Review of Research on Ceramic Surface Defect Detection Based on Deep Learning

Abstract

Ceramic surfaces are directly related to product quality and safety in industry, and any minor defects may affect performance. Therefore, surface defect detection has important practical significance. Traditional detection methods have limitations, while deep learning methods bring new opportunities. Although there have been many studies on ceramic surface detection, most of them focus on traditional image processing methods or single-angle deep learning applications. This article proposes a detection scheme that combines multi-perspective image acquisition and improved deep learning models for complex environments in industrial production lines, with a particular focus on small-sample, imbalance, and small-target defects. In ceramic defect detection, defects are often diverse, small in size, and difficult to collect, which can lead to insufficient model training and low recognition accuracy when using deep learning methods for defect detection. In addition, industrial production requires the high real-time performance of detection systems, which must respond quickly while ensuring accuracy to meet efficient and stable quality control requirements. Therefore, data imbalance, small samples, small targets, and real-time issues are particularly critical in ceramic defect detection. This article first introduces the basic steps and current situation of data preparation. It then explores solutions to the imbalanced-sample problem in ceramic surface defect detection using methods such as data augmentation, sample distribution optimization, network structure improvement, and loss function design. Additionally, it reviews the small-sample problem in ceramic surface defect detection through approaches like data augmentation, transfer learning, unsupervised learning, and network structure optimization. This article also elaborates on methods to enhance the detection accuracy of small-target defects on ceramic surfaces, including adding attention mechanisms, improving features, and optimizing network structures. Finally, it discusses improvements in the real-time performance of model defect detection from two perspectives: enhancing lightweight models and integrating and optimizing network modules. This article summarizes solutions for implementing ceramic surface defect-detection technology and explores future research directions in this field.

1. Introduction

Ceramic is a solid material made from natural or synthetic inorganic non-metallic materials, usually formed through high-temperature sintering. It possesses excellent physical, chemical, and structural properties [1]. These properties enable its extensive application in fields such as construction, industry, and healthcare, with broad application prospects [2,3,4]. Ceramic products often exhibit surface defects during manufacturing, primarily caused by technical and operational constraints, which directly compromise product quality. Therefore, in order to ensure the qualification rate and reliable quality, it is necessary to conduct surface defect detection on ceramic products.

Defects in ceramic products can generally be understood as areas of deficiency, imperfection, or difference compared to normal samples. Common surface defects mainly include shrink glaze [5], spots [6], bubbles [7], cracks [8], pinholes [9], and scratches [10], as shown in Figure 1. Surface defect detection involves inspecting the surface of a sample to identify such defects. These defects not only affect the esthetics of ceramic products but may also impact their sealing performance and durability. Manual surface inspection methods employed by quality inspectors suffer from low efficiency, high labor intensity, low accuracy, and poor real-time performance, failing to meet the increasingly high-quality standards in the industrial manufacturing process [11].

Traditional ceramic surface defect-detection methods, such as threshold segmentation, edge detection, and template matching, are effective in some cases. However, they are sensitive to illumination, noise, and surface reflections, which affects detection accuracy. These methods rely on manually set rules or features, making it difficult to handle complex defects. Moreover, they lack automation and efficiency when dealing with diverse defects and are easily influenced by human intervention and subjective factors. In recent years, numerous scholars have started to use object detection methods in deep learning to detect ceramic surface defects. Deep learning methods can automatically learn and extract complex high-dimensional features through the training of large-scale datasets, effectively addressing various defect-detection problems. Deep learning models have strong generalization capabilities and can efficiently and accurately identify various defects on ceramic surfaces under complex backgrounds and different illumination conditions, reducing human intervention and improving detection accuracy and robustness [9,12,13].

Birlutiu [14] addressed the numerous problems associated with manual inspection during ceramic visual inspection by using deep learning methods for surface defect detection, improving detection efficiency and providing a more efficient and objective solution for quality control in the porcelain industry. Mariyadi [15] applied an artificial neural network with resilient backpropagation (RProp) to achieve the automated detection of multiple surface defects on ceramic tiles. Recently, Dang [16] proposed a railway damage detection method based on pencil lead fracture-triggered ultrasound and the adversarial autoencoder (AAE). The method uses PLB to generate broadband ultrasound waves as the excitation source, AAE to learn the signal characteristics of healthy tracks, and uses hidden variable space Jensen–Shannon divergence to construct unsupervised damage indicators, overcoming the high cost and operational complexity of traditional ultrasound detection and achieving fast and accurate damage diagnosis. He [17] proposed a label noise robust framework based on bounded neural networks (BNNs), which integrates the global features of sensor data through a multi-basis model with shared weights, and designed a constraint function that combines Box–Cox loss and an implicit weighting mechanism to limit the upper-bound loss of noise samples while maintaining cross-entropy learning efficiency. Building on these advances, this paper adopts deep learning-based object detection methods and models to perform more accurate and efficient ceramic defect detection based on their characteristics. It further analyzes key issues in the surface defect-detection process, including data preparation, imbalanced-sample detection, small-sample detection, small-target detection, and real-time detection methods, as shown in Figure 2.

2. Data Preparation

In the task of ceramic surface defect detection, data preparation is a crucial link for improving model performance. High-quality sample preprocessing can not only enhance the accuracy of the model but also improve its robustness, enabling it to perform more stably in practical applications.

