A Novel Surface Defect Detection Method for Carbon Fiber Composite Plates Based on Super-Resolution Reconstruction and U-Net Network

Yixiong Feng; Peiyan Pan; Qi Kong; Bingtao Hu; Zhenghao Sun; Junliang Wang; Jianrong Tan

doi:10.1016/j.eng.2025.09.033

Engineering ›› 2026, Vol. 63 ›› Issue (8) :123 -136. DOI: 10.1016/j.eng.2025.09.033

Research

research-article

A Novel Surface Defect Detection Method for Carbon Fiber Composite Plates Based on Super-Resolution Reconstruction and U-Net Network

Author information +

History +

PDF (4619KB)

Abstract

Rapid and accurate detection of surface defects has become critical with the increasing demand for highly reliable carbon-fiber composite plates (CFCPs) in advanced manufacturing. This study proposes a dual-stage enhancement framework to detect subtle defects in CFCPs, addressing the limitations of conventional down--sampling and feature--extraction methods, especially when samples are limited and defects are subtle. The framework highlights the critical role of expert knowledge in defect detection and allows effective parameter transfer between a task-specific super-resolution reconstruction module and a residual, multiscale fusion semantic segmentation network. Experiments on a digital--radiography CFCPs data set demonstrate that this method markedly amplifies weak defect signatures and pinpoints their locations with high fidelity. The findings exhibit significant gains in precision, recall, F1--score, and mean intersection—over--Union relative to U-Net, SegNet, and other baselines. In small--sample conditions, the proposed model nearly doubles the performance of the canonical U-Net. This framework offers a broadly applicable solution for automated micro--scale defect inspection across composite-material systems and other advanced--manufacturing contexts.

Graphical abstract

Keywords

Carbon fiber composite plates / Defect detection / Super-resolution reconstruction / Residual learning / Semantic-segmentation / Digital radiography

Cite this article

Download citation ▾

Yixiong Feng, Peiyan Pan, Qi Kong, Bingtao Hu, Zhenghao Sun, Junliang Wang, Jianrong Tan. A Novel Surface Defect Detection Method for Carbon Fiber Composite Plates Based on Super-Resolution Reconstruction and U-Net Network. Engineering, 2026, 63 (8) : 123-136 DOI:10.1016/j.eng.2025.09.033

登录浏览全文

4963

注册一个新账户忘记密码

1. Introduction

The growing aerospace and defense industries are driving demand for advanced, lightweight materials, making carbon fiber composite plates (CFCPs) highly suitable [1]. Their low density, high specific strength, and corrosion-resistant properties have led to their widespread use in applications, ranging from satellite solar panels and aircraft engine fan blades to wind turbine blades, vehicle structural components, and microelectronic packaging chips [2,3]. Due to limitations in material technology, manufacturing processes, and harsh service conditions, CFCPs often develop surface or internal defects, including cracks, pores, and inclusions, that can accumulate over time and weaken structural integrity [4,5]. Therefore, early and accurate detection of these flaws, especially subtle ones, is critical for ensuring reliability in practical applications [6].

Nondestructive testing (NDT) techniques have attracted widespread attention and are widely used to efficiently identify and monitor the aforementioned damage and defects during production and operation [7]. However, the CFCPs’ manufacturing complexity and cluttered surfaces, intricate internal structures, and weak defect signals often obscure these minute flaws amidst background noise, creating these detection challenges [8,9].

To detect and monitor various damage and defects in CFCPs during production and service, several NDT techniques have been developed [10,11], including ultrasonic testing (UT) [12], infrared thermography (IIT) [13], optical inspection [14], and eddy current testing [15,16]. For example, Wang et al. [17] addressed the interlayer defects detection in carbon fiber composites by proposing a modal decomposition imaging method using air-coupled Lamb wave detection. Wong et al. [18] developed a method that employed a two-dimensional (2D) ultrasonic array sensor to measure the elastic characteristics of carbon fiber reinforced polymer (CFRP) plates. Caminero et al. [19] utilized phased array ultrasonic technology to locate simulated inclusion defects and impact damage in CFRP precisely.

However, UT of shallow surface defects is often complicated by coupling agents, curved geometries, and noise from wave propagation in thin CFRP laminates, making defect identification more challenging [20,21]. IIT has gained popularity in NDT in recent years. It detects variations in infrared radiation between defective and intact regions, allowing for the identification of subsurface flaws without damaging the materials [13].

Avdelidis et al. [22] demonstrated that pulsed thermography (PT) is applicable to diverse structural defects, including those in CFRPs, honeycomb cores, and sandwich composites. Building on this, Azizinasab et al. [23] developed a transient response phase extraction method for PT using “local reference pixel vectors” enabling reliable defect detection in CFRPs. Similarly, Dong et al. [24] employed infrared lock-in thermography for delamination inspection in CFRPs and incorporated deep learning to enhance subsurface defect characterization. In related work, Yang et al. [25] integrated induction IIT with thermal-wave radar to identify faults in CFRP wind turbine blades.

Overall, IIT provides quick contact-free detection over large surfaces, offering intuitive and efficient methods for identifying delamination, debonding, and impact damage in composite materials. However, limited detector performance can obscure deeper defects, reducing resolution.

While optical inspection techniques can produce high-resolution images in the visible spectrum, balancing efficiency and accuracy in industrial settings, especially where minor defects, complex morphologies, and limited data remain challenging. Cui et al. [14] proposed a vision-assisted strategy for extracting and quantifying delamination in CFCPs, including a robust delamination extraction algorithm and a comprehensive evaluation framework. However, conventional optical methods often face practical limitations in real-world production environments, where surface flaws in CFCPs can be challenging to detect due to constrained data, intricate geometries, and illumination.

Eddy current testing identifies defects by measuring changes in coil impedance and induced voltage, arising from interactions between the alternating magnetic field of the coil and the magnetic field produced by currents in defective regions. This technique is commonly used to identify surface, near-surface cracks, fiber fractures, and delamination in CFCPs. Zou et al. [26] applied eddy current thermography to evaluate CFRP-steel interfaces, inducing eddy currents in the steel plate to detect and characterize interface failures. However, due to the skin effect, eddy currents decay rapidly through the material thickness, and the excitation frequency further limits penetration depth, making it challenging to identify deeper defects in composite materials.

X-ray inspection is a widely used industrial NDT method, using its strong penetration capability to visualize internal structures and fine surface defects [27]. Using this approach, X-rays pass through the object being examined, and the attenuated signals are converted into electronic data. These data are then processed into real-time grayscale images, offering a 2D view of the object’s internal architecture [28,29].

Recently, digital radiography (DR) has become increasingly popular in the NDT field, thanks to its real-time imaging, strong penetration power, relatively simple operation, and convenient storage of inspection data. Gao et al. [30] focused on the NDT of aluminum conductor composite core (ACCC) cables. They introduced the LineSpyX robot, which incorporates a portable X-ray emitter and a DR detection unit to provide a unified NDT solution for imaging the carbon fiber core within ACCC cables. These images are analyzed using deep learning and manual diagnosis to identify defects.

