A Scoping Review: Applications of Deep Learning in Non-Destructive Building Tests

Abstract

Background: In the context of rapid urbanization, the need for building safety and durability assessment is becoming increasingly prominent. Objective: The aim of this paper is to review the strengths and weaknesses of the main non-destructive testing (NDT) techniques in construction engineering, with a focus on the application of deep learning in image-based NDT. Design: We surveyed more than 80 papers published within the last decade to assess the role of deep learning techniques combined with NDT in automated inspection in construction. Results: Deep learning significantly enhances defect detection accuracy and efficiency in construction NDT, particularly in image-based techniques such as infrared thermography, ground-penetrating radar, and ultrasonic inspection. Multi-technology fusion and data integration effectively address the limitations of single methods. However, challenges remain, including data complexity, resolution limitations, and insufficient sample sizes in NDT images, which hinder deep learning model training and optimization. Conclusions: This paper not only summarizes the existing research results, but also discusses the future optimization direction of the target detection network for NDT defect data, aiming to promote intelligent development in the field of non-destructive testing of buildings, and to provide more efficient and accurate solutions for building maintenance.

1. Introduction

In modern urbanization, concrete materials have become the main construction material for most buildings due to their good mechanical properties and cost-effectiveness [1]. However, during long-term service, buildings inevitably face multiple challenges such as material aging, environmental erosion, and load accumulation, leading to various forms of deterioration of concrete materials, such as cracks, delamination, spalling, voids, leakage, and corrosion of reinforcement [2]. These deterioration phenomena not only seriously affect the structural integrity and safety, but also pose a great challenge to the regular inspection and maintenance of buildings.

Regular inspection and maintenance are an essential step to ensure the safety and longevity of a building and to detect potential problems. However, in practice, we are often faced with a lack of systematic maintenance inspections, reliance on temporary repairs for external damage, and missing historical inspection reports [3]. Traditional wall defect detection methods, such as the manual visual inspection method and knocking method, although simple to operate, are limited by low efficiency, poor accuracy, and strong subjectivity, and make it difficult to meet the urgent needs of modern building maintenance. As a result, advanced NDT techniques are leading a new trend, evolving from mere defect identification to a comprehensive assessment of the nature, number, location, and size of defects, as well as the impact of defects on the performance of the component under test [4]. Although NDT technology has made significant progress, the final damage detection results still require professional judgment, making it difficult to achieve rapid detection.

In this context, the great potential of deep learning in automatic image analysis and computer vision tasks provides new ideas for building condition inspection. Through deep learning technology, we can achieve building condition detection more efficiently and accurately, injecting new vitality into the field of building maintenance. Deep learning models can be classified according to the number of stages in the detection process, whether they rely on an anchor box, and a labeling scheme, such as single-stage detection models (e.g., YOLO and SSD), two-stage detection models (e.g., Faster R-CNN and Mask R-CNN), anchor box-based models (e.g., Faster R-CNN), anchor box-free models (e.g., CornerNet), region suggestion-based models such as Faster R-CNN, and key point-based models (e.g., CenterNet). These classification methods provide a rich applicational perspective of deep learning in the field of building detection. Therefore, in this paper, we delve into the application of deep learning in non-destructive testing of building conditions, to provide smarter and more efficient solutions for regular building maintenance through a literature review.

This paper reviews the application progress of deep learning in the field of building non-destructive testing (NDT). Section 2 describes the methods used in the thesis. Section 3 discusses the combination of image-based NDT technology and deep learning, such as infrared thermal imaging, ground-penetrating radar, and ultrasonic detection, and integrates the application of multiple inspection methods to improve inspection accuracy. In Section 4, future optimization directions for NDT images are presented, aiming to address the issues of data complexity and scarcity. Finally, in Section 5, we summarize the application progress and development trend of deep learning in NDT.

2. Methods

The reporting process adhered to the PRISMA-ScR checklist, which is provided in the Supplementary Materials.