Reference [18] emphasizes the importance of data preparation, pointing out that different preprocessing steps have significant differences in their applicability to various learning algorithms. Selecting appropriate preprocessing methods according to the specific task and algorithm characteristics is the core of data preparation. Reference [19] analyzed the relationship between volatility—defined here as the variability or diversity in data distribution and sample characteristics—and data characteristics through data balancing experiments. The study suggests that such volatility helps improve prediction performance and enhances the understanding of the effects of preprocessing. Reference [20] reviewed the key steps of data preprocessing and augmentation. Preprocessing includes data cleaning, noise processing, data integration, transformation, and reduction, aiming to improve data quality and diversity. Augmentation techniques are divided into two categories: data deformation based on geometric and color transformations, and oversampling based on methods such as Mixup and SMOTE, which are used to generate diverse samples and alleviate data imbalance. Meanwhile, techniques to prevent overfitting, such as transfer learning and Dropout, are also mentioned. These methods work together in model training to improve its generalization ability. After preprocessing operations, defects can be made more prominent, facilitating better identification by the model and significantly enhancing the accuracy and robustness of the ceramic defect-detection model. Similarly, data augmentation techniques in ceramic image processing are not only used to expand datasets but can also be customized for challenges such as complex surface textures and strong reflectivity. Geometric and color transformations help enhance the model’s adaptability to changes in position and lighting, while methods such as Mixup and SMOTE can alleviate sample imbalance, thereby improving the model’s robustness and detection accuracy in complex scenes. In addition, Tian [21] used laser radial scanning combined with the Hough transform linear detection algorithm to locate stripe positions, identified stripe fracture features through connected domain analysis, quantified texture spacing and color depth differences, and achieved a multidimensional analysis of three-dimensional texture states. This is also a new method for preparing data in the process of ceramic defect detection.

Currently, open-source datasets for ceramic surface defects are scarce and lack unified standardization. Many datasets are customized for specific research or enterprises and have not formed widely recognized standards. Creating standardized datasets related to ceramics is highly challenging due to the significant differences in ceramic types and processes, the complex and diverse forms of appearance and defect expression, the subjective definition of defects, the scarcity and imbalance of real defect samples, and the difficulty in unifying collection and labeling standards. These factors can seriously affect the consistency and universality of the dataset. With the increasing demand for intelligence, there is hope for the establishment of a unified and standardized open-source dataset in the field of ceramic surface defect detection.

3. Detection Methods for Imbalanced Samples in Ceramic Surface Defects

The problem of imbalanced samples is common in ceramic defect datasets, as the frequency of defects such as adhesion, foreign objects, and black spots is much lower than that of normal or slightly defective samples, resulting in significant differences in the number of different defect categories in the data. This can easily cause the model to lean toward the majority class and ignore key defect categories. In response to this issue, researchers have continuously explored optimization strategies to effectively alleviate the impact of class imbalance and improve the model’s adaptability and detection ability in complex scenarios. Reference [22] introduced methods for dealing with class imbalance problems in deep learning, which can be roughly divided into three categories: preprocessing, special-purpose algorithms, and postprocessing. Preprocessing methods include random oversampling, data generation by GANs, and the combination of sampling and transfer learning. Special-purpose algorithms cover weighted cross-entropy, integrated cost matrices, and the methods of learning the classifier after representation learning. Postprocessing methods mainly involve changing the threshold based on prior class probabilities. In the field of ceramic surface defect detection, researchers have combined these strategies and proposed various methods from aspects such as data augmentation, sample distribution optimization, network structure improvement, and loss function design, significantly enhancing the model’s ability to identify minority defect classes and its overall detection performance.

Huang [23] proposed a K-means clustering balance strategy for the problem of inter-class and intra-class imbalance in ceramic substrate defect-detection datasets. It improves the model’s ability to identify minority defects by performing oversampling and undersampling operations more reasonably when the sample distribution is uneven, as shown in Figure 3. This strategy first uses the K-means clustering algorithm for the unsupervised clustering of samples, calculates the mean value of each class based on the clustering results, and sets a balance target value. Then, it calculates the difference between each subclass and the balance target value. During the training process, the speed of sample increase or decrease in each training round is controlled by hyperparameters, thereby performing oversampling on minority classes or undersampling on majority classes to gradually balance the model during training. Wu [24] addressed the problem of imbalanced-sample identification in daily use ceramic cup surface defect detection by using data augmentation methods such as flipping, clockwise rotation, saturation change, random cropping, and adding noise to expand the dataset information, increase the number of different types of samples, enable the model to learn more feature information, and improve the identification ability of minority class samples. He [25] aimed to solve the problems of low contrast, large intra-class differences, and high inter-class similarity in the surface defect detection of ceramic substrates. By designing a global information aware GIGP and contextual semantic extraction CSFE, gradually mining deep discriminative features, and combining the MSIG-FPN module, including HRR refocusing task-related features and CSDF cross-stage semantic fusion to achieve multi-path semantic interaction, the feature-fusion effect is effectively improved.