For CFCPs, DR can penetrate multiple carbon fiber layers, revealing critical information about internal or near-surface defects, offering insights into defect distribution and morphology. Nonetheless, imaging still faces challenges. Limitations such as the resolution of the imaging system, photon shot noise, and the complex interlayer structure of composite materials can degrade image quality, leading to noise, low contrast, and difficulty in detecting minute defect signals [30]. Detecting and enhancing these weak defect features in DR images remains a key challenge for enhancing detection accuracy and reliability.

Recently, machine learning and deep learning have become vital tools for challenging industrial inspection tasks, thanks to their powerful feature extraction and pattern recognition capabilities in image processing. Researchers have investigated super-resolution reconstruction approaches without modifying existing imaging hardware [31] to overcome the limitations of inadequate image resolution and contrast in defect detection. By applying deep convolutional networks to perform pixel-level reconstructions on low-resolution images, high-definition outputs can be generated, often matching or surpassing the original hardware’s resolution limits. Deep learning-based reconstruction models are more flexible and adaptable than conventional super-resolution methods, which rely on prior knowledge and precise image registration, and this makes the former a highly promising alternative for high-magnification image reconstruction [32].

In industrial inspection, Ijjeh et al. [33] introduced a deep learning-based super-resolution method to quickly obtain full-wavefield Lamb wave data from low-resolution measurements, surpassing traditional compressed sensing (CS) methods. Wu et al. [34] developed a CS-based post-processing algorithm for super-resolution thermal imaging to address the issue of limited high-speed imaging in uncooled infrared sensors for NDT. By integrating high-resolution infrared images with pulse phase thermography (PPT), they validated the approach on adhesively bonded CFRP joints, demonstrating highly effective visualization of subsurface defects.

Furthermore, as deep learning has been successfully applied to image semantic segmentation and object detection, researchers have turned to convolutional networks, generative adversarial networks (GANs) [35,36], and other models for defect recognition and segmentation, making significant progress [37,38]. Huang et al. [39] proposed a fully convolutional network (FCN)-based algorithm for metro shield tunnels, enabling rapid and accurate defect identification. The goal is to address the low efficiency of traditional hand-crafted visual inspections for detecting tunnel cracks and leakage. In their work, Feng et al. [40] developed the cross-position aggregation network (CPANet), a framework tailored to few-shot segmentation of defects on strip steel surfaces. By integrating cross-position information exchange, auxiliary support guidance, and a spatial attention compression strategy, the network effectively captures multiscale context and sets a new performance benchmark on the FSSD-12 dataset. Building on ResNet-50 and FCN, Li et al. [41] developed FCS-Net for precise crack segmentation in full-scale steel bridge images with complex backgrounds. To address diverse textures and defect types in fabrics, Liu et al. [42] introduced a multi-stage approach combining GANs with a deep semantic segmentation network, continuously updating an existing fabric defect dataset to improve detection accuracy. Despite the progress of FCN- and GAN-based methods across diverse industrial inspection tasks, further improving segmentation speed and accuracy remains critical, especially for minor, morphologically complex defects in noisy or cluttered environments [37].

U-Net, a typical encoder-decoder architecture, has shown remarkable extensibility and portability in NDT [43,44]. Its design preserves global and local image features while restoring spatial resolution and positional information through skip connections, thereby improving defect localization. The U-Net is particularly suitable for detecting surface defects in CFCPs [45]. For example, Liu et al. [46] deployed a U-Net-based segmentation method for conveyor belt imaging to mitigate the adverse effects of adhesion and blurred edges. Similarly, Wang et al. [47] integrated attention mechanisms and residual modules into a U-Net (smoke-U-Net) for detecting forest fire smoke. Wang et al. [48] used a three-dimensional (3D) U-Net segmentation model in in-situ X-ray computed tomography (CT) for carbon fiber composite inspection. This approach accurately captures microstructural features and damage within carbon fiber composites, effectively addressing the low contrast and multidirectional complexities that challenge traditional segmentation techniques.

Recently, a variety of deep-learning segmentation architectures have been extensively explored for defect identification in composite materials [39,42,46]. However, most existing approaches directly adopt general-purpose segmentation architectures without sufficiently adapting to the specific characteristics of CFCPs, such as complex surface textures, weak signals from micro-scale defects, and limited datasets. Consequently, these traditional techniques are susceptible to interference from subtle defect signals and complex backgrounds, severely compromising their performance in accurately identifying minor defects, especially micro-cracks and micro-pores. Additionally, deep learning-based super-resolution enhances image resolution without hardware upgrades, making it attractive for industrial inspection [32], [33], [34]. Nevertheless, existing research has predominantly focused on visible or infrared imaging, with little empirical validation for defect detection in CFCPs using DR [42]. These DR images pose two additional obstacles: First, the oriented fibers create directional artefacts that can mislead super-resolution models; second, these DR images suffer from Poisson noise and low contrast, which complicate end-to-end training [30].

This study proposes a novel dual-stage joint enhancement framework specifically designed to boost subtle defect signals and preserve critical defect features in CFCPs. The first stage used task-specific super-resolution reconstruction to enhance the subtle defect signals significantly. Subsequently, we developed an improved semantic segmentation network based on a deep residual-multiscale feature fusion U-Net architecture to extract these enhanced defect features precisely. Unlike existing techniques that treat super-resolution reconstruction and segmentation as isolated procedures, this approach uniquely integrates domain-specific expert knowledge into the conventional machine learning pipeline, facilitating the cooperative interaction between the two stages. This technique ensures high-fidelity preservation and transmission of defect information, effectively overcoming the commonly encountered problem of subtle signal loss in DR-based CFCPs inspections.

Moreover, this paper refines the canonical U-Net architecture by innovatively embedding deep residual modules and a multi-scale feature-fusion mechanism. Residual connections mitigate the gradient vanishing issues in training U-Net architectures with small-scale datasets. Simultaneously, the multiscale fusion strategy significantly enhances the model’s detection of subtle defects across different scales. This approach overcomes the shortcomings of single-scale segmentation networks, which lack sufficient sensitivity to defects at varying scales. Compared to mainstream methods, this approach demonstrates significantly improved accuracy and robustness in micro-scale defect detection, especially under limited data and high-noise conditions, highlighting its practicality and advantages.

The rest of this paper is structured as follows: Section 2 presents the proposed approaches, detailing the deep residual-based super-resolution reconstruction and improved U-Net network structure, along with the evaluation criteria. Section 3 describes the overall process of CFCPs defect detection. Section 4 showcases results from various stages of network and data processing. Section 5 provides a comparative analysis of the experimental results. Section 6 presents the conclusions of the entire manuscript.

2. Methods

A novel deep learning-based framework for super-resolution reconstruction and semantic segmentation was introduced, effectively addressing the complex challenge of surface defect detection in CFCPs, especially in complex backgrounds with limited sample sizes. The overall structure of the framework is depicted in Fig. 1.

2.1. Human expertise in the framework

Human involvement serves as an intermediary between image super-resolution reconstruction and defect semantic segmentation. This framework includes annotating defect regions in sample images and configuring experimental sensitivity parameters for the super-resolution reconstruction network and the modified U-Net network. These discrete steps highlight the importance of human expertise within the overall framework.