2.1. Review Question

Our research revolves around the following key questions: What are the strengths and limitations of NDT techniques in construction? What is the current status of deep learning applications in NDT, and what are the main challenges, future trends, and optimization directions?

2.2. Eligibility Criteria

This study includes publications from the last decade, with a focus on the last five years of the literature to reflect the latest developments in NDT and deep learning in the construction field. The most recent search was conducted on 18 February 2025.

2.3. Exclusion Criteria

When writing our review article, the exclusion criteria include the following: the literature that is not relevant to the topic of NDT and deep learning research areas; the literature published more than a decade ago, unless it is considered classic literature or grounded theory; the literature from non-academic journals, conference proceedings, or non-peer-reviewed preprints such as arXiv; and the literature that lacks full-text access or duplicates content.

2.4. Search Strategy

We systematically searched four databases, namely, Web of Science, IEEE Xplore, ASCE, and CNKI, using keywords such as “deep learning”, “nondestructive testing”, “building inspection”, “concrete deterioration”, and “image processing”. Initially, we obtained 547 articles. After eliminating duplicates, early publications, and the irrelevant literature, the remaining literature was screened and evaluated in full text. The studies were finally determined to be included in the systematic review. The flow chart of the PRISMA method is shown in Figure 1.

2.5. Data Extraction and Data Synthesis

In the data extraction stage, we collected key information from the literature, including publication year, authors, research objectives, research methodology, main findings, and impacts. We also recorded the types of deep learning models, input/output data, and their main results. In the data synthesis stage, a thematic synthesis approach was used to organize the data into categories such as infrared thermography, ground-penetrating radar, ultrasonic detection technology, synthesis of multiple NDT methods, and deep learning-assisted inspection.

3. Discussion

3.1. Deep Learning-Assisted Infrared Thermography Technology

Infrared thermography (IRT) is a non-contact detection method; it can quickly scan a large area, capture the thermal radiation of the object [5], and convert it into a visualized thermal image to achieve rapid detection of surface and internal material defects, including heat loss defects, water infiltration of building envelopes, as well as internal defects such as voids, cracks, delamination, and so on [6,7,8,9]. The device is shown in Figure 2a [10]. The technology is divided into two categories, passive and active, according to the heat source: passive IRT naturally captures the thermal radiation emitted by the object itself, and is suitable for surface defect detection and thermal efficiency assessment, while active IRT excites deep thermal waves with the help of an external heat source and is specialized in the detection of internal defects. At the application level, IRT technology can be divided into qualitative and quantitative modes. Qualitative IRT focuses on intuitive temperature distribution analysis and the interpretation of pixel values in thermal images, and is often used for the detection of defective areas; quantitative IRT delves deeper into the subtle changes in temperature values and material properties to achieve the quantitative assessment [11].

However, the accuracy and efficiency of infrared thermal imaging technology are susceptible to interference from environmental factors such as atmospheric conditions and solar radiation. To address this challenge, the aerial infrared thermal imaging (AIR_T) method [12] is proposed, which combines the UAV detection technology with the mobility of data acquisition equipment, and solves the monitoring problem in complex or inaccessible locations (such as under bridges and behind high-rise buildings), as shown in Figure 2b [13]. In addition, by optimizing the flight plan and data acquisition strategy, flexibly adjusting the camera angle, flight distance, and trajectory, fast and efficient data acquisition is realized, and the adverse impact of environmental factors on data acquisition is effectively reduced. Ma et al. [14] analyzed the impact of camera angle, distance, and illumination on the detection of the fall of tiles outside the building, and provided suggestions on the flight area and time of the UAV, as shown in Figure 2c, thereby improving the data quality. In the field of thermal image processing, Deng et al. [15] fused IRT and Mask R-CNN algorithms for an intelligent inspection system to achieve automatic classification of defects in internal concrete structures, as illustrated in Figure 3. Cheng et al. [16] proposed a deep learning model based on an encoder–decoder architecture to achieve pixel-level segmentation of layered regions in thermal images. Ali et al. [10] demonstrated the effectiveness of the YOLOv7 algorithm in improving building automation potential and the accuracy of building facade heat loss detection. Currently, the fusion of IRT and deep learning algorithms mostly focuses on qualitative analysis, and there have also been studies proposing a lightweight internal damage segmentation network (IDSNet) for real-time region segmentation of thermal images of concrete slabs to quantify internal damages such as delamination, cracks, voids, and honeycombing [17].

