, a spatial attention module (SPAM) is added in our proposal before the detection head to focus on the discriminative regions. The SPAM adopts deformable convolutional network (DCN) with spaced residual connection for better detecting of leakage areas. Adjustments were made to generation of the prior anchor by analyzing the defect region’s aspect ratio distribution, through calculations. To further improve the model’s performance in dark conditions, we generate low-light images from those collected in normal light conditions. The proposed TDD is evaluated based on a public data set [9] that is collected from real work in tunnels, and the results demonstrate its effectiveness and accuracy in detecting water leakage under challenging conditions. The following are the key innovations and contributions of our work.

1) The D-FPN is first proposed, which is different from other FPN variants. It can fuse more feature details of the backbone output, and has a better ability to recognize features on the edges and cracks of the water leakage areas.

2) We propose a SPAM that adopts DCN with spaced residual connection for spatial attention (SPA) operation, working as a feature-optimization module. The application of SPA allows clearer focus on the area of the leakage spot.

3) To support possible low-light conditions in the tunnels, we generate training samples by reducing the exposure of images to make the model adaptive to low-light scenarios.

4) Our method achieves promising performance, as demonstrated by experimental results that show an average precision (AP) of 56.8% based on real images collected during shield tunnel lining maintenance [9].

This paper is structured as follows. Section 2 discusses the basics of computer vision and related research, and Section 3 elaborates on our proposal. Section 4 covers the experimental environment and technical approach, while Section 5 presents a comprehensive analysis of evaluation results and comparisons with other methods. Lastly, Section 7 concludes the paper and outlines potential future work.

2 Related work

Shield tunnel structure is complex. Structural damage may cause accidents and threaten the safety of lives and property [1,10,11]. To avoid disasters, an efficient automated regular maintenance technology is necessary.

The use of AI technology in maintenance and overhaul works can be traced back to as early as 1980s. Kawata et al. [12] developed an expert system using AI and database technologies to achieve automatic inspection and diagnosis of tunnels. With further developments of computer hardware and image processing technologies, recent automated solutions use images captured by lidar, RGB-D cameras, infrared cameras, charge-coupled device/complementary metal-oxide-semiconductor (CCD/CMOS) cameras, etc. Yu et al. [13] designed a tunnel inspection robotic system, which consists of a robot with a CCD camera and image processing algorithms. Hu et al. [14] achieved water leakage edge extraction using a canny filter on gray value images, thus deploying a traditional image processing method that can be used for most defect detection problems with special edge features, as demonstrated in Fig.2. With the spread of smartphones, it is becoming possible to detect water leaks with the phone camera.

Compared to conventional image processing algorithms mentioned earlier, deep learning-based solutions have significant advantages in terms of accuracy and generalizability. Most of the defect detection works are based on classification models. Yokoyama and Matsumoto [15] developed a detection method using a 6-layer classifier, which can determine whether there are cracks in the sliding windows. Similarly, Protopapadakis et al. [16] designed a convolutional neural network-based (CNN-based) classifier for tunnel defects inspection, and achieved 76.0% accuracy on a self-built data set.

However, a classification model is not a satisfactory solution when the cracks are very small. Many previous works are based on pixel-level image segmentation technology [17] for realizing the automation of crack inspection. Pixel-level methods have benefits for eventual quantification. Detection and quantification are two separated tasks. In some engineering scenarios, such as water leakage detection, action is required immediately after detection. Where engineer only needs to determine the approximate location of the leakage area and there are no strong analytical requirements, direct and rapid methods, rather than segmentation methods, can be applied. In water leakage inspection tasks, using object detection technology, defects in the wall or other materials, such as brick cracks and pipe cracks, can be quickly located and identified. Object detectors are generally categorized into two-stage and one-stage detectors. Two-stages detectors generate proposals first, then classify and adjust the proposals to obtain more accurate results, such as Fast R-CNN [18,19], Cascade R-CNN [20], DetectoRS [21]. One-stage detectors make predictions directly without using proposal boxes during the training and inference procedures, such as single shot multibox detector [22], YOLO series [23–26], RetinaNet [27] and so on. Cha et al. [28] proposed a method using Faster R-CNN for detecting multiple types of damage/abnormalities in civil structures, achieving high detection accuracy in a wide range of data, proving the feasibility of object detection in defect detection. Xia et al. [29] proposed a real-time one-stage object detector for railway sleeper crack detection, achieving an accuracy of 98.1%. Xue et al. [30] proposed a structure from motion (SfM)-deep learning method, which could realize three-dimensional (3D) visualization of leakage defects by constructing a high-quality 3D tunnel model onto which the image texture, after defect segmentation, was projected. For leakage detection, the irregular edge patterns of cracks and seeps are crucial features for detection. Guo et al. [31] proposed a road crack detection method using the original edges of images as additional features to produce more accurate crack and leakage boundaries.