First, it is undeniable that the dataset significantly influences the performance and outcome of supervised learning algorithms [49], [50], [51]. In defect detection tasks, the network learns to identify defect characteristics through repeated training on labeled samples, thus building a prediction model. Due to the complex image backgrounds of CFCPs datasets, non-defective regions are often mistakenly classified as defects, making accurate labeling of defect areas crucial. To address this problem, the “LabelMe” tool meticulously delineated the defect boundaries [52,53]. After annotations, the images were converted into labels for model training and validation. The specific annotation procedure is illustrated in Fig. 2.

Second, network structure parameters profoundly impact the algorithm’s stability in deep learning and are the key to developing a high-quality learning model [54]. Since defect detection steps are relatively independent, various parameters cannot be optimized automatically. Therefore, human analysis and decision-making are necessary to coordinate parameter tuning across the framework, enhancing the system’s stability and performance.

2.2. Deep residual model

CFCPs often exhibit numerous small-scale and sparsely distributed surface defects, making it challenging for conventional down-sampling operations and traditional feature extraction mechanisms in semantic segmentation algorithms to capture sufficient discriminative features. To overcome this limitation, this paper integrates residual learning into the super-resolution reconstruction stage and the semantic segmentation process. The overall architecture is illustrated in Fig. 3(a).

Deep neural networks often face training challenges such as gradient vanishing and gradient explosion, especially in deep architectures. Residual networks address these issues by incorporating residual blocks [55], including skip connections that allow gradients to propagate directly across layers. This design facilitates better gradient flow, stabilizes training, and improves convergence in deep architectures. Notably, residual connections stabilize training and enhance feature representation. By incorporating input features into the output of convolutional layers, residual blocks enhance low-level detail information, such as defect edges and delicate textures, while retaining and propagating to higher layers. This capability is crucial to defect localization tasks and guarantees that critical visual cues are preserved during deep feature abstraction [56].

To detect micro-defects in low-resolution DR images, we embed residual modules in the super-resolution reconstruction and the segmentation branches. Unlike the baseline U-Net with simple skip connections, our architecture employs deep residual learning to enhance gradient flow and preserve fine-grained spatial information. Recent multiscale architectures such as CPANet [40] emphasize global context modeling. In contrast, our approach integrates global context preservation with heightened sensitivity to local, micro-scale defects, aligning with the complex textures and low annotation in typical CFCP datasets.

Within the multiscale fusion framework, residual skip connections facilitate the incorporation of fine-grained features into higher-level semantic abstractions. This integration enables the network to recover detailed high-resolution representations from low-resolution inputs to delineate defect contours, enhancing prediction accuracy and segmentation robustness. Additionally, residual modules substantially enhance training stability by alleviating gradient degradation [55,56]. To boost feature extraction capacity, we stack multiple residual units, a deeper residual network, offering richer feature hierarchies, faster convergence, and improved segmentation performance, particularly in low-contrast, fine-defect detection scenarios [57].

The deep residual module, shown in Fig. 3(b), consists of 16 two-layer residual modules, each comprising two 3 × 3 convolutions with an identical number of output channels. For an input X, each layer’s learning process is represented by the mapping X + F(X). In this scenario, the natural X + F(X) tends to approximate 0, revealing that as the number of layers increases, the error F(X) remains minimal. This approach has fewer parameters than traditional convolutional neural networks with the same number of layers, thereby mitigating degradation issues during the down-sampling process [58].

Additionally, the piecewise linear function rectified linear unit (ReLU) is applied between successive convolutional layers, helping the network learning complex relationships within the data [59].

2.3. Super-resolution reconstruction

Efficient and accurate image-resolution enhancement remains central to super-resolution research. Building upon the super resolution generative adversarial network (SRGAN) paradigm [60], our approach fuses sub-pixel convolution (pixel shuffle) with adversarial learning to address the exacting demands of CFCPs DR imagery.

First, the generator employs sub-pixel convolution modules to up-sample images. These modules produce multi-channel feature maps during the convolution stage and spatially rearrange these features according to predefined rules, doubling the image’s spatial dimensions. Unlike transposed convolution or bilinear interpolation, sub-pixel convolution better preserves high-frequency details and minimizes the blurring that arises during the rearrangement operation. Multiple sub-pixel convolution stacked modules enable larger-scale resolution enhancement. As a result, the generator output closely resembles accurate high-resolution images and restores critical textures.

A discriminator is introduced for adversarial training based on this structure. Moreover, it differentiates between authentic high-resolution images and super-resolution outputs, compelling the generator to iteratively refine the quality of its samples throughout training. Unlike generic super-resolution methods such as SRCNN or the standard SRGAN, designed for natural-image content, our approach is optimized for DR images of CFCPs. Our technique develops a domain-adaptive super-resolution reconstruction strategy that accentuates fine-scale details without introducing perceptible noise or spurious defects. The structure of the proposed method is shown in Fig. 4.

Experiments were performed using three widely recognized benchmark datasets: SET 5, SET 14, and BSD 100, for training and verification [58]. Peak signal to noise ratio (PSNR) and structural similarity (SSIM) were used as metrics to evaluate the similarity between the reference and the processed images [58,61]. The PSNR is defined as follows:

(1)$\mathrm{PSNR}=10 \log _{10} \frac{\mathrm{MaxValhe}^{2}}{\mathrm{MNE}}=10 \log _{10} \frac{\left(2^{\text {lits}}-1\right)^{2}}{\mathrm{MNE}}$

In this context, bits represents the number of bits each pixel utilizes for storage, and the mean squared error (MSE) serves as a measure of reconstruction fidelity. MaxValue is the maximum possible pixel value of the image. Its expression is given by:

(2)$\mathrm{MSE}=\frac{1}{n} \sum_{i=1}^{n}\left|g\left(X_{i} ; \theta\right)-g_{i}\right|^{2}$

where n is the total number of pixels in the image, g(X_i; θ) signifies the predicted data value of the ith pixel, g_i is the original data value of the ith pixel.

Since subjective evaluation is challenging, the SSIM is designed to encompass three performance parameters: luminance, contrast, and structure, aligning with the characteristics of the human visual system (HVS) [60]. The mean value quantifies luminance, contrast by the variance after mean normalization, and structure by the correlation coefficient. If the input images are x and y, these components are defined as $l(x, y), c(x, y) $, and $s(x, y)$ as follows:

(3)$l(x, y)=\frac{2 \mu_{x} \mu_{y}+c_{1}}{\mu_{x}^{2}+\mu_{y}^{2}+c_{1}}$

(4)$c(x, y)=\frac{2 \sigma_{x} \sigma_{y}+c_{2}}{\sigma_{x}^{2}+\sigma_{y}^{2}+c_{2}}$

(5)$s(x, y)=\frac{\sigma_{x y}+c_{3}}{\sigma_{x} \sigma_{y}+c_{3}}$

where $\mu_{x}$ and $\mu_{y}$ denote the mean values of x and y, $\sigma_{x}$ and $\sigma_{y}$ represent the variances of x and y, $\sigma_{x y}$ indicate the covariance of x and y.