Figure 2. (a) Handheld infrared thermography equipment [10]; (b) UAV thermal imaging [13]; and (c) UAV flight area proposal map [14]. Reproduced with permission from Ref. [10]. Copyright 2024 Elsevier. Reproduced with permission from Ref. [13]. Copyright 2023 Elsevier. Reproduced with permission from Ref. [14]. Copyright 2024 Elsevier.

Figure 2. (a) Handheld infrared thermography equipment [10]; (b) UAV thermal imaging [13]; and (c) UAV flight area proposal map [14]. Reproduced with permission from Ref. [10]. Copyright 2024 Elsevier. Reproduced with permission from Ref. [13]. Copyright 2023 Elsevier. Reproduced with permission from Ref. [14]. Copyright 2024 Elsevier.

Figure 3. Schematic diagram of IRT combined with Mask R-CNN algorithm for detecting internal defects in concrete structures [15]. Reproduced with permission from Ref. [15]. Copyright 2023 Elsevier.

Figure 3. Schematic diagram of IRT combined with Mask R-CNN algorithm for detecting internal defects in concrete structures [15]. Reproduced with permission from Ref. [15]. Copyright 2023 Elsevier.

3.2. Deep Learning-Assisted Ground-Penetrating Radar Technology

Ground-penetrating Radar (GPR) technology has demonstrated its unique application advantages in many key engineering fields, such as pavement defect detection, bridge quality assessment, railway monitoring, and tunnel lining inspection [18,19,20]. The technology generates detailed images of underground structures through the propagation and reflection of high-frequency electromagnetic waves in the medium, providing intuitive information about the structure’s interior. In the GPR image, the scattering hyperbolic vertices become the key indicators to accurately mark the location of the target object, and this feature greatly enhances the ability of engineers to judge the location of various defects or objects inside the structure. In the inspection of concrete structures, GPR not only accurately locates reinforcement bars and evaluates corrosion conditions [21,22], but also effectively reveals cracks, leaks, grouting defects, voids and other internal defects [23,24], while assisting in the measurement of material delamination and the thickness of the protective layer. The basic principle and images of GPR inspection are shown in Figure 4 [22,25]. To meet various depths of detection requirements, high-frequency antennas, with their high-resolution advantage, become the preferred choice for detecting subtle objects (such as defects and reinforcement bars) in concrete, road, and bridge surface structures. Although low-frequency antennas have slightly lower resolution, they can penetrate deeper into the ground and provide valuable information for deep structure detection [26].

With the advancement of image processing technology, researchers have been exploring ways to improve the accuracy of GPR data analysis. Dinh et al. [27] effectively reduced false anomalies by combining traditional processing techniques, such as image migration, with CNN automation algorithms, but the algorithms require a large amount of training data and do not apply to all types of bridge coverages. Subsequently, Liu et al. [28] introduced migration learning to optimize SSD model training, which alleviated the dependence on a large amount of training data. However, in terms of quantitative rebar size detection, it is difficult to achieve the required accuracy standard by directly assessing rebar size from GPR images. For this reason, Chang et al. [29] achieved quantitative detection of rebar radius within 7% error by constructing an electromagnetic physical model considering rebar radius. Many factors can affect the detection results of ground-penetrating radar (GPR). For example, the thickness of the overlay, the moisture content in the structure, and the spacing of the rebar all have an impact on the GPR signal. It is worth noting that the strong reflectivity of metallic materials leads to a shielding effect of the reinforcement layer in the concrete on the GPR signals, masking the echoes of the damage under the reinforcement and making it difficult to accurately identify the damage signals. This shielding also complicates the detection of steel pipe grouting compactness, making it challenging to assess the integrity of such structures using GPR alone. Focusing on the interference problem of steel reinforcement in defect detection, Wang et al. [30] designed a deep neural network integrating a spatial attention module and a residual starting block, as shown in Figure 5a, aiming to strengthen the network’s ability to identify weak defect signals and achieve an accurate reconstruction of multiple defect reflections. With this design, the rebar signals in the GPR image were effectively filtered out and the echoes of the defects under the rebar layer were enhanced, which improved the recognition accuracy from 0.208 to 0.850. After that, they proposed an unsupervised rebar clutter suppression model based on CycleGAN by combining the contrast feature encoder and similarity feature encoder to address the problem of fewer images in the actual project, as shown in Figure 5b, which effectively reduces the misdetection and omission of the defect [31].