3 Methods

3.1 Overall architecture

The overall pipeline architecture of the proposed TDD is depicted in Fig.3, which contains a total of five primary parts.

1) Low-light images are first generated and added into the training set for data augmentation, as shown in Fig.3(a). Reversed Zero-DCE is used, and the data augmentation aims to improve the model’s applicability in inspection work under low-light condition, without extra computational cost.

2) Fig.3(b) demonstrates the backbone of our proposal. The ConvNeXt-S [32] is chosen as the backbone for feature extraction, because it is computationally efficient while also maintaining satisfactory performance.

3) Fig.3(c) shows the developed detection neck, called the D-FPN. It is more suitable for water leakage detection than the commonly used FPN [33].

4) Fig.3(d) is another innovation in this study; the SPAM, which is composed of several SPA blocks to improve the model’s ability to focus on the defects.

5) The final part of TDD is shown in Fig.3(e). Similar to other two-stage object detectors, a decoupled head is used for better prediction of boxes.

3.2 Backbone

The backbone and detection neck make up the feature extraction section of the object detector. The backbone network aims to extract the image features in stages, and the representation quality of the features will significantly affect the final detection performance. Therefore, a strong backbone network is necessary.

ConvNeXt [32] adopts some of the latest ideas and techniques of Transformer networks into the existing modules of CNN networks, thus combining the advantages of both networks. Achieving high performance while maintaining fast speed is also very important, so there is a trade-off between them. Our strategy is to use ConvNeXt-S to satisfy the following requirements: 1) it is stage-designed, and 2) it is lightweight.

3.3 Detection neck

The detection neck is a common structure that serves to aggregate the features of different stages of the backbone, such as FPN [33]. The information loss in the upsampling process of FPN is serious, and negatively affects the edge and texture information of defects. Instead of using a sequence of upsampling, the proposed D-FPN directly deconvolutes the backbone feature map and then concatenates it with the output of the previous layer, as shown in Fig.3(c). The feature for the neck layer ‘nl’ can be represented as

(1)f^{\mathrm{n}\mathrm{l}}=Deconv \left(F^{\mathrm{b}\mathrm{l}-1}\right)\cup F^{\mathrm{b}\mathrm{l}},

where Deconv represents the deconvolution operation, F^bl denotes the backbone layer feature corresponding to ‘nl’.

3.4 Spatial attention module

The D-FPN feature fusion strategy is able to significantly reduce the loss of detailed features. In the commonly used FPN, where the upper-level features are obtained by fusing the upsampled lower-layer feature and the backbone output, a lot of details are lost in the up-sampling process. In D-FPN, features are obtained by deconvolution of backbone outputs, which means the upper-levels are generally not affected by the lower-level features. In addition, the feature pyramid formed by deconvolution of the stage-by-stage feature map will contain noisy features.

In this direction, SPAM is proposed to remove the noise from the background and focus attention on the objects. The SPAM is composed of several SPA blocks, as demonstrated in Fig.4. Structurally, SPAM is a residual network with an odd number of blocks, and its residual structure exists between the first and last blocks. In terms of data flow, for the ith (i > 1) SPA block in SPAM, the output features A_i^{\mathrm{o}\mathrm{u}\mathrm{t}} can be calculated as

(2)A_i^{\mathrm{o}\mathrm{u}\mathrm{t}} =\{\begin{array}{ll}SPA\left(A_{i-1}^{\mathrm{o}\mathrm{u}\mathrm{t}}+A_{i-2}^{\mathrm{o}\mathrm{u}\mathrm{t}}\right),& i\mathrm{i}\mathrm{s}\mathrm{o}\mathrm{d}\mathrm{d},\\ SPA\left(A_{i-1}^{\mathrm{o}\mathrm{u}\mathrm{t}}\right),& i\mathrm{i}\mathrm{s}\mathrm{e}\mathrm{v}\mathrm{e}\mathrm{n},\end{array}

where SPA(·) represents a SPA operation. Generally, a SPA operation is denoted as

(3)A_{\mathrm{a}\mathrm{t}\mathrm{t}\mathrm{n}}=spa(A)\cdot A.