In addition, researchers have proposed a variety of effective methods from the perspectives of network structure improvement and loss function optimization. Carvalho [26] addressed the imbalanced sample problem in ceramic tableware detection by combining similar defects into broader categories to reduce the number of classes and using a weighted loss function to assign different weights to different classes, alleviating the class imbalance problem to a certain extent and improving the classification model’s ability to identify minority defect classes. Hang [5] used an improved focal loss to calculate the heatmap loss during RPN training for the sample imbalance problem in sanitary ceramic surface defect detection. By setting balance factors, the losses of difficult and easy samples, as well as positive and negative samples, were better balanced, enabling the network to pay more attention to difficult samples and positive samples. Ye [27] proposed a channel and spatial joint attention method (CSAM) for the problem of texture background interference in complex texture tile defect detection. CSAM uses a local convolution with a kernel size of 3 to model the local channel dependencies, solving the interference of the texture background on defect detection to a certain extent, enabling the model to better focus on defect features, and indirectly facilitating the detection of different types of defects. To a certain extent, the problem of differences in the detection performance of some defect types caused by sample imbalance was solved. Cao [28] designed a network loss function composed of bounding box regression loss, confidence loss, and classification loss for the imbalanced sample problem in tile surface defect detection. The bounding box regression loss adds a modulation term obtained by calculating the Wasserstein distance of the Gaussian model between the predicted box and the true box based on the CIoU loss function. The confidence loss and classification loss use the binary cross-entropy loss function, balancing the score differences in the targets of different scales and classes and enabling the model to treat different types of defect samples more equally during the learning process.

In summary, through data augmentation, sample distribution optimization, network structure improvement, and loss function design, the problem of sample imbalance in ceramic defect detection has been effectively alleviated, and the detection accuracy and generalization ability of the model have been improved. Among them, data augmentation is simple and efficient but may introduce noise, sample distribution optimization is suitable for extreme imbalance but complex processing, structural improvement enhances capability but has a high computational cost, and loss function design strikes a balance between efficiency and performance. These methods can be further optimized to achieve better results in ceramic defect detection.

4. Detection Methods for Small Samples in Ceramic Surface Defects

In the field of ceramic surface defect detection, due to the difficulty of data acquisition during the industrial production process, the problem of small samples restricting the model’s ability to learn complex features often occurs. On the one hand, some defects are extremely rare in production, making it difficult to collect sufficient samples; on the other hand, defect labeling often relies on experienced manual quality inspection experts, resulting in high labeling costs and low efficiency, further limiting the acquisition of high-quality defect data. However, deep learning models can effectively learn and train with limited samples through techniques such as data augmentation and transfer learning, achieving the accurate detection of ceramic surface defects. Reference [29] proposed a series of solutions for the problem of data scarcity in deep learning, including data augmentation, generative models, transfer learning, semi-supervised learning, and data simulation techniques. These methods reduce the dependence on a large amount of labeled data and improve the detection accuracy and generalization ability.

4.1. Data Augmentation Methods

Currently, researchers have proposed a variety of data augmentation methods, from traditional data augmentation techniques to the generation of synthetic samples by generative adversarial networks (GANs) and then to hybrid strategies combined with transfer learning. These methods have significantly improved the richness of the dataset and the generalization ability of the model. They have effectively alleviated the small-sample problem, providing more reliable data support and model performance guarantee for ceramic defect detection. Cumbajin [30] addressed the problem of insufficient samples during ceramic part manufacturing by using data augmentation techniques such as rotation and flipping to expand the dataset. At the same time, through transfer learning, they initialized the network with the weights of a model pre-trained on other large-scale data and then fine-tuned it, enabling the model to converge quickly on limited ceramic part defect samples and achieve good performance, reducing the dependence on a large number of samples. Niu [31] addressed the problem of insufficient samples in sanitary ceramic defect detection by using four offline data augmentation methods, image generation, image mosaicking, image fusion, and image rotation mosaicking, to increase the number of training samples, alleviating the small-sample problem to a certain extent from the data level. Chu [32] addresses the issue of limited coating defect samples by implementing five data augmentation strategies, including rotation, translation, brightness variation, mirroring, and copy–paste, to increase the sample size, enrich the image background, and enhance the robustness of model training. Wang [33], when facing the detection of tile surface defects, expanded the capacity of the training dataset through operations such as translation, scaling, and rotation in response to the limited capacity of the original dataset and the possible overfitting problem. At the same time, to eliminate the impact of background defects on the experimental results, they performed an opening operation on the image, first eroding to remove noise and then dilating to remove background plate defects. By comparing the processing of images with different-sized convolutional kernels, they found that an 80 × 80 convolutional kernel had a better processing effect, thereby improving the model’s generalization ability and better detecting tile defects. After processing, 7200 images were obtained. Wang [34] expanded the dataset of ceramic ring defects from 780 to 1560 through methods such as scaling and translation when faced with defect detection in ceramic rings. This increased the number of samples to a certain extent, alleviated the problem of small samples, and improved the adverse effects of small samples on model training to a certain extent, thereby enhancing the efficiency of ceramic surface defect detection. Tang [35] addressed the small-sample problem in ceramic filter surface defect detection by using the pix2pixHD generative adversarial network model to expand the ceramic filter surface defect samples. By drawing “fake” defect labels on normal samples and using the generative adversarial network to generate samples similar to real defects, the problem of insufficient defect sample numbers was solved, providing sufficient sample data for the training of deep learning models.

Among the above methods, traditional data augmentation techniques such as rotation and scaling have played a certain role in increasing sample size and alleviating overfitting, but their generated sample variations are limited, and it is difficult for them to fully simulate the true features of complex defects. In contrast, GAN can generate defect images with more visual realism, effectively expanding data diversity and distribution range, thereby significantly improving model performance under small-sample conditions. However, the quality of GAN-generated samples highly depends on training quality and generation strategy. If the generation is inaccurate, it may introduce false features, affecting the model’s generalization ability.