3.3. Deep Learning-Assisted Ultrasonic Detection Technology

Ultrasonic inspection methods are widely used in thickness measurement, delamination detection, and the detection of concrete members, tunnel linings, under-pavement voids, and reinforcing steel by the analysis of ultrasonic propagation characteristics in the medium, such as reflection, refraction, and scattering phenomena [32,33,34]. Combined with advanced imaging techniques such as synthetic aperture focusing technology (SAFT) (shown in Figure 6 [35,36]), ultrasonic transverse wave tomography (USWT) [33], and the total focusing method (TFM) [37], ultrasonic inspections can acquire high-resolution image datasets, which greatly facilitates the accurate identification of defects. During ultrasonography, the adjustment of parameters affects the accuracy of imaging, among which the physical parameter of wave velocity is crucial for the depth accuracy of internal defect localization [36]. Figure 7 [38] shows the imaging effects at different wave speeds.

Although the combination of ultrasound imaging and deep learning models has been less studied for the automatic identification of internal concrete defects compared to thermography and GPR images, significant results have been achieved. The first attempt to process ultrasound images using Convolutional Neural Networks (CNNs) was made by Słoński et al. [39] to achieve the classification of a single type of defects in a simulated PVC pipe using a pre-trained VGG-16 model. Kuchipudi et al. [40] further advanced by optimizing defect detection and segmentation of ultrasonic images using a region-based CNN model combined with the RolAlign technique, identifying a wide range of defects in concrete including rebar debonding, delamination, and cracking, with an optimal mean accuracy (mAP) of up to 0.98. Yang et al. [41] utilized a deep neural network with an encoder-decoder architecture, combined with a feature fusion strategy and residual module optimization to achieve pixel-level non-destructive detection of cracks as low as 1 mm width inside reinforced concrete structures, and globally display the crack location and distribution through the alignment technique, with a crack length quantification error of only 6.22%. In addition, the study by Wu et al. [38] focuses on significantly improving image resolution with enhanced recognition of small target defects, integrating the variational mode decomposition and bilateral filtering techniques and introducing the improved YOLOv5s-CBAM network model, as illustrated in Figure 8a, which has been tested in the laboratory with good results. However, the model does not have high detection accuracy for tunnel lining scenes.

Ultrasonic testing still has many limitations in the identification of internal defects of concrete structures. First of all, the detection equipment has high requirements for surface smoothness and is subject to the adjustment of equipment parameters and the resolution of imaging technology. Secondly, there is insufficient research on the automatic detection of various defect characteristics in concrete structures by imaging. Currently, most of the related research work mainly focuses on simulated detection under laboratory conditions, while relatively few studies are conducted in real application scenarios, resulting in limited model generalization capability and a lack of necessary field dataset support for training optimization. Although existing studies have made some progress in simulated laboratory environments, how to effectively translate these results into models and data suitable for real-world applications is still an urgent problem to be solved.

3.4. Integration of Multiple Inspection Methods

The diversity and complexity of NDT techniques are reflected in their inability to single-handedly address the challenges of all types of defects, with each technique possessing unique strengths and limitations. It is these characteristics that have led to the integration of multi-technology systems as an important strategy to ensure the success and accuracy of inspection results.