The constants $c_{1}$, $c_{2}$,and $c_{3}$ are introduced to prevent system errors caused by a zero denominator and stabilize the calculations when the denominators approach 0, simplifying the SSIM index as follows:

(6)$\operatorname{SSIM}(x, y)=\frac{\left(2 \mu_{x} \mu_{y}+c_{1}\right)\left(2 \sigma_{x y}+c_{2}\right)}{\left(\mu_{x}^{2}+\mu_{y}^{2}+c_{1}\right)\left(\sigma_{x}^{2}+\sigma_{y}^{2}+c_{2}\right)}$

2.4. U-Net

The U-Net network, an enhancement over the FCN network structure, was originally designed for image compression. U-Net’s flexibility has made it a powerful tool in medical image segmentation, and recent studies also report its promising performance in defect detection under intelligent manufacturing scenarios [62,63].

As a typical encoder-decoder network, U-Net can extract fundamental information from an image and restore it to its original size using deconvolution and pooling [43]. U-Net used a breakthrough skip-connection mechanism that preserved information extracted by a down-sampling convolutional layer within a given receptive field. This information is fed back to the corresponding up-sampling deconvolution layer, resulting in more complete information recovery [44]. Additionally, the network’s number of convolutional and pooling layers can be adjusted according to specific scenarios, offering the U-Net considerable adaptability [64].

Given the relatively subtle CFCPs’ defects, it is crucial to gather more information about the defect areas during image feature extraction for subsequent network learning. In the basic U-Net network, the pooling operations associated with convolutional kernels during down-sampling progressively enlarge the receptive field of deeper layers. However, merely stacking additional layers without modifying kernel design fails to provide the network with access to richer image information. To mitigate information loss during convolutional pooling, the U-Net architecture was adopted as the baseline model, with a residual module integrated into the down-sampling path to enhance its feature learning capability [55].

The feature map from the corresponding down-sampling position can be fused within the channel during up-sampling at each network level. The integration of shallow and deep features allows the model to maintain fine-grained spatial details, ultimately leading to more precise image segmentation. Fig. 5 illustrates the structure of the improved network, featuring major components similar to U-Net: down-sampling (left side), skip-connection, and up-sampling (right side).

The U-Net network’s down-sampling process uses a standard convolutional architecture. The encoding path consists of repeated 3 × 3 convolutions (without padding) with ReLU activation, followed by a 2 × 2 max pooling at a stride 2. Each down-sampling operation doubles the number of feature maps. The expansive path mirrors this process: feature maps are first up-sampled, then passed through a 2 × 2 convolution (“up-convolution”) that reduces the channel count by half. These are concatenated with the corresponding cropped feature maps from the encoder, after which two additional 3 × 3 convolutions with ReLU activation are applied. Cropping is required to compensate for boundary pixel loss introduced by convolution. Finally, a 1 × 1 convolution projects the resulting 64-dimensional feature representation into the target number of classes. Overall, the network comprises 23 convolutional layers [44].

To optimize multivariate outputs, the enhanced U-Net leverages gradient descent combined with randomized initialization strategies to strengthen its generalization capacity. Stochastic elements are commonly introduced during deep neural network training, for instance, through dropout, gradient noise injection, or asynchronous stochastic optimization.

Computers cannot represent samples with infinite-dimensional precision, providing the neural network with more data than expected [65]. Constraining the flow of information within the network and introducing stochasticity encourages a more compact encoding, forcing the model to learn succinct representations. The network updates parameters by reducing the average loss error over the entire training dataset. Using stochastic gradients in mini-batches is an efficient method to avoid local minima during training.

Considering the limited dataset size, this study aims to develop a training model with minimal loss and high accuracy by increasing the network iterations. However, this approach significantly increases the risk of model overfitting, where the model fitting function coefficients become excessively large, and the function values change drastically over small intervals. To mitigate this approach, weight decay (i.e., L₂ regularization) is introduced to constrain the norm of the parameters, thereby balancing model loss and hyperparameters. L₂ regularization (C) is expressed as follows:

(7)$C=C_{0}+\frac{\lambda}{2 n}\|\omega\|_{2}^{2}$

where C₀ represents the original cost function, ω denotes the weight, and λ is the weight decay coefficient. Taking the partial derivative with respect to ω, we obtain the following expression:

(8)$\frac{\partial C}{\partial \omega}=\frac{\partial C_{0}}{\partial \omega}+\frac{\lambda}{n} \omega$

Thus, the change trend of ω is as follows:

(9)$\omega \rightarrow\left(1-\frac{\eta \lambda}{n}\right) \omega-\eta \frac{\partial C_{0}}{\partial \omega}$

where η is a constant characterizing the learning rate, or the speed of change.

The loss function measures the deviation between predicted outputs and ground truth labels, where lower values reflect improved robustness and generalization of the model. Cross-entropy loss, frequently used in semantic segmentation, serves as an effective objective for binary and multi-class problems. Its mathematical form is given as follows:

(10)$\operatorname{loss}=-\frac{1}{N} \sum_{i=1}^{N}\left[y_{i} \log \left(\dot{y}_{l}\right)+\left(1-y_{i}\right) \log \left(1-\dot{y}_{l}\right)\right]$

where N is the number of pixels, $y_{i}$ is true label, and $\dot{y}_{l}$ is the prediction probability.

3. Data preprocessing

Our objective is to accurately detect weak defects in the small samples of CFCPs. Deep learning-based defect detection requires the network to continuously learns the features of defect areas from samples and labels to develop a predictive model. Accurate and precise labeling is critical, as it directly influences the model’s quality, particularly when the sample size is limited.

3.1. Dataset

Representative samples are shown in Fig. 6. A total of 85 images are included in the dataset, each with a resolution of 256 × 256, obtained from the Shanghai Satellite Equipment Research Institute (China) using a high-sensitivity digital ray detection system, is depicted in Fig. 6. The system operates with a maximum tube voltage of 150 kV, a maximum tube current of 50 mA, a focal size of 100 µm, and a maximum matrix size of 1000 × 1000.

Metallographic experiments of the CFCPs identified two main defect types: voids and raised foreign matter. Following meticulous manual labeling, 60 images were allocated for training, while an additional 25 were set aside for validation and testing. The size of each image is 100 mm × 100 mm.

3.2. Super-resolution construction of CFCPs images

3.2.1. Steps of super-resolution reconstruction

Our approach initially utilized a large-scale public image dataset (SET 5, SET 14, and BSD 100) for preliminary training of our super-resolution reconstruction and segmentation models more efficiently. While these generic images differ significantly from CFCPs in appearance and structural patterns, they provide a rich source of low-level features (edges, contours, and basic textures). These learned features are applicable in early layers of deep networks, serving as a foundation for high-resolution image reconstruction. This paper conducts domain-specific fine-tuning on our in-house CFCPs dataset to tailor the model to the unique surface texture and defect features of CFCPs. Experimental comparisons reveal that models pre-trained on large-scale public data can achieve superior performance in detecting micro-scale defects. This result suggests that while the final semantic layers must be refined on domain-specific samples, the core representation capabilities acquired from extensive generic data remain beneficial. The dataset reconstruction proceeds in three main steps:

(1) Step 1: downsampling a large number of real-world image samples from three public datasets. The low-resolution image L obtained in this step and the high-resolution image H before down-sampling form an image pair for subsequent model training.