4.2. Methods Based on Transfer Learning, Unsupervised Learning, and Network Structure Optimization

Transfer learning, unsupervised learning, and network structure optimization techniques are also effective means to alleviate the small-sample problem. By using models pre-trained on large-scale datasets or adopting unsupervised learning methods, better fitting and generalization abilities can be obtained with limited samples. At the same time, optimizing the network structure can improve the stability of training and the accuracy of detection.

Zhang [36] addressed the small-sample problem of ceramic surface defects by using the method of a pre-trained model, using the weights trained on the COCO dataset as the initial parameters, alleviating the impact of small-sample data on model training and improving the model’s generalization ability. Jia [7], in the defect detection of the ceramic additive manufacturing process with a small-scale dataset, designed a dual-branch structure, separating the serial branch feature-extraction network from the parallel branch attention weight calculation network, enabling the model to focus more precisely on specific regions of the feature map, reducing the impact of outliers on the training results, and improving the stability of the model’s training on small-sample datasets. Li [37], for the small-sample problem of texture tile defect detection, based on the idea of transfer learning, initialized the corresponding weights of the improved model with the shallow backbone weights of the original YOLOv3 model pre-trained on the COCO dataset. They set up two groups of experiments for comparison, one with randomly initialized weights and the other with pre-trained weights. The results showed that importing pre-trained weights accelerated the model convergence and improved the network’s fitting ability for small-sample datasets. Li [38], for the problem of insufficient tile surface defect samples, adopted an unsupervised learning method based on convolutional autoencoders, only requiring a large number of unlabeled normal tile samples for training, avoiding the dependence on a large number of manually labeled defect samples and alleviating the small-sample problem to a certain extent. From this, it can be concluded that unsupervised methods avoid dependence on any labeled data, but they may have difficulty in handling complex defect patterns compared with fine-tuned pre-trained models.

By applying a variety of data augmentation techniques, the small-sample problem in ceramic surface defect detection has been effectively alleviated, and the robustness of the model has been improved. However, data augmentation and generative models still face challenges such as insufficient sample quality and diversity, which affect detection performance. Transfer learning, unsupervised learning, and network structure optimization techniques provide effective paths to solve the small-sample problem, but the limitations of relying on pre-trained datasets and the insufficient effectiveness of unsupervised learning in complex scenarios still need to be overcome. Focusing on improving the diversity and authenticity of generated samples and optimizing the application of unsupervised learning and transfer learning can further enhance the model’s performance and adaptability.

5. Detection Methods for Small Targets in Ceramic Surface Defects

Traditional detection methods struggle to handle complex issues, such as the weak features of small targets and vulnerability to background interference, resulting in detection accuracy that fails to meet high-standard requirements. In recent years, intelligent detection technologies based on deep learning have witnessed rapid development. Through innovations in network structures, feature processing, attention mechanisms, as well as loss functions and training strategies, the detection performance has been significantly enhanced [39]. Meanwhile, the comprehensive application of these technological innovations provides an efficient solution for detecting surface defects in ceramic-like products, promotes the in-depth development of intelligent detection technologies, and injects new impetus into improving industrial production efficiency and product quality.

5.1. Network Structure Optimization

To address the challenging problem of detecting small-target defects on the surfaces of ceramic-like products, different researchers have optimized the network structure. This includes methods such as enhancing feature-extraction capabilities, improving activation functions, and optimizing anchor box designs, which have alleviated the difficulties in detecting small-target defects on ceramic surfaces.

Li [37] added a convolutional autoencoder composed of feature encoding and decoding modules to the front-end of the Darknet-53 in the YOLOv3 network. Additionally, they optimized the anchor boxes by using the K-means clustering algorithm to classify the width and height of the defects. This improvement increased the model’s average accuracy by 5%, enabling it to more effectively detect tiny defects, such as holes and scratches on textured tiles. Teng [40] increased the number of layers in the backbone network and the number and scale of feature maps based on the YOLOv3 network. They also employed up-sampling to fill in the feature maps, increased the prior boxes of the sanitary ceramic defect dataset from 9 to 5, and changed the Leaky ReLU activation function to the PReLU activation function, improving the detection ability for small-target defects on the surfaces of sanitary ceramics.