By using spa(A) to generate an attention map, we can then map it back to the original feature map to generate a new feature map that incorporates SPA. Specifically, to achieve SPA, we adopt the deformable convolution [34], the theory of which is to add an offset to the sample location of the original convolution that is determined by the image features. Thus, the process of SPA is

(4)A_{\mathrm{a}\mathrm{t}\mathrm{t}\mathrm{n}}=DCN(A)\cdot A,

where the DCN represents the deformable convolution.

The SPAM enhances feature maps to focus on the spatial locations of foreground objects through SPA. The incorporation of SPA enhances the precision and robustness of object detection. The proposed technique is a valuable tool for promoting the performance of object detection models in complex applications.

3.5 Low-light generation

Poor light levels are common in shield tunnel inspection work. There are many ways to address the low-light problem, mostly by means of light enhancement. As an algorithm that needs to be deployed in the real working environment, processing time of light enhancement can affect detection speed. However, training models to learn image features in such scenes by adding low-light data does not affect the inference speed of the algorithm.

There is a lot of work on low-light enhancement, but few studies on light reduction. Direct exposure adjustment of compressed images (jpeg, jpg, png, etc.) can bring about information loss, because no protection is made for consistency between pixels and channels. To create an effective light attenuation method, we implemented the inverse from satisfactory light enhancement algorithm, i.e., reverting from normal light to low light. Zero-DCE [35] is a good light enhancement model that uses a pixel-wise adjustment curve and takes full account of multiple dimensions to avoid loss of information caused by adjusting the pixel values. The curve is defined as

(5)LE(x)=x+A(x)\cdot x(1- x),

where LE(·) is light enhancement function, x denotes the input pixel value, and A is a parameter map that has the same dimensions as the input image.

The exposure control loss L_exp is derived with this curve,

(6)L_{\text{exp}}=\frac{1}{M}\sum _{k=1}^M\left|Y_k-E\right|,

where M represents the number of nonoverlapping local regions, Y is the average intensity value of a local region in the enhanced image, and E is the well-exposedness level which is set to 0.6 by default. Specifically, to reverse the Zero-DCE operation, we adjust its proposed monotonically increasing LE-Curve to monotonically decreasing. The curve is formulated as

(7)LE\mathrm{\_}r evers e\text{(}x\text{)} =1-x+A\text{(}x\text{)}\cdot x\text{(}x-1\text{)}.

In addition, the exposure control loss parameter E in Zero-DCE is modified as 0.2. We use this inverse operation to train the water leakage image with normal illumination, and then generate the low-light images. Fig.5 demonstrates the pipeline of the generation process, and the produced images are shown in Fig.6. All of the experiments using low-light generation are conducted using the low-light generated images together with the original images for model training.

4 Experiment configuration

4.1 Data set

The open-access data set [9] we use to evaluate the performance is proposed for fully-supervised water leakage inspection. Following the strategy in Ref. [9], we use 3555 images for training, 1000 images for validation, and 100 images for testing, as listed in Tab.1. Also, we add 1200 low-light images which are generated using reverse Zero-DCE as extra training set.

4.2 Performance metrics

We assessed our method using AP as the evaluation metric, which is derived from the area under the precision-recall (PR) curve. The precision and recall scores are first obtained by

(8)\begin{array}{rl}& Precoldid=\frac{TP}{TP+FP},\\ & recall=\frac{TP}{TP+FN},\end{array}

where TP, FP, and FN indicate true positive, false positive, and false negative, respectively. Fig.7 gives an illustration of AP calculation. Precision is a measure of how many items that the algorithm claims to be positive are actually TP. Recall is a measure of how many of the ground truth objects were successfully detected by the algorithm. AP considers both precision and recall when evaluating the performance of an object detection algorithm. It provides a thorough assessment of its effectiveness at various confidence levels; the higher the AP, the better performance of the model.