In the practice of integrated application, different NDT techniques have their own strengths due to their differences in the depth of detection. The study by Lagüela et al. [42] achieved comprehensive localization from road surface geometry to internal lesions through the combined application of TLS, GPR, and IRT. Similarly, the joint application of IRT and GPR in building moisture detection provides a reliable basis for moisture tracking and depth assessment through the dual analysis of surface moisture movement and internal water accumulation points [43]. In addition, the integrated utilization of data between different NDT techniques enables mutual validation and complementarity, with IRT helping to confirm whether scattering and attenuation in GPR signals are caused by delamination or moisture, thus reducing misjudgments. Meanwhile, the depth information of GPR provides three-dimensional spatial localization of temperature anomalies for IRT, enabling more accurate identification and assessment of internal corrosion and damage areas [44]. For different inspection objects, each technology also shows unique value. In concrete masonry wall monitoring, for example, non-contact IRT, passive acoustic emission AE, and active ultrasonic UT technologies work in concert to comprehensively capture structural damage information through a cross-validation mechanism [45]. In the detection of defects in steel pipe concrete, in response to the possible bias of the ultrasonic method in the detection of dehollowing defects, Chen et al. [46] proposed an improved technique combined with IRT, which was able to more accurately identify dehollowing and internal defects in steel pipe concrete columns. This was further demonstrated in the study by Gucunski et al. [47], who combined five NDE techniques (e.g., impingement echo, ground-penetrating radar, etc.) for the long-term monitoring of bridge decks and not only confirmed corrosion as the main cause of deck deterioration but also demonstrated the high reproducibility and inter-correlation of these techniques.

With the continuous enrichment of NDT tools and methods, more possibilities for multi-technology system integration have also been opened up. Building information modeling (BIM) models are a key tool in the construction industry, detailing the geometry, material properties, internal spatial layout, and semantic information of a building [48]. Integration of BIM models with sensing devices and drone technology is driving the digital transformation of the construction industry. Data collected by drones are used to identify building defects and generate accurate 3D models. At the same time, the BIM model provides critical information to the drone to assist in planning flight paths [49] and creates a two-way interaction. Unlike static representations of building information modeling (BIM), digital twins (DT) enable real-time data interaction between virtual models (built with modeling, simulation, computation, and analysis tools) and actual physical objects, such as buildings, by integrating technologies such as laser scanning, sensors, and the Internet of Things [50]. This need to visualize project data in real time has driven the widespread use of DT in the construction industry [51,52]. In the operation and maintenance phase of the building lifecycle, Yiğit et al. [53] used drone photogrammetry and structure from motion (SfM) algorithms to create high-precision 3D digital twins, and combined deep learning techniques to automatically detect cracks in cultural heritage buildings. To address the high computing resource and technological requirements for large-scale data processing and real-time processing, Pantoja-Rosero et al. [54] proposed an end-to-end automated process for generating lightweight damage-enhanced digital twins (DADT). The method uses multi-view images and artificial intelligence technology to map damage information into 3D models, significantly improving the efficiency of real-time modeling and data processing, making it more suitable for real-time applications and large-scale deployments. However, the current lack of unified modeling and implementation standards has undoubtedly increased the complexity of data integration, technology implementation, and application [55]. Despite significant advances in hybrid inspection systems, a combination of cost, environmental suitability, and technological complexity needs to be considered in the future to expand the application of NDT technologies [2].

4. Future Trends and Possible Research Directions

4.1. Deep Learning Background

Deep learning (DL), an important branch of artificial intelligence, has revolutionized the field of image processing since AlexNet’s victory in the ImageNet Large-Scale Visual Recognition Challenge (ILSVRC) in 2012 [56]. Deep learning creates a hierarchical representation of input information (e.g., images) through a multilayer network that is automatically adapted during training to learn and extract discriminative features for efficient and accurate image object detection and localization [57]. In contrast, traditional image processing relies on manually extracted features and machine learning techniques, which not only require domain expertise but also struggle to cope with the high variability and complexity inherent in visual data [58]. While these methods perform reasonably well in specific tasks, they are limited in accuracy and cumbersome, while deep learning-based target detection algorithms (e.g., defect recognition) have demonstrated much higher efficiency, driving the field of image processing.