Chen [41] added SK modules after the three effective feature layers from the backbone to the neck in the YOLOv5 target detection network. These modules are used to dynamically select features with different receptive fields, enabling the model to better adapt to and detect small-target defects. Lei [42] replaced the spatial pyramid pooling module in the original network with a dilated spatial convolution pooling pyramid module in the backbone network of YOLOv5 and added an ASFF mechanism after the neck network, improving the detection ability for small-target defects on the surfaces of flat ceramic membranes. Ding [43] introduced a small-target-detection layer after the CoordAtt and CSP2_1 modules based on the YOLOv5 network. They increased the 160 × 160 feature map, as well as the number of image cells and anchors, enhancing the detection ability for defects such as cracks and dirt spots on the surfaces of ceramic double-layer cups. Zhang [36] added a small-target-detection layer to the YOLOv5 network and removed the large-target-detection layer in the original network structure. Cao [28] introduced the CARAFE module into the backbone network of the YOLOv5l model to address the problem of insufficient small-target-detection accuracy in tile surface defect detection. The CARAFE module compresses the channels of the feature map to focus on useful information and improve computational efficiency. Then, it enters the encoding module to construct a recombination kernel operator. Through pixel-shuffling operations, a high-resolution feature map is obtained. Finally, the softmax activation function is applied to normalize and activate the recombination kernel weights, and a high-resolution feature map is generated by weighted summation, thereby enhancing the detection accuracy of small-target defects. In summary, CARAFE can up-sample low-resolution feature maps through content-aware methods, generating higher resolution and clearer structured feature maps. Compared to traditional up-sampling methods, CARAFE considers a wider range of contextual information, making the edge and texture information of small targets more complete, which is beneficial for the model to more accurately identify small targets. Pan [6] obtained the position and size information of objects from the dataset, determined the optimal anchor box size and number using the K-means clustering algorithm, optimized the anchor boxes with a genetic algorithm, and applied the optimized anchor boxes to the training and inference of YOLOv5, improving the problem of the difficult detection of small defects in the detection of ceramic disk surface defects.

Wu [24] added deformable convolutions to the backbone part of YOLOv8 to address the problem of insufficient small-target-detection accuracy in the detection of daily use ceramic cup surface defects. They replaced the SPPF module with the RFB module to fuse multi-scale context information and proposed the DWC2f module based on depth-separable convolutions and BottleNeck to replace the C2f module in the original network. As a result, the detection accuracy increased from 77.3% to 85.1%, improving the detection accuracy for small-target defects and the detection accuracy of daily use ceramic cup surface defects. Jia [7] utilized the differential Siamese network with differential information and multi-prediction head design. The structure of the Siamese network is shown in Figure 4, which effectively improved the detection accuracy of small-target defects such as shortages and collapses during the re-coating process of ceramic additive manufacturing. In addition, in order to address the challenges faced by multi-scale and small-sized defect detection on ceramic tableware surfaces, Sun [44] innovatively designed a new type of convolution module that combines non-step convolution and spatial depth conversion to address the problems of fine-grained information loss and insufficient feature extraction in traditional convolution. By preserving feature map resolution and reassembling spatial information, it enhances the detection ability of small targets.

5.2. Feature Processing Improvement

Improvements in feature processing, such as up-sampling, feature fusion, and the design of multi-scale detection heads, can enhance the feature expression ability and provide a more efficient solution for the precise positioning and identification of small-target defects.

Reference [45] fused the features extracted from the Conv3_3, Conv4_3, and Conv5_3 convolutional layers in the VGG-16 network. The Conv3_3 feature map was down-sampled by max-pooling, and the Conv5_3 feature map was up-sampled. After making the sizes of the former two consistent with that of Conv4_3, they were added together. The fused feature map had visible object contours and rich information, enhancing the detection ability for small targets. Based on the YOLOv5 network structure, Ding [43] added an up-sampling layer in the neck part of the model and performed feature-layer fusion with the CSP1_1 module of the backbone network, integrating high-level semantic information and low-level detail features and enhancing the model’s feature expression ability for small targets on the ceramic surface. Wang [34] replaced the nearest-neighbor up-sampling operator with the CARAFE operator to increase the receptive field and added a new feature-fusion layer. This layer was output after 4 times of down-sampling by the backbone network and then fused with the 8 times down-sampled feature map, finally generating a 160 × 160-pixel feature map, which improved the model’s detection ability for small-target defects in ceramic rings. Zhu [46] introduced the LSKNet module (Large Selection Kernel Network) into the YOLOv8 network model. It used deep convolutional kernels and a spatial selection mechanism to dynamically adjust weights and receptive fields, enhancing the model’s feature-extraction ability for multi-scale targets. In the detection head, they removed the large-scale detection head and added a small-scale detection head, enhancing the network model’s ability to identify tiny defects on the tile surface. Ye [27] proposed a selective feature-fusion method based on the PVTv2 network in the Transformer. By combining shallow spatial features and deep semantic features from different layers, this method generates a fused feature map, effectively improving the detection of small defects, such as white spots and black spots on ceramic surfaces.

Han [47] enhanced the detection accuracy of small targets in the ResNeXt-SSD network through the feature-fusion module and the utilization of shallow-layer features, improving the detection effect for small-target defects such as pores and karst caves. Chen [48] proposed the Inception–SSD network model based on the SSD network model. It used a multi-branch parallel feature-extraction method with different-sized convolutional kernels, and the Inception structure that fuses all the features replaced the four convolutional layers of VGG-16 in the SSD network for feature extraction, enabling the model to extract more feature information. Compared with the SSD network model, the Inception-SSD network model increased the mean average precision (mAP) for the detection of small-target defects such as cracks and protrusions by 13.83% and 15.22%, respectively, effectively enhancing the detection accuracy of small defects on the ceramic surface. Dong [49] proposed a new method for detecting the surface defects of complex textured ceramic tiles based on a high light-reflection background. In response to the problem of texture interference in traditional algorithms and convolutional neural networks, the RANSAC algorithm was used for feature corner matching and rigid stitching to construct a complete surface image with high light areas as the background. Threshold segmentation and morphological filtering were combined to achieve defect extraction, and the detection effect of small-target defects on ceramic surfaces was significant.