(2) Step 2: train the model. Using the proposed network structure, the low-resolution image L, obtained in the initial step, is first reconstructed into SH, which is subsequently compared against the original image H. Afterward, the error derived from this comparison adjusts the parameters, and through repeated iterations, the training process converges to a pre-trained model with minimized error.

(3) Step 3: We reconstruct the low-resolution images of the dataset using the trained model to obtain high-resolution images. The training process was accelerated using an RTX 2060 graphics card (NVIDIA, USA). During the public dataset training, the variation of the MSE loss function is shown in Fig. 7. As the training progresses, the loss function gradually converges as shown in Fig. 7, converging to 0.0165-0.017 at the end.

3.2.2. Optimal magnification for super-resolution reconstruction

The initial CFCPs’ surface image resolution is 256 × 256 pixels. Therefore, super-resolution reconstruction is employed to enlarge the images, with magnification factors set to 2, 4, and 8, adjusted by the number of pixel shuffle modules. However, excessively large magnifications impose a greater computational burden on the super-resolution reconstruction network. Moreover, the subsequent U-Net semantic segmentation network increases the risk of gradient explosion [35].

Fig. 8 presents the comparative outcomes obtained from the standard U-Net and its enhanced variant. Then, we verified the original image (256 × 256), twice-enlarged image (512 × 512), four-times-enlarged image (1024 × 1024), and eight-times-enlarged image (2048 × 2048).

Section 4.1 presents the evaluation criteria for the prediction results, including pixel accuracy (PA), class pixel accuracy (CPA), mean pixel accuracy (MPA), and mean intersection over union (MIoU) [41]. As observed in Fig. 8, the prediction results using the improved U-Net training model outperform those obtained with the standard U-Net network training model for images with the same magnification.

A comparison between Figs. 8(a) and (b), and Figs. 8(c) and (d) reveal that the improved U-Net network outperformed PA and CPA for the same magnified images under different iterations. The improved U-Net leads to a smaller range of values, a higher minimum value, and a larger median value. The prediction results for images without magnification, double magnification, quadruple magnification, and eight-times magnification are significantly higher than those for eight-times magnified images. The former exceeds 0.95 and the latter ranges between 0.85 and 0.9, indicating higher pixel classification accuracy and better performance for images magnified two or four times using the improved U-Net network.

In Figs. 8(e) and (f), for images with no magnification or eight-times magnification, the MPA of the model trained with the enhanced U-Net exhibits greater variability but a higher median value. The upper part of the probability map has a larger width, suggesting a higher probability of values above 0.9. For images magnified two or four times, the predicted CPA values using the improved U-Net network show smaller fluctuations, a larger median, and a minimum value above 0.92. Overall, applying the improved U-Net network and magnifying images by factors of two or four produces better results.

Comparing Figs. 8(g) and (h), the MIoU value remains unchanged by different training models for images not magnified or magnified two or four times. However, for images magnified by a factor of eight, the values start to fluctuate. Specifically, the MIoU value for the original image is about 0.85, approximately 0.74 for twice and four-times magnified images, and about 0.83 for eight-times magnified images.

While validation results are similar, we opt for the approach that is faster and consumes fewer resources. Therefore, a magnification factor of 2 for the CFCPs dataset is chosen, resulting in a fourfold increase in resolution for the images acquired by the high-sensitivity digital ray detection system.

Fig. 9 shows the super-resolution reconstruction effect of the public data during the training process, with 100 iterations. Fig. 9(a) displays the original images, and Fig. 9(b) shows a low-resolution image acquired from a down-sample of the original images. With more iterations increases, the quality of super-resolution reconstruction progressively improves. As illustrated in Fig. 9(c), the block noise is substantially reduced in the final restored image, yielding a much smoother visual result. The evaluation criteria of each public dataset in the training and reconstruction process are shown in Table 1.

The PSNR values for all three public datasets are above 25, and the SSIM values exceed 0.65, demonstrating the effectiveness of the super-resolution reconstruction model obtained through training.

Fig. 10 compares a CFCP image before and after super-resolution reconstruction, based on the fine-tuned model. The comparison highlights the original low-resolution input (left) and the improved high-resolution result achieved through super-resolution (right).

4. CFCPs defect detection result

4.1. Evaluation criterion

In image segmentation, algorithm performance is generally assessed using metrics such as PA and different forms of MIoU. Four main evaluation criteria are introduced to verify the model accuracy as follows:

(11)$\mathrm{PA}=\frac{\sum_{i=0}^{k} P_{i i}}{\sum_{i=0}^{k} \sum_{j=0}^{k} P_{i j}}$

(12)$\mathrm{MPA}=\frac{1}{k+1} \sum_{i=0}^{k} \frac{P_{i i}}{\sum_{j=0}^{k} P_{i j}}$

(13)$\mathrm{CPA}=\frac{P_{i i}}{\sum_{j=0}^{k} P_{i j}}$

(14)$\mathrm{MIoU}=\frac{1}{k+1} \sum_{i=0}^{k} \frac{P_{i i}}{\sum_{j=0}^{k} P_{i j}+\sum_{j=0}^{k} P_{j i}-P_{i i}}$

For a better explanation, the following assumptions are made: there are k+1 classes (from L₀ to L_k, including an empty class or background), and P_ij represents the number of pixels that belong to the class i but are predicted to be class j. Thus, P_ii indicates the true positives, while P_ij and P_ji are interpreted as false positives and false negatives, respectively.

4.2. Experiment setup

Four sets of control experiments are established to thoroughly compare the effectiveness of the proposed method. Table 2 details the datasets and semantic segmentation networks used in each experimental group. The method’s efficacy is confirmed by controlling a single variable and comparing validation results across these four groups. We collected 85 image samples, with 60 used for training and 25 reserved for validation and evaluation. Every experiment was carried out independently on a 64-bit system with a 7th-generation Intel processor (Intel, USA), 16 GB of memory, and Python 3.10.

4.3. Result and comparison

Table 2 presents the experimental conditions for Experiments 1-4. This study compares experiments with identical conditions to those with varying conditions, aiming to verify the separate and combined effects of super-resolution techniques and the improved U-Net network.

By analyzing Experiments 1 (based on original images + basic U-Net) and 2 (reconstructed images + basic U-Net), as well as Experiments 3 (original images + improved U-Net) and 4 (reconstructed images + improved U-Net), this study evaluates the significance of super-resolution reconstruction on defect detection performance, irrespective of the U-Net variant used. The comparisons between Experiments 1 and 3, and between Experiments 2 and 4, further illustrate the contribution of the improved U-Net to defect detection, regardless of the dataset reconstruction. Finally, comparing Experiments 1 and 4 highlights the substantial performance gains achieved through the combination of super-resolution reconstruction and the improved U-Net network for detecting defects in CFCPs.

First, this paper conducted a systematic evaluation of two CNN architectures: the basic U-Net and an improved U-Net incorporating a residual module for detecting CFCPs’ defects. Experiments were conducted on both the original dataset and its super-resolution reconstructed counterpart, with training iterations set at 100, 200, and 300 epochs. The training loss, as depicted in Figs. S1 and S2 in Appendix A, detailed in Table 3, provides key insights into the convergence behavior and final model performance.