Innovations in network architecture are crucial to cope with requirements such as complex datasets. Convolutional Neural Networks (CNNs), since their introduction by LeCun, have demonstrated a wide range of applications in fields such as non-destructive inspection, as a nonlinear mapping model with supervised learning [27]. With the increase in network depth, residual network (ResNet) [59] effectively mitigates the gradient vanishing problem by introducing residual connections, supports the training of deeper networks, and significantly improves the model performance. In addition, the densely connected convolutional network (DenseNet) [60] further enhances performance through cross-layer feature reuse. Currently, deep learning target detection models are mainly classified into two categories: candidate region methods (e.g., R-CNN, Fast R-CNN, and Faster R-CNN) classify and extract features by pre-selecting potential target regions and combining them with CNNs; regression methods (e.g., SSD and YOLO series) use a single end-to-end CNN to directly predict target bounding boxes and categories, which significantly improves the detection speed. The YOLO family, as the mainstream of single-stage target detection algorithms, has been widely used in industry with continuous evolution from YOLOv1 to YOLOv11. In order to operate efficiently on resource-constrained devices, lightweight network architectures (e.g., MobileNet [61,62,63] and ShuffleNet [64,65]) significantly reduce the computational and parametric quantities of the models by introducing techniques such as depthwise separable convolution and group convolution, enabling deep learning techniques to be applied in more practical scenarios. In addition, extensions such as the attention mechanism provide new solutions for deep learning performance optimization. The attention mechanism directs the model to focus on important features by assigning different weights to different information, thus optimizing feature extraction and representation and improving model performance [66]. This mechanism covers a variety of types such as channel attention, spatial attention, self-attention, and combinations of the above attention types to improve the accuracy of the model in handling complex tasks [67].

In the field of non-destructive testing, we rely on more complex data types, such as thermograms, radargrams, and ultrasound images. These images reveal key information such as the temperature distribution, internal structure, and material properties of an object, but they are usually more complex than RGB images, often accompanied by noise and artifacts, and limited in resolution, coupled with a small number of samples that require precise annotation and pre-processing by professionals, all of which increase the difficulty of processing by deep learning models. In addition, due to the different detection objects and purposes, it is usually necessary to select a specific sensor, and even the same sensor needs to adjust parameters (such as dielectric constant and frequency) for different detection objects. These factors increase the difficulty of deep learning model processing. To further advance the application of deep learning in the field of building inspection, data from multiple sensors need to be integrated to improve the integrity and accuracy of inspections and at the same time, improve the data standardization and sharing platform, and promote the accumulation of high-quality data sets to adapt to the resource-constrained detection environment.

4.2. Data Enhancement

In order to overcome the difficulties of data processing in the field of NDT, brightness and contrast adjustment, perspective transformation, flipping, cropping, scaling, and panning are used as the basic means of supervised data enhancement [68], as shown in Figure 9a [69]. The basic image processing methods can improve the generalization ability and accuracy of the model in the target detection task to a certain extent, but there is a limitation in the expansion of the data diversity. Because of this, YOLO-v5 introduces mosaic and mix-up data enhancement techniques, as shown in Figure 9b,c [69], which effectively enrich the background environment of defect detection and enhance the model’s ability to deal with complex scenes (e.g., small, occluded, and blurred targets) [70].

Meanwhile, the rise of Generative Adversarial Networks (GANs) has injected new power into the field of data enhancement. GANs are capable of generating realistic synthetic data through adversarial training of generators and discriminators. Among them, Deep Convolutional Generative Adversarial Network (DCGAN), as a variant of GAN, combines the variational auto coder (VAE) and Adam optimizer, which is successfully applied to generate realistic asphalt pavement crack images, effectively alleviating the constraints of sample scarcity on model training [71]. To deal with the collapse problem of DCGAN and the limitation of convolutional GAN in generating complex geometric patterns, Liu et al. [72] designed the self-attention mechanism of Self-Attention Generative Adversarial Network-W (SAGAN-W), which generated GPR images with clear hyperbolic curves and could simulate the true backscattering characteristics of defects. In addition, Yang et al. [35] developed a Conditional Generative Adversarial Network (CGAN) variant for generating ultrasound B-scan images that was trained using real B-scan images and their corresponding truthful image data, thus improving the generalization ability of the model.