Wang [33] also improved the Faster R-CNN network by introducing the Adaptive Spatial Feature Fusion (ASFF) algorithm. This algorithm added weight coefficients to multiple different feature maps as inputs, achieving the effective fusion of different detection results. It retained the semantic information of high-level features and took into account the details of low-level features, enabling the more precise detection of tile defects. Man [50] improved the Faster R-CNN network by using the ResNet-50 residual network and the feature pyramid network (FPN). This enhanced the feature-extraction and detection ability for small-area defects such as pinholes in metallized ceramic rings, significantly improving the detection accuracy and positioning accuracy. Compared with the original VGG-16 structure, the mAP increased by 17.5%. In addition, Liu [51] proposes to enhance the model’s perception ability of small defects on the surface of ceramic tiles by deepening the high-resolution feature-fusion layer of the feature pyramid network (FPN) and the path aggregation network (PAN).

Li [38] proposed a lightweight convolutional autoencoding image reconstruction network LR-CAE. They fused the corresponding convolutional-layer feature maps and de-convolutional-layer feature maps, integrating high-level semantic information and low-level texture information, improving the reconstruction ability of the traditional convolutional autoencoding structure (Convolutional Attention Encoder, CAE) for high-resolution content. This enabled the network to better reconstruct image details and was helpful for the more accurate detection of tile surface defects. Jia [7] effectively improved the detection accuracy of surface defects such as bubbles, depressions, and scratches during the ceramic additive manufacturing process through multi-image fusion and multi-scale feature-fusion methods. Hang [5] constructed a five-layer feature pyramid based on the MobileNetV3 backbone network, enabling the network to detect targets on feature maps of different scales and having a certain adaptability to the detection of small-target defects on the surfaces of sanitary ceramics.

5.3. Attention-Mechanism Improvement

The application of attention mechanisms in the field of small-target detection has always been a research hotspot. Reference [52] systematically reviewed the development of attention mechanisms and affirmed their role in enhancing the model’s attention to important information and suppressing irrelevant or redundant features. In the field of ceramic surface defect detection, optimizing the attention-mechanism modules of deep-learning models from multiple perspectives can improve the defect-detection performance and enhance the recognition ability of small-target defects on ceramic surfaces.

Based on the YOLOv5 network, both Ding [43] and Zhang [36] introduced the CA mechanism after the Concat operation of the third up-sampled feature and the CSP1_1 feature, better meeting the requirements of ceramic quality inspection. Chen [41] introduced the efficient channel attention (ECA) in the three branches of the detection part (Detect), making the features extracted by the model more precise and the detection of small defects more accurate. Lei [42] added the Coordinate Attention (CA) mechanism to the backbone network, more accurately detecting the small defects of flat ceramic membranes. Tang [35] introduced the CBAM attention mechanism into the backbone network of the YOLOv5 model, to a certain extent solving the problem of high false-detection rates caused by small defects on the surfaces of ceramic filters. Guan [8] introduced the CBAM attention-mechanism module after the first convolutional layer of the backbone network, solving the problem of small defect sizes in ceramic rings. Wang [34] introduced the CBAM attention-mechanism module after the last C3 layer of the backbone network, improving the quality inspection efficiency and accuracy during the production of ceramic rings. Pan [53] introduced the Global Attention Mechanism (GAM) into the backbone network and neck of YOLOv5 to enhance feature extraction, enabling the model to pay more attention to important features and improving the problems of the false detection of small shadows and the missed detection of obvious defects. Pan [6] added the ECA module to the feature-extraction module of the YOLOv5s network structure, better detecting the small defects on the surfaces of ceramic disks.

Based on the Transformer network, An [54] proposed inserting a CBAM module composed of a channel attention module (CAM) and a spatial attention module (SAM) at the front-end of the network. They also improved the self-attention module to an L2 multi-head self-attention module, which is of great significance for improving the accuracy of the quality inspection of ceramic-bearing products.

Wang [33] proposed adding a convolutional block attention hybrid module (CBAMM) composed of a channel attention module, a spatial attention module, and a class attention module (class attention, CA) connected in parallel to the Faster R-CNN network. This made the network pay more attention to the content, position, and class information of the target object, improving the detection accuracy. Tang [55] selected a convolutional neural network based on ResNet34 and inserted ECA modules at different stages of ResNet34 and between residual modules. That is, attention-mechanism modules were added between the two convolutional layers of the Basicblock residual module and between different stages, enabling the model to better adapt to the detection task of ceramic hole-filling defects. Chen [56] added an ECA mechanism module after each effective feature layer in the improved network model. By calculating the adaptive convolutional-kernel size of each channel, performing a 1D convolutional operation on each channel to obtain weights, and multiplying the weights by the original input feature layer, the convolutional neural network was enabled to actively focus on the feature points with important information, effectively improving the model’s detection accuracy for small targets.

In summary, references [36,42,43] used the CA mechanism, references [6,41,55,56] used the ECA mechanism, references [8,33,34,35,54] used the CBAM attention mechanism, and reference [53] used the GAM. Among them, CA focuses on combining channel and location information modeling to improve the model’s perception ability of target location; ECA emphasizes achieving efficient channel attention extraction at low computational costs; CBAM focuses on channel and spatial dimensions to improve the accuracy of feature expression; the GAM integrates multidimensional spatial and channel information to enhance overall semantic modeling capabilities. They optimize the model’s ability to focus on key features at different levels.