For the unreconstructed dataset, both U-Net variants exhibited a rapid decline in training loss during the initial stage, achieving near-convergence by approximately epoch 50. In the standard U-Net (Table 3), the loss values at epoch 50 were nearly invariant across different iteration counts (0.0139, 0.0142, and 0.0144 for 100, 200, and 300 epochs, respectively), suggesting that the initial convergence rate is relatively independent of the total number of iterations. However, the final stabilized loss decreased markedly with additional iterations, from 0.0092 at 100 epochs to 0.0031 at 300 epochs, indicating that extended training is critical for refining the model.

Similarly, the improved U-Net (Table 3) demonstrated a consistent loss at epoch 50 (0.0136, 0.0138, and 0.0139 for 100, 200, and 300 epochs, respectively), while the final loss decreased from 0.0069 to 0.0024 as the number of epochs increased. Under equivalent training conditions, the improved U-Net had a faster convergence speed and achieved approximately 70%-80% of the final loss observed with the basic U-Net, indicating its superior performance, particularly at 300 epochs.

The experimental results were extended to a super-resolution reconstructed dataset, where the images were magnified fourfold. As shown in Fig. S2, training on these reconstructed images led to a rapid decline in loss during the initial phase, with convergence observed at slightly later epochs, approximately epoch 80 for the standard U-Net (Table 3) and epoch 60 for the improved U-Net (Table 3). The final loss values obtained using the reconstructed dataset were lower than those from the original dataset, with an approximate improvement of 20%. This enhancement suggests that super-resolution reconstruction preserves the convergence dynamics and contributes to a more precise defect detection model without imposing a significant computational burden.

Incorporating a residual module in the improved U-Net architecture facilitates faster convergence and lower final loss values. Although the number of training iterations has little influence on the early convergence speed, longer training ultimately leads to better model performance. The observed reduction in final loss, especially when combined with super-resolution image reconstruction, offers a promising strategy for enhancing defect detection accuracy in carbon fiber materials. Such improvements are crucial for industrial applications where precision and reliability are paramount.

Our findings indicate that although increasing the number of training iterations does not accelerate early-stage convergence, it is essential for achieving an optimal and stable loss. Moreover, the improved U-Net architecture, when coupled with super-resolution preprocessing, demonstrates superior performance over the standard U-Net, making it a compelling approach for high-precision defect detection in carbon fiber materials.

5. Comparative study

To assess the effectiveness of the proposed algorithm, multiple network architectures were evaluated on self-constructed CFCPs defect datasets with image resolutions of 256 × 256 and 1024 × 1024. This paper analyzes the impact of resolution changes on defect detection accuracy and highlights the advantages of the proposed algorithm under these conditions. By comparing it with traditional U-Net, FCN, SegNet, and PSP-Net, we examine the accuracy and stability of the improved U-Net, particularly with insufficient samples, highlighting its potential for carbon fiber defect detection tasks.

As shown in Table 4, at a resolution of 256 × 256, all network models converge well, achieving high PA and class MIoU. However, the improved U-Net outperforms all other models across all indicators, particularly in key performance metrics such as PA (0.9685), MPA (0.4759), and MIoU (0.4814), significantly surpassing other networks. The CPA of 0.9795 for the improved U-Net indicates its ability to maintain high consistency in PA across categories. In contrast, while traditional U-Net and other common networks (e.g., SegNet and PSP-Net) exhibit relatively stable results, their accuracy and stability are slightly inferior.

As shown in Table 4, when the resolution increases to 1024 × 1024, image details are significantly enhanced, improving the model’s performance. The improved U-Net demonstrates notable advantages in detail capture, accuracy, and stability. Its PA (0.9813) and MIoU (0.4923) are the highest among all methods, confirming its excellent performance with high-resolution images. In terms of detail capture and defect segmentation accuracy, the improved U-Net exhibits greater robustness and precision than other networks. The higher resolution allows the improved U-Net to accurately extract subtle features, crucial for precise carbon fiber defect detection.

Similar to the results at 256 × 256 resolution, the improved U-Net shows comprehensive advantages in accuracy (PA) and stability (MIoU) at 1024 × 1024 resolution. Specifically, among the evaluated models, the improved U-Net shows the best results in terms of MIoU and CPA.

Through a comprehensive analysis of the experimental results at both 256 × 256 and 1024 × 1024 resolutions, the improved U-Net excels in all performance metrics, particularly PA, MIoU, and CPA. When working with low- or high-resolution images, the improved U-Net consistently provides stable and high-precision defect segmentation results. By incorporating residual modules and super-resolution technology, the improved U-Net maintains high accuracy and stability, even in small sample scenarios, offering an effective solution for carbon fiber defect detection.

Fig. 11 presents the defect detection results for a representative image from the dataset under various experimental configurations. In Fig. 11(a), potential defect locations (labeled 1-4) are highlighted with yellow dotted lines, indicating that training with the reconstructed dataset in conjunction with the improved U-Net architecture yields the most accurate detection performance. This approach successfully identifies all four potential defect sites, capturing details that other methods often overlook. Compared to Figs. 11(b) and (c), the detection outcome in Fig. 11(e) demonstrates that our approach improved detection accuracy by approximately 50% over the basic U-Net. It has a more robust capability to isolate and recognize subtle or weak defects. Compared to Fig. 11(d), our method demonstrates superior sensitivity in detecting weak defect signals and enhanced capability in suppressing complex texture noise. Such precise detection is critical for CFCPs that typically exhibit complex surface backgrounds and require high sensitivity to minor flaw indications.

6. Conclusion and outlook

This study proposed an innovative dual-stage framework for detecting surface defects on CFCPs, effectively combining super-resolution reconstruction with an enhanced deep semantic segmentation architecture. Unlike conventional methods, which rely exclusively on down-sampling, and traditional feature extraction, which is inadequate for identifying subtle, micro-scale defects, the proposed method integrates a customized super-resolution module to amplify weak defect signals and preserve critical structural details, facilitating subsequent segmentation tasks.

Our approach innovatively introduces deep residual blocks and multiscale feature fusion within the classical U-Net segmentation model. Residual connections ensure stable gradient propagation and preserve low-level defect features, such as boundaries and textures, enabling precise localization of subtle anomalies. Moreover, the multiscale fusion strategy effectively captures defects across varying scales, seamlessly bridging global context with local details.

Extensive experimental evaluation on our in-house CFCPs digital-radiography dataset demonstrates the preeminent performance of our framework, achieving state-of-the-art results in precision, recall, F1-score, and mean MIoU. Our approach exhibits robust performance in severely limited annotated samples, underscoring its practical value in industrial inspection scenarios where acquiring extensive labeled datasets is challenging. While developed specifically for CFCPs' surface defect detection, the underlying methodology presents promising adaptability for broader defect-inspection applications across other composite materials and advanced manufacturing contexts.