In the development of image super-resolution techniques, the emergence of Super-Resolution Generative Adversarial Network (SRGAN) significantly improves the quality of image reconstruction, which not only improves the resolution of the image but also preserves the detailed features of the image, such as edges and textures [73]. Aiming at the challenges of low resolution and weak features in infrared images, Regional Super-Resolution Generative Adversarial Network (RSRGAN) effectively improves the recognition rate of small infrared targets by fusing the regional context network with the improved SRGAN, while reducing the computational cost [74]. The Lightweight Thermal Image Super-Resolution (LTSR) model, on the other hand, further improves detection efficiency through multi-scale feature extraction and latent representation learning [75]. Pozzer et al. [76] employed Very Deep Super-Resolution (VDSR) technology in conjunction with the histogram equalization (HE) method to process fusion images, original visible light images, and thermal images. This combination effectively enhanced image resolution and contrast, enabling more accurate capture of details such as exposed rebar, peeling areas, and boundaries. For more accurate capture of details such as exposed rebar, peeling areas, and boundaries in GPR image processing, the unsupervised Generative Adversarial Network model based on CycleGAN achieves automatic conversion of cluttered GPR images to clutter-free images without relying on pairs of training data, which solves the problem of data set scarcity [77]. Finally, in ultrasound image synthesis, an innovative GAN framework successfully synthesizes high-resolution ultrasound images through auxiliary sketch guidance, progressive training strategy, and feature loss function, providing richer and more accurate image resources for non-destructive testing [78].

4.3. Model Adaptation

4.3.1. Transfer Learning

Transfer learning has shown significant advantages in addressing the challenges of insufficient data and changing scenarios, especially in the field of object detection. By effectively utilizing models and knowledge accumulated in related fields or previous tasks, it can greatly reduce the workload of data collection and labeling, avoid model overfitting problems caused by small sample data, and significantly improve the performance and robustness of target detection [79]. Figure 10 illustrates the difference between traditional machine learning and TL techniques for multitasking [80].

In practice, commonly used pre-trained CNN models, such as AlexNet, U-Net, VGG-16, etc., which have been trained on large-scale datasets, can be applied to new problems through migration learning [81]. Fine-tuning is a common method of migration learning, which applies to cases where the source domain is similar to the target domain. For example, using SSD models pre-trained on the PASCAL VOC2007 dataset and fine-tuned to fit the newly acquired GPR dataset, the average accuracy of rebar detection can reach 90.9% [28]. However, the fine-tuning approach may be less applicable in the face of more complex domain offset problems, such as variations in visible data due to differences in illumination conditions and image appearance. To overcome this problem, lv et al. [82] proposed an unsupervised migration learning framework, which improves the generalization ability of the model by constructing intermediate domains to simulate the target domain properties and combining it with an illumination-aware label fusion strategy, which achieves automatic adaptation to the target domain without additional manual annotation. It is worth noting that the success of migration learning depends largely on the similarity between the source and target domains, as well as the knowledge distillation and alignment strategies during the migration process. Therefore, in concrete crack detection and other specific application scenarios, researchers need to choose the most suitable model and method according to the data characteristics; e.g., comparing the performance of Alexnet, VGG16, VGG19, and ResNet-50 models in terms of validation and testing accuracy and training time, it is found that the ResNet-50 model maintains a high accuracy rate with the shortest training time and is the most suitable for concrete crack image classification [83].