5.4. Loss Function and Training Strategy Optimization

To improve the detection accuracy of small-target defects on the surfaces of ceramic-like products, scholars have also conducted numerous studies on loss functions, activation functions, and training strategies. These improvement methods focus on enhancing the model’s ability to pay attention to tiny targets and effectively improve the detection accuracy and robustness.

Li [37] used a structural similarity measurement function to construct a new loss to replace the mean-square error loss during the training of the CAE module in the YOLOv3 network. This new loss comprehensively considered image brightness, contrast, and structural similarity and could effectively detect holes and scratches on textured tiles. Pan [53] improved the bounding box regression loss function with α-IoU in YOLOv5, making the model pay more attention to small targets during training. Reference [54] proposed a two-stage training strategy based on the Transformer, improving the image super-resolution reconstruction accuracy and enhancing the defect-detection accuracy.

Wu [24] calculated the classification loss using VFL Loss and calculated the bounding box regression loss by combining CIoU Loss and DFL Loss. By adjusting the three weight parameters, λ₁, λ₂, and λ_3, to balance the losses, the detection accuracy was finally increased from 77.3% to 85.1%, improving the detection accuracy for small targets. Zhu [46] replaced the original IoU loss function with the DIoU loss function, which considered the normalized distance between the centers of the predicted box and the target box, enhancing the ability to identify tiny defects on the tile surface. Li [38] used the ReLU activation function in each network layer and adopted BN (Batch Normalization) and Dropout techniques to prevent overfitting. To avoid the model overfitting to the defect area, they proposed first using the structural similarity criterion loss for initial training, then using the mean-square error loss to train the network and using the structural similarity criterion loss as the iteration termination condition in subsequent training. This improvement method had a good detection effect on small defects such as holes and scratches on tiles, with a significant improvement compared to classical models. Jia [7] optimized the loss function by combining target-scale information, including using focal loss and smooth L1 loss, effectively improving the detection accuracy of small targets. Cumbajin [30] evaluated three network architectures, AlexNet, VGG, and ResNet, and three training techniques, Train from Scratch (TFS), transfer learning (TL), and transfer learning with fine-tuning (FT), using a CNN convolutional neural network to address the problem of the insufficient detection accuracy of small-target defects during the manufacturing process of ceramic parts. They determined that the combination of ResNet and FT training could improve the detection accuracy of small targets.

For the detection of small-target defects in ceramic-like products, researchers have significantly enhanced the detection accuracy and model adaptability by introducing multiple attention mechanisms, feature fusion, multi-scale receptive fields, and feature pyramid networks. Notably, remarkable progress has been made in the identification of small-target defects. However, existing research still has drawbacks such as high computational complexity and difficult deployment, and some methods lack extensive verification and universality. To further improve the detection accuracy, it is advisable to consider adopting newer network architectures like YOLOv8 and explore lightweight designs and cross-scene applicability. In addition, the optimization of loss functions and training strategies has also demonstrated excellent performance in enhancing robustness. Nevertheless, issues such as long training times and high computational resource consumption still need to be addressed, and the generalization ability of different datasets requires further verification.

6. Real-Time Detection Methods for Ceramic Surface Defects

In the high-speed production environment of ceramic products, real-time monitoring and rapid feedback are crucial for ensuring product quality and production efficiency. Through algorithm optimization and hardware acceleration, deep-learning models can achieve the high-speed detection of ceramic surface defects, significantly shortening the detection time and ensuring that products are accurately detected in a timely manner. This not only helps enterprises handle defective products promptly and prevent defective products from entering the market but also improves the overall efficiency and flexibility of the production line. Therefore, enhancing real-time performance is of great significance for guaranteeing quality, improving production efficiency, and promoting intelligent manufacturing.

6.1. Lightweight Model Improvement

To improve the detection speed, researchers have significantly reduced the computational load and the number of parameters of the model by introducing lightweight architectures and replacing complex modules, thereby enhancing the detection efficiency. Reference [57] introduced various convolution methods of lightweight convolutional neural networks, multiple lightweight network models, and optimization methods, providing ideas for scholars to improve the real-time performance of ceramic surface defect detection.

Chen [56] replaced the VGG network in the SSD model with the lightweight MobileNet network for backbone feature extraction to address the problem of slow running speed when the SSD model detects defects on ceramic curved surfaces. Six effective feature layers were extracted, thus improving the real-time performance of the model. Han [47] selected the lightweight ResNeXt network to replace the VGG-16 network in the traditional SSD algorithm as the feature-extraction network to solve the real-time performance problem of ceramic surface defect detection, improving the detection speed of ceramic surface defects. Chen [48] replaced the traditional convolution in the additional feature-extraction layer with depth-separable convolution based on the SSD network model to address the problem of insufficient real-time performance in the anti-interference detection research of 3D-printed ceramic surface defects. This reduced the computational load and the number of parameters of the network and improved the speed. Hang [5] selected the lightweight backbone network MobileNetV3_Large to address the real-time performance problem in the detection of sanitary ceramic surface defects. Through depth-separable convolution, the number of network parameters was effectively reduced, and a lightweight detection head structure with a channel attention module was designed, as shown in Figure 5. This significantly improved the network’s detection speed and met the real-time performance requirements.