Despite these advances, several limitations remain. A discrepancy exists between the domain of widely available general-purpose datasets and that of specialized DR images used in industrial defect inspection. Addressing this challenge through sophisticated domain adaptation and targeted data augmentation strategies constitutes an important direction for future research. Due to imaging constraints, our current dataset lacks explicit spatial reference information, preventing precise quantification of defect dimensions. Subsequent work will focus on refining experimental procedures by embedding appropriate spatial scales, enabling more accurate defect size evaluations, and enhancing practical applicability.

References

Publishing order | Descend order by publishing year | Descend order by cited within

[1]	Fu Y, Zhang Y, Chen H, Han L, Yin X, Fu Q, et al. Ultra-high temperature performance of carbon ﬁber composite reinforced by HfC nanowires: a promising lightweight composites for aerospace engineering. Compos Part B Eng 2023; 250(1):110453.

[2]	Waller MD, Mcintyre SM, Koudela KL. Composite materials for hybrid aerospace gears. J Am Helicopter Soc 2020; 65(4):1-11.

[3]	Wagner T, Heimbs S, Franke F, Burger U, Middendorf P. Experimental and numerical assessment of aerospace grade composites based on high-velocity impact experiments. Compos Struct 2018; 204(15):142-52.

[4]	Liang J, Liu L, Qin Z, Zhao X, Li Z, Emmanuel U, et al. Experimental study of curing temperature effect on mechanical performance of carbon ﬁber composites with application to ﬁlament winding pressure vessel design. Polymers 2023; 15(4):982.

[5]	Cieciela˛K KK, Zaleski K. Effect of depth surface defects in carbon ﬁbre reinforced composite material on the selected recurrence quantiﬁcations. Adv Mater Sci 2020; 20(2):71-80.

[6]	Naranjo-Lozada J, Ahuett-Garza H, Orta-Castañón P, Verbeeten WMH, Sáiz-González D. Tensile properties and failure behavior of chopped and continuous carbon ﬁber composites produced by additive manufacturing. Addit Manuf 2019; 26:227-41.

[7]	Wang Q, Zhou H, Liu M, Li X, Hu Q. Study of the skin depth and defect detection in carbon ﬁber composites with terahertz waves. Optik 2019; 178:1035-44.

[8]	Zhao M, Meng L, Ma L, Ma L, Yang X, Huang Y, et al. Layer-by-layer grafting CNTs onto carbon ﬁbers surface for enhancing the interfacial properties of epoxy resin composites. Compos Sci Technol 2018; 154(18):28-36.

[9]	Yao SS, Jin FL, Rhee KY, Hui D, Park SJ. Recent advances in carbon-ﬁber-reinforced thermoplastic composites: a review. Compos Part B Eng 2018; 142 (1):241-50.

[10]	Wu J, Zhou D, Wang J. Surface crack detection for carbon ﬁber reinforced plastic materials using pulsed eddy current based on rectangular differential probe. J Sens 2014; 2014:727269.

[11]	Qiu B, Li M, Zhang X, Chen Y, Zhou S, Liang M, et al. Carboxymethyl cellulose sizing repairs carbon ﬁber surface defects in epoxy composites. Mater Chem Phys 2021; 258:123677.

[12]	Chen Y, Ye L.Path-dependent progressive failure analysis for 3D-printed continuous carbon ﬁbre reinforced composites. Chin J Mech Eng 2024; 37 (1):72.

[13]	Linke M, Chakraborty S, Göbel H, Lammering R. On the accuracy of monitoring inter-ﬁbre fractures in composites using passive infrared thermography extended by microscopic analysis. Nondestruct Test Eval 2021; 36(6):654-66.

[14]	Cui JC, Liu W, Zhang Y, Han L, Yin P, Li Y, et al. A visual inspection method for delamination extraction and quantiﬁcation of carbon ﬁber reinforced plastic (CFRP). Measurement 2022; 196:111252.

[15]	Grimberg R, Savin A, Steigmann R, Bruma A. Eddy current examination of carbon ﬁbres in carbon-epoxy composites and Kevlar. Int J Mater Prod Technol 2006; 27:221-8.

[16]	Yin W, Withers PJ, Sharma U, Peyton AJ. Noncontact characterization of carbon-ﬁber-reinforced plastics using multifrequency eddy current sensors. IEEE Trans Instrum Meas 2009; 58(3):738-43.

[17]	Wang B, Shi W, Zhao B, Tan J. A modal decomposition imaging algorithm for ultrasonic detection of delamination defects in carbon ﬁber composite plates using air-coupled Lamb wave. Measurement 2022; 195:111165.

[18]	Wong V, Hu Y, Tham ZW, Chen YF, Liu M, Lim KE, et al. Measurement of elastic constant matrix of carbon ﬁber composites with an ultrasonic 2D-array transducer. IEEE Sens J 2022; 22(6):5562-70.

[19]	Caminero MA, García-Moreno I, Rodríguez GP, Chacón JM. Internal damage evaluation of composite structures using phased array ultrasonic technique: impact damage assessment in CFRP and 3D printed reinforced composites. Compos Part B Eng 2019; 165:131-42.

[20]	Meola C, Boccardi S, Carlomagno GM, Boffa ND, Monaco E, Ricci F. Nondestructive evaluation of carbon ﬁbre reinforced composites with infrared thermography and ultrasonics. Compos Struct 2015; 134:845-53.

[21]	Stone DEW, Clarke B. Ultrasonic attenuation as a measure of void content in carbon-ﬁbre reinforced plastics. Non Destr Test 1975; 8(3):137-45.

[22]	Avdelidis NP, Almond DP, Dobbinson A, Hawtin BC, Ibarra-Castanedo C, Maldague X. Aircraft composites assessment by means of transient thermal NDT. Prog Aerosp Sci 2004; 40(3):143-62.

[23]	Azizinasab B, Hasanzadeh RP, Hedayatrasa S, Kersemans M. Defect detection and depth estimation in CFRP through phase of transient response of flash thermography. IEEE Trans Industr Inform 2022; 18(4):2364-73.

[24]	Dong Y, Zhao B, Yang J, Cao Y, Cao Y. Two-stage convolutional neural network for joint removal of sensor noise and background interference in lock-in thermography. NDT & E Int 2023; 137:102816.

[25]	Yang R, He Y, Mandelis A, Wang N, Wu X, Huang S. Induction infrared thermography and thermal-wave-radar analysis for imaging inspection and diagnosis of blade composites. IEEE Trans Industr Inform 2018; 14 (12):5637-47.

[26]	Zou X, Wang L, Wang J, Liu J, Ma H, Bao Y. Nondestructive evaluation of carbon ﬁber reinforced polymer (CFRP)-steel interfacial debonding using eddy current thermography. Compos Struct 2022; 284:115133.

[27]	Nikishkov G, Nikishkov Y, Makeev A. Finite element mesh generation for composites with ply waviness based on X-ray computed tomography. Adv Eng Softw 2013; 58:35-44.

[28]	Nikishkov Y, Airoldi L, Makeev A. Measurement of voids in composites by X-ray computed tomography. Compos Sci Technol 2013; 89:89-97.

[29]	Rique AM, Machado AC, Oliveira DF, Lopes RT, Lima I. X-ray imaging inspection of ﬁberglass reinforced by epoxy composite. Nucl Instrum Methods Phys Res Sect B 2015; 349:184-91.