The AP selects the predicted boxes with the ground-truth intersection over union (IoU) greater than the threshold as a positive sample. In this paper, AP₅₀ refers to the AP at 50% IoU threshold, AP denotes the average of AP₅₀ to AP₉₅. The calculation of IoU is

(9)IoU=\frac{Are{a}_{\mathrm{I}\mathrm{n}\mathrm{t}\mathrm{e}\mathrm{r}\mathrm{s}\mathrm{e}\mathrm{c}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}}}{{Area}_{\mathrm{U}\mathrm{n}\mathrm{i}\mathrm{o}\mathrm{n}}},

where Area_Intersection is the area of the intersection between the predicted and ground truth regions, and Area_Union is the area of their union, as shown in Fig.8.

4.3 Implementation details

The experiments were carried out on a computer equipped with two NVIDIA RTX 3090 GPUs (24 GB memory). The modeling process used Pytorch framework, and the low-light generation methods adopted the source codes provided by Zero-DCE.

For model training of TDD, we utilized the AdamW optimizer [36] with a batch size of 16 and an initial learning rate of 0.0001 for training. Considering that many cracks and leaks have slender shapes, to accommodate crack aspect ratios that are too small or too large, we manually added scales of 0.1 and 10 to the list of base anchors when generating the prior anchors. For performance comparison, some state-of-the-art detectors were selected. Specifically, all of the compared detectors were initialized using pre-trained weights on Microsoft COCO [37]. The experimental results were obtained with the inputs of size 1333 pixels × 800 pixels.

4.4 Data augmentation

In the scenarios of low light condition, employing data augmentation techniques can enhance the model’s performance without an increased computational burden. Coarse labeling of the samples has previously been observed from the available water leakage data set. Many samples have different labeling styles. As indicated in Fig.9, large leaky areas are labeled into small areas or labeled as a whole, which increases the difficulty of model learning. In this direction, a straightforward approach is to perform data augmentation on the ground-truth bounding boxes, and this paper uses BBoxJitter and BBaug [38] to improve the diversity of the training samples.

5 Results

5.1 Detection results

The TDD was evaluated on the data set collected by Xue et al. [9], it achieves an AP_0:5:0:95 score of 56.8%, and an AR_0:5:0:95 (average recall) score of 72.4%. Some detection results are visualized in Fig.10. Specifically in Fig.10(a) we selected samples with noise addition and rotation transformation. Fig.10(b) shows the detection results in normal light conditions in the validation set, and Fig.10(c) demonstrates the detection results on the samples with insufficient illumination. As can be observed, TDD is able to accurately detect water leakage in various conditions.

For deployment in practical work, an optimal confidence threshold for predicting should be found. A good model should have a high detection rate, i.e., recall rate; in most cases, it is better to have a small probability of false detection than to have a missed detection of water leakage spot. According to this requirement, an F2-score indicator is employed, defined as

(10)F\beta =(1+{\beta }^2)\frac{P\times R}{{\beta }^2\times P+R}, \beta = 2,

where R (recall) has a significant impact on the results. The higher the recall rate (detection rate), the higher the F2-score. Therefore, by setting different inference confidence thresholds, we can find the highest F2-score, which is the parameter with the best detection rate. As we can observe from Tab.2, TDD has the highest F2-score when the confidence threshold is set to 0.3. This means the model keeps the detections with a higher confidence than 0.3 as positive samples, which should be selected for real shield tunnel lining inspection work.

Visualization allows for a clear comparison between the feature maps produced by SPAM and those produced without SPAM. We randomly select three samples from the validation sets to visualize feature maps for demonstrating SPAM’s effectiveness. We list the attention heatmap to the feature map and annotate bounding boxes for better visualization in Fig.11, where Fig.11(a) displays the original images, while Fig.11(b) shows the features before SPAM, Fig.11(c) shows the features processed by SPAM. We can clearly observe that the feature background is very clean after SPAM processing, and the attention is well focused on the object range during feature extraction.

5.2 Comparisons

Our goal is to automatically and effectively detect water leakage areas with fewer computational burden. The proposed method adopts a relatively lightweight backbone ConvNeXt-S, but achieves satisfactory performance, which demonstrates the advantages of our method for water leakage detection of shield tunnel lining. Tab.3 lists three kinds of methods that are selected for comparison.

The depth of the backbone significantly determines the performance of the method. Tab.3 indicates that the performances of heavy backbones (ResNeXt-101, ConvNeXt-L, SwinT Large, and Dual-SwinT Base) are higher than those of lighter backbones (SwinT Small, ConvNeXt-S). Our method brings a significant performance improvement with the lightweight ConvNeXt-S, achieving state-of-the-art performance.