6.2. Network Module Integration and Optimization

The integration and optimization of network modules are important means to improve the defect-detection performance. By integrating advanced modules or making targeted improvements to existing networks, it is possible to significantly enhance real-time performance while improving detection accuracy. Researchers have optimized feature extraction, classifier design, and the overall model architecture for different scenarios and targets, solving the problem of insufficient real-time performance in ceramic defect detection and providing efficient and reliable solutions for industrial applications.

Wang [58] processed the problem of weak real-time performance in the detection of ceramic-clad copper board defects based on the Faster R-CNN network. By comparing the feature-extraction networks of VGG, ResNet-v1, ResNet-v2, and Inception-ResNet-v2, they selected the ResNet-v1 network. In addition, they optimized the region proposal network, adjusted the anchor box parameters to four groups using the K-means clustering algorithm, and then the RPN generated anchor boxes on each pixel of the feature map, obtained regression parameters and target scores through convolution, and generated candidate regions after correction, cropping, and non-maximum suppression, improving the detection speed to a certain extent.

Wang [59] modified the YOLOv5 classifier in response to the low real-time performance in the tile defect-detection scenario. They changed the output dimension from 255 to 30, reducing the number of network parameters of the model and the computational load, and improving the efficiency of defect detection during the tile production process. Chen [41] replaced the “CBS” structure in the backbone part of the YOLOv5 network with the DepRes module composed of depth-separable convolution and a residual network to address the problem of insufficient real-time performance in the multi-scale surface defect detection of 3D-printed ceramic parts. Pan [6] optimized the network structure by introducing the ECA mechanism module based on the YOLOv5s network to address the real-time detection problem in the detection of ceramic disk surface defects. They reduced the number of model layers, the size of the weight model, and the number of network parameters, accelerating the detection speed of the model to a certain extent and alleviating the pressure of real-time detection.

Ma [60] optimized the grayscale image conversion algorithm to improve the image preprocessing speed for the real-time defect detection of advanced ceramic parts, adopted frame-skipping tracking detection to reduce the computational load, and combined it with the YOLOv3 algorithm improved by SKNet. They achieved real-time detection while ensuring the detection accuracy. Yang [61] introduced the Inception structure and depth-separable convolution to construct the YOLOv3-DS to address the problem of low real-time performance in the detection of ceramic medicine bottle defects. While improving the detection accuracy, its frames per second (FPS) increased to 37.4, meeting the real-time performance requirements of enterprises for the detection of ceramic medicine bottle defects and effectively solving the real-time problem of ceramic medicine bottle defect detection. Huang [10] proposed integrating the classification model and the YOLOv3 target-detection model to form a two-stage detection model to address the problem of insufficient real-time performance in the ceramic substrate defect-detection system. This integration reduced the number of images to be processed by YOLOv3 and thus the number of times YOLOv3 needed to perform target detection, improving the detection speed.

The improvement of lightweight models and the integration and optimization of network modules have improved the speed and real-time performance in the detection of ceramic surface defects, reduced the consumption of computational resources, and enhanced the adaptability and stability of detection. However, when facing complex defects, the model has detection errors, especially in cases of data imbalance or high noise. It is necessary to improve the accuracy in complex environments through model fusion and optimization of lightweight algorithms, further reduce the computational complexity, and improve the detection efficiency.

The main problems and corresponding strategies in ceramic surface defect detection are shown in

Table 1. Summary of ceramic surface defect-detection problems.

Problem	Improvement Methods	Advantages	Disadvantages	Future Research Directions
Imbalanced sample detection problem	Data augmentation, clustering sampling, and loss function optimization	Improved minority class recognition and overall accuracy	Limited performance under extreme imbalance	Synthetic data generation and intelligent resampling
Small-sample detection problem	GAN-based augmentation, pre-training, unsupervised learning, and structure design	Reduced overfitting and better generalization and accuracy	Poor robustness and challenges in complex environments	Semi/self-supervised learning and novel augmentation
Small-target-detection problem	Attention mechanisms, multi-scale feature fusion, and network optimization	Enhanced feature representation and better detection	High computational cost and long training time	Lightweight attention and new loss functions
Real-time detection problem	Lightweight networks, module optimization, and simplified structures	Faster detection with acceptable accuracy	Accuracy may drop in complex scenarios	Efficient architecture design and algorithm optimization

.

7. Conclusions

Deep learning has brought revolutionary progress to ceramic surface defect detection, but there are still several key bottlenecks that need to be overcome to achieve industrial implementation. Currently, preprocessing techniques such as normalization and data augmentation have improved model performance, but the lack of standardized open-source datasets remains a major obstacle. Resampling and focus loss methods have alleviated the problem of class imbalance, but the stability of models under extreme imbalance still needs further research. In small-sample scenarios, although transfer learning is effective, its dependence on pre-trained models and the insufficient diversity of defective samples limit its generalization ability. Attention mechanisms and network optimization have improved the accuracy of small-object detection, but the computational cost hinders practical deployment. Lightweight models have improved real-time performance, but their performance remains unstable in complex defects or noisy environments.

Future research should focus on the following: first, developing synthetic data generation techniques to address severely imbalanced data; second, exploring semi-supervised/self-supervised learning for efficient data training; third, designing low computational overhead small-object detection architectures; fourth, deeply integrating material science knowledge to match industrial practical needs. The ultimate goal is to transform laboratory accuracy into robust systems that can be used on production lines, promoting the intelligent upgrading of the ceramic industry.