[30]	Gao Y, Song G, Li S, Zhen F, Chen D, Song A. LineSpyX: a power line inspection robot based on digital radiography. IEEE Robot Autom Lett 2020; 5 (3):4759-65.

[31]	Zhou F, Li X, Li Z. High-frequency details enhancing densenet for super-resolution. Neurocomputing 2018; 290(17):34-42.

[32]	Dong C, Loy CC, He K, Tang X. Image super-resolution using deep convolutional networks. IEEE Trans Pattern Anal Mach Intell 2016; 38(2):295-307.

[33]	Ijjeh AA, Ullah S, Radzienski M, Kudela P. Deep learning super-resolution for the reconstruction of full waveﬁeld of Lamb waves. Mech Syst Signal Process 2023; 186:109878.

[34]	Wu X, Zhou B, Huang F, Lin P, Cao R. Super-resolution thermal imaging using uncooled infrared sensors for non-destructive testing of adhesively bonded joints. IEEE Sens J 2022; 22(14):14415-23.

[35]	Cao W, Liu Q, He Z. Review of pavement defect detection methods. IEEE Access 2020; 8:14531-44.

[36]	Zhang Q, Barri K, Babanajad SK, Alavi AH. Real-time detection of cracks on concrete bridge decks using deep learning in the frequency domain. Engineering 2020; 7(12):1786-96.

[37]	Guo Y, Liu Y, Georgiou T, Lew MS. A review of semantic segmentation using deep neural networks. Int J Multimed Inf Retr 2018; 7(2):87-93.

[38]	Jiang B, An X, Xu S, Chen Z. Intelligent image semantic segmentation: a review through deep learning techniques for remote sensing image analysis. J Indian Soc Remote Sens 2023; 51(9):1865-78.

[39]	Huang H, Li Q, Zhang D. Deep learning based image recognition for crack and leakage defects of metro shield tunnel. Tunn Undergr Space Technol 2018; 77:166-76.

[40]	Feng H, Song K, Cui W, Zhang Y, Yan Y. Cross position aggregation network for few-shot strip steel surface defect segmentation. IEEE Trans Instrum Meas 2023; 72:1-10.

[41]	Li Z, Zhu H, Huang M. A deep learning-based ﬁne crack segmentation network on full-scale steel bridge images with complicated backgrounds. IEEE Access 2021; 9:114989-97.

[42]	Liu J, Wang C, Su H, Du B, Tao D. Multistage GAN for fabric defect detection. IEEE Trans Image Process 2020; 29:3388-400.

[43]	Liu Z, Cao Y, Wang Y, Wang W. Computer vision-based concrete crack detection using U-Net fully convolutional networks. Autom Construct 2019; 104:129-39.

[44]	Falk T, Mai D, Bensch R, Çiçek Ö, Abdulkadir A, Marrakchi Y, et al. U-Net: deep learning for cell counting, detection, and morphometry. Nat Methods 2019; 16 (1):67-70.

[45]	He Y, Mu X, Wu J, Ma Y, Yang R, Zhang H, et al. Intelligent detection algorithm based on 2D/3D-UNet for internal defects of carbon ﬁber composites. Nondestruct Test Eval 2024; 39(4):923-38.

[46]	Liu X, Zhang Y, Jing H, Wang L, Zhao S. Ore image segmentation method using U-Net and Res_Unet convolutional networks. RSC Adv 2020; 10(16):9396-406.

[47]	Wang Z, Yang P, Liang H, Zheng C, Yin J, Tian Y, et al. Semantic segmentation and analysis on sensitive parameters of forest ﬁre smoke using smoke-Unet and landsat-8 imagery. Remote Sens 2021; 14(1):45.

[48]	Wang Y, Chen Q, Luo Q, Li Q, Sun G. Characterizing damage evolution in ﬁber reinforced composites using in-situ X-ray computed tomography, deep machine learning and digital volume correlation (DVC). Compos Sci Technol 2024; 254:110650.

[49]	Wang Y, Narayanan A, Wang D. On training targets for superv ised speech separation. IEEE/ACM Trans Audio Speech Lang Process 2014; 22(12):1849-58.

[50]	Uddin S, Khan A, Hossain ME, Moni MA. Comparing different supervised machine learning algorithms for disease prediction. BMC Med Inform Decis Mak 2019; 19(1):281.

[51]	Huang G, Song S, Gupta JN, Wu C. Semi-supervised and unsupervised extreme learning machines. IEEE Trans Cybern 2014; 44(12):2405-17.

[52]	Torralba A, Russell BC, Yuen J. LabelMe: online image annotation and applications. Proc IEEE 2010; 98(8):1467-84.

[53]	Russell BC, Torralba A, Murphy KP, Freeman WT. LabelMe: a database and web-based tool for image annotation. Int J Comput Vis 2008; 77(1-3):157-73.

[54]	Stanley KO, Clune J, Lehman J, Miikkulainen R. Designing neural networks through neuroevolution. Nat Mach Intell 2019; 1(1):24-35.

[55]	Zhao M, Zhong S, Fu X, Tang B, Pecht M. Deep residual shrinkage networks for fault diagnosis. IEEE Trans Industr Inform 2020; 16(7):4681-90.

[56]	Szegedy C, Ioffe S, Vanhoucke V, Alemi A. Inception-v4, inception-ResNet and the impact of residual connections on learning. In:Proceedings of the Thirty-First AAAI Conference on Artiﬁcial Intelligence; 2017 Feb 4-9; San Francisco, CA, USA. Palo Alto, CA: AAAI Press; 2017. p. 4278-84.

[57]	Mohammad-Alikhani A, Nahid-Mobarakeh B, Hsieh M. One-dimensional LSTM-regulated deep residual network for data-driven fault detection in electric machines. IEEE Trans Ind Electron 2024; 71(3):3083-92.

[58]

Ledig C, Theis L, Huszár F, Caballero J, Cunningham A, Acosta A, et al. Photo-realistic single image super-resolution using a generative adversarial network. In: Proceedings of the 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR); 2017 Jul 21-26; Honolulu, HI, USA. Piscataway: IEEE; 2017. p. 105-14.

[59]	Dubey SR, Singh SK, Chaudhuri BB. Activation functions in deep learning: a comprehensive survey and benchmark. Neurocomputing 2022; 503:92-108.

[60]	Hua Z, Yi S, Zhou Y. Medical image encryption using high-speed scrambling and pixel adaptive diffusion. Signal Process 2018; 144:134-44.

[61]	Tong YB, Zhang QS, Qi YP. Image quality assessing by combining PSNR with SSIM. J Image Graph 2006; 11(12):1758-63.

[62]	Gao L, Shen W, Li X. New trends in intelligent manufacturing. Engineering 2019; 5(4):619-20.

[63]	Gan Y. Research on the innovative development of new materials science and technology in China. Engineering 2024; 32:12-5.

[64]	Liu F, Wang L. UNet-based model for crack detection integrating visual explanations. Constr Build Mater 2022; 322(7):126265.

[65]	Guan S, Khan AA, Sikdar S, Chitnis PV.Fully dense UNet f or 2-D sparse photoacoustic tomography artifact removal. IEEE J Biomed Healt h Inform 2020; 24(2):568-76.