We further compare our method with previous work [9] and other advanced detectors, including GFLv2 [39], and YOLOv7 [26]. These detectors have shown satisfactory object detection performance on public data set; however, their performance is inconsistent in water leak detection due to the data difference.

5.3 Ablation experiments

1) D-FPN. The contribution of the D-FPN is demonstrated in Tab.4. We can observe that detection neck plays an important role in object detection. Without neck as the feature aggregator, the performance drops significantly. By adding the 4-layer FPN as the default neck, the performance is improved by 13.3%. A further improvement of 0.2% was achieved by replacing it with the 3-layer D-FPN as is done in our proposal.

2) SPAM. We also evaluate the SPAM’s contribution, and the results are listed in Tab.5. As can be observed, SPAM brings a significant performance improvement, and the performance drops by 6.1% if our method is implemented without SPAM. Contribution of the SPAM is also indicated in Fig.11, and has been discussed in Subsection 5.1.

3) Low-light generation. Regarding low-light generation, we first test the detection performance on low-light images without using generated low-light images, the results are listed in Tab.6. Evidently, the addition of generated low-light images improves the model’s performance in both normal light and low light conditions, suggesting that the proposed low-light generation technique effectively enhances the model’s robustness to detect water leakage areas in low-light conditions.

6 Discussion

6.1 Spatial attention module

The SPAM is developed as an intermediate module between the neck and head, with the purpose of improving the feature quality before prediction. It is pluggable, with SPA that can well attenuate the background noise features and focus attention on the object. Tab.7 shows results of plugging the same SPAM into different models to evaluate its compatibility. As can be observed, the performance improvement of SPAM is considerable, especially in small models with less feature extraction capability.

The SPAM is lightweight, and is formed by an odd number of SPA block. We attempt to find an optimal result by stacking different counts of SPA blocks. We intuitively illustrate, in Fig.12 and listed in Tab.8, the experimental results of stacking different numbers of blocks, which clearly show that the performance is optimal when stacking up to 5 or 7 SPA blocks, increasing the total computational effort by only 11.7%.

6.2 Strategy and methods

In this paper, we propose a fully-supervised water leakage detection method. In practical engineering scenarios, labeled image data of leakage areas or other tunnel defects are not readily available, thus there is a high probability of data inadequacy. Insufficient data will lead to insufficient training of the fully-supervised model, so that it has unsatisfactory detection precision in practical applications. In semi-supervised deep-learning tasks [41–43], the model is trained on a smaller set of labeled data along with a larger set of unlabeled data, these learning strategies can improve the model ability to recognize and localize objects accurately.

This paper proposes an object detection-based water leakage and other wall defect inspection method, which is a bounding box level approach. In most cases, object detection methods have advantages in real-time; however, some recent works [44,45] also achieve real-time segmentation. For specific tasks that require quantification analysis, object detector has insufficient capabilities. In potential future works, segmentation-based methods are worth studying.

6.3 Imaging condition

The employed data set does not contain images taken from a distant position, thus there is no significant multiscale problem. In practical applications, the camera could be placed in any position, and thus causes different object size distributions. To solve this problem, data augmentation is an efficient way to avoid precision decrease. For example, multiscale training and testing can be used to improve the model’s ability to detect objects of different sizes, thus simulating objects that are at different distances. Mosaic [46] also makes the object samples have different size distributions by piecing together training images.

7 Conclusions

Water leakage detection of shield tunnel is a crucial task in ensuring the safety and reliability of structures such as tunnel lining. In this study, we introduce a novel approach for detecting water leakage by utilizing deep-learning based object detection techniques. Our method obtains an AP score of 56.8%, which outperforms similar work by 5.7%. The combined use of SPAM and D-FPN dramatically improves the performance of the lightweight models, outperforming models using large backbones at only 1/3 of the computational cost. Meanwhile, the model is further adapted to the low-light environment, and its robustness is enhanced by simulating the light problems that exist in inspection and maintenance work.