4.3.2. Multimodal Data Fusion Technology

Multimodal data fusion technology breaks through the limitations of traditional isolated data processing by extracting visual attributes from data from different sensors and achieves a more comprehensive information feature system construction [84]. Currently, deep learning-based fusion methods mainly include frameworks such as Autoencoder, Convolutional Neural Networks (CNNs), and Generative Adversarial Networks (GAN). M²FNet, as an innovative multimodal fusion network, combines the local feature extraction capability of CNNs with the global characterization capability of Transformer through its dual-network architecture, extracts features from the raw signal data and B-scan and top-scan GPR image data, and achieves multimodal feature fusion through the cross-attention mechanism, which significantly improves the learning ability of GPR data [80], as shown in Figure 11a. Meanwhile, aiming at the comprehensive detection of metal internal and surface defects, the multi-modal automatic defect detection (M-ADR) system constructed by the fusion of ultrasonic and pulsed thermal imaging (PT) data via coordinate transformation technology realizes the optimization of defect identification and characterization, as shown in Figure 11b, and the detection accuracy rate is up to 91.46%, far exceeding the traditional model and single-mode technology [85]. For the fusion of infrared and visible images, Faegheh et al. [86] proposed the EFASPP U-Net deep network, consisting of two independent encoders (dealing with visible and thermal images, respectively) and a decoder, which combines a standard convolutional layer with a null convolutional layer to achieve feature-level fusion and avoid the loss of information for small targets. Figure 11c shows the block combining the visible and thermal feature maps. In addition, Wang et al. [13] used GAN’s Infrared–Visible Image Fusion (IVIF) technology and relied on the Resnet generator (as shown in Figure 11d) and PatchGAN discriminator to integrate thermal imaging and visual information, and built a deep neural network to achieve efficient case segmentation of building facade defects (cracks, spalling, etc.) in IVIF images.

5. Conclusions and Outlook

This paper reviews and summarizes the progress of the application of deep learning algorithms in the field of non-destructive building inspection, and combs through the latest frontiers of image-based non-destructive inspection techniques, especially infrared thermal camera, ground-penetrating radar, and ultrasonic inspection techniques. Each of these techniques has its strengths and limitations, so exploring the integrated application of multiple inspection methods has become an important way to improve inspection accuracy. In addition, precise adjustment of the parameters of non-destructive sensors is essential to ensure the accuracy of data acquisition. For precast concrete members with known defect distribution, non-destructive testing (NDT) images can be used to confirm the defect conditions by cross-referencing with known defects. However, in real-world scenarios where unknown damage exists, such as in building walls, NDT, as an indirect means, often needs to be combined with traditional testing to provide more comprehensive information. At the same time, the distinction of damage types in NDT images often relies on professional experience, which poses a challenge to deep learning models in damage matching and recognition, requiring the models to possess higher accuracy and generalization ability.

With the continuous development of deep learning and NDT technology, the integration of NDT and deep learning shows new trends in the future. On the one hand, data enhancement methods focus on image generation and super-resolution techniques based on Generative Adversarial Networks (GAN), aiming to improve the diversity, resolution, and feature retention of NDT image data, solve the problem of sample scarcity and labeling, and provide richer and more accurate image resources for deep learning models. On the other hand, the development of deep learning models focuses on the optimization of transfer learning techniques and the innovative application of multimodal data fusion techniques. Through migration learning, the dependence on new domain datasets is reduced, and the adaptability and generalization ability of the model are improved. Although some progress has been made in the field of visible light images, the application in the fields of infrared thermal imaging, radar 2D imaging, and ultrasound imaging needs to be further strengthened. In addition, further work is required in terms of combining multimodal data fusion technology, innovative network architecture, and fusion strategy, making full use of different sensor data resources and achieving more accurate information feature extraction and fusion, in order to improve the performance and robustness of tasks such as target detection, and to cope with the challenges of complex and changing scenes.

However, this study has certain limitations. It does not delve deeply into the theoretical aspects of deep learning techniques, nor does it provide a comprehensive discussion on the specific deep learning models suitable for different NDT methods and their optimization strategies. Future research should focus on extending the application of deep learning to more NDT methods to improve the quantitative accuracy of defect sizes, ensure intuitive and accurate inspection results, and provide reliable data support for building structural health assessment and maintenance.