References

[1] Li C, Wang M, Gao X, Yang B. Research on on-site monitoring and measurement technology of tunnel. IOP Conference Series: Earth and Environmental Science, 2020, 455(1): 012154

[2] Ma D, Fang H, Wang N, Zhang C, Dong J, Hu H. Automatic detection and counting system for pavement cracks based on PCGAN and YOLO-MF. IEEE Transactions on Intelligent Transportation Systems, 2022, 23(11): 22166–22178

[3] Zhou Q, Qu Z, Li Y X, Ju F R. Tunnel crack detection with linear seam based on mixed attention and multiscale feature fusion. IEEE Transactions on Instrumentation and Measurement, 2022, 71: 1–11

[4] Meng S, Kuang S, Ma Z, Wu Y. MtlrNet: An effective deep multitask learning architecture for rail crack detection. IEEE Transactions on Instrumentation and Measurement, 2022, 71: 1–10

[5] Ali R, Chuah J H, Talip M S A, Mokhtar N, Shoaib M A. Structural crack detection using deep convolutional neural networks. Automation in Construction, 2022, 133: 103989

[6] Chen J, Sun S, Zhang L B, Yang B, Wang W. Compressed sensing framework for heart sound acquisition in internet of medical things. IEEE Transactions on Industrial Informatics, 2022, 18(3): 2000–2009

[7] WangWYu XFangBZhaoD YChenY WeiWChenJ. Cross-modality LGE-CMR segmentation using image-to-image translation based data augmentation. IEEE/ACM Transactions on Computational Biology and Bioinformatics, 2022

[8] ChenJGuo ZXuXZhangL BTengY ChenYWoniak MWangW. A robust deep learning framework based on spectrograms for heart sound classification. IEEE/ACM Transactions on Computational Biology and Bioinformatics, 2023

[9] Xue Y, Cai X, Shadabfar M, Shao H, Zhang S. Deep learning based automatic recognition of water leakage area in shield tunnel lining. Tunnelling and Underground Space Technology, 2020, 104: 103524

[10] DingHLiu SCaiSXiaY. Big data analysis of structural defects and traffic accidents in existing highway tunnels. In: Proceedings of the 3rd International Conference on Information Technology in Geo-Engineering. Cham: Springer Cham, 2019, 189–195

[11] LiCChenW DengRHan Q. Overview of tunnel detection technology. In: Proceedings of International Conference on Maintenance Engineering. Cham: Springer Cham, 2021, 81–90

[12] Kawata H, Matsumoto Y, Hachiga A, Okuaki S. Development of expert system (TIMES-1) for tunnel inspection and diagnosis. Railway Technical Research Institute, Quarterly Reports, 1989, 30(3): 143–148

[13] Yu S N, Jang J H, Han C S. Auto inspection system using a mobile robot for detecting concrete cracks in a tunnel. Automation in Construction, 2007, 16(3): 255–261

[14] HuCZhuH LiX. Detection of tunnel water leakage based on image processing. Information Technology in Geo-Engineering, 2010, 254–262

[15] Yokoyama S, Matsumoto T. Development of an automatic detector of cracks in concrete using machine learning. Procedia Engineering, 2017, 171: 1250–1255

[16] Protopapadakis E, Voulodimos A, Doulamis A, Doulamis N, Stathaki T. Automatic crack detection for tunnel inspection using deep learning and heuristic image post-processing. Applied Intelligence, 2019, 49(7): 2793–2806

[17] Lewis J, Cha Y J, Kim J. Dual encoder–decoder-based deep polyp segmentation network for colonoscopy images. Scientific Reports, 2023, 13(1): 1183

[18] GirshickR. Fast R-CNN. In: Proceedings of the IEEE International Conference on Computer Vision. New York: IEEE, 2015, 1440–1448

[19] Ren S, He K, Girshick R, Sun J. Faster R-CNN: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017, 39(6): 1137–1149

[20] CaiZVasconcelos N. Cascade R-CNN: Delving into high quality object detection. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. New York: IEEE, 2018, 6154–6162

[21] QiaoSChen L CYuilleA. Detectors: Detecting objects with recursive feature pyramid and switchable atrous convolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE, 2021, 10213–10224

[22] LiuWAnguelovD ErhanDSzegedy CReedSFuC YBergA C. SSD: Single shot multibox detector. In: Proceedings of Computer Vision–ECCV 2016: 14th European Conference. Cham: Springer Cham, 2016, 21–37

[23] RedmonJDivvala SGirshickRFarhadiA. You only look once: Unified, real-time object detection. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. New York: IEEE, 2016, 779–788

[24] RedmonJFarhadi A. Yolov3: An incremental improvement. 2018, arXiv: 1804.02767

[25] GeZLiuS WangFLi ZSunJ. Yolox: Exceeding Yolo series in 2021. 2021, arXiv: 2107.08430

[26] WangC YBochkovskiy ALiaoH Y M. YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. 2022, arXiv: 2207.02696

[27] LinT YGoyal PGirshickRHeKDollár P. Focal loss for dense object detection. In: Proceedings of the IEEE International Conference on Computer Vision. Proceedings of the IEEE International Conference on Computer Vision, 2017, 2980–2988

[28] Cha Y J, Choi W, Suh G, Mahmoudkhani S, Büyüköztürk O. Autonomous structural visual inspection using region-based deep learning for detecting multiple damage types. Computer-Aided Civil and Infrastructure Engineering, 2018, 33(9): 731–747

[29] Xia B, Cao J, Zhang X, Peng Y. Automatic concrete sleeper crack detection using a one-stage detector. International Journal of Intelligent Robotics and Applications, 2020, 4(3): 319–327

[30] Xue Y, Shi P, Jia F, Huang H. 3D reconstruction and automatic leakage defect quantification of metro tunnel based on SfM-deep learning method. Underground Space, 2022, 7(3): 311–323

[31] Guo J, Markoni H, Lee J D. BARNet: Boundary aware refinement network for crack detection. IEEE Transactions on Intelligent Transportation Systems, 2022, 23(7): 7343–7358

[32] LiuZMaoH WuC YFeichtenhofer CDarrellTXieS. A convnet for the 2020s. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE, 2022, 11976–11986

[33] LinT YDollár PGirshickRHeKHariharan BBelongieS. Feature pyramid networks for object detection. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. New York: IEEE, 2017, 2117–2125

[34] DaiJQiH XiongYLi YZhangGHuHWeiY. Deformable convolutional networks. In: Proceedings of the IEEE International Conference on Computer Vision. New York: IEEE, 2017, 764–773

[35] GuoCLiC GuoJLoyC C HouJKwongS CongR. Zero-reference deep curve estimation for low-light image enhancement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE, 2020, 1780–1789

[36] LoshchilovIHutter F. Decoupled weight decay regularization. 2017, arXiv: 1711.05101

[37] LinT YMaire MBelongieSHaysJPeronaP RamananDDollár PZitnickC L. Microsoft COCO: Common objects in context. In: Proceedings of the European Conference on Computer Vision. Cham: Springer Cham, 2014, 740–755

[38] ZophBCubuk E DGhiasiGLinT YShlensJ LeQ V. Learning data augmentation strategies for object detection. In: Proceedings of Computer Vision–ECCV 2020: 16th European Conference. Cham: Springer Cham, 2020, 566–583

[39] LiXWangW HuXLiJ TangJYang J. Generalized focal loss v2: Learning reliable localization quality estimation for dense object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE, 2021, 11632–11641

[40] Liang T, Chu X, Liu Y, Wang Y, Tang Z, Chu W, Chen J, Ling H. CBNet: A composite backbone network architecture for object detection. IEEE Transactions on Image Processing, 2022, 31: 6893–6906

[41] DoulamisNDoulamis A. Semi-supervised deep learning for object tracking and classification. In: Proceedings of the 2014 IEEE International Conference on Image Processing (ICIP). New York: IEEE, 2014, 848–852

[42] BaurCAlbarqouni SNavabN. Semi-supervised deep learning for fully convolutional networks. In: Proceedings of the Medical Image Computing and Computer Assisted Intervention MICCAI 2017: 20th International Conference. Cham: Springer Cham, 2017, 311–319

[43] ChenYZhu XGongS. Semi-supervised deep learning with memory. In: Proceedings of the European Conference on Computer Vision (ECCV). Cham: Springer Cham, 2018, 268–283

[44] Choi W, Cha Y J. SDDNet: Real-time crack segmentation. IEEE Transactions on Industrial Electronics, 2020, 67(9): 8016–8025

[45] Kang D H, Cha Y J. Efficient attention-based deep encoder and decoder for automatic crack segmentation. Structural Health Monitoring, 2022, 21(5): 2190–2205

[46] BochkovskiyAWangC YLiaoH Y M. Yolov4: Optimal speed and accuracy of object detection. 2020, arXiv: 2004.10934

RIGHTS & PERMISSIONS

Higher Education Press