Small object detection is a pervasive challenge in surface defect inspection tasks. In typical industrial scenarios such as solder balls and scratches on PCBs, delamination and porosity in composite materials, and micro-cracks in silicon wafers, defects are inherently small in size, have subtle textures, and are easily obscured by background noise. This leads to weak feature representation and ambiguous localization within deep neural networks, significantly impacting the recall rate and precision of detection models. The definition of a small object generally falls into two categories: a relative scale based on the area ratio of the object to the entire image (e.g., an area less than 0.12% or 0.03%), and an absolute scale based on the object’s pixel dimensions, the specifics of which depend on the dataset. For instance, the DOTA aerial dataset defines objects with 10–50 pixels as small [8], while the general-purpose MS COCO dataset categorizes objects with resolutions below 32 × 32 pixels as small [9]. Notably, these definitions primarily originate from the field of general object detection. In high-resolution industrial imaging (e.g., PCBs, wafers), an object spanning hundreds of pixels may still be considered an ‘effectively small target’ if it occupies an extremely low proportion of the entire image (<0.01%). To address the challenges of small object detection, current research primarily focuses on technological directions such as multi-scale feature fusion, context information enhancement, and super-resolution reconstruction. (1)

Multi-Scale Feature Fusion-based Methods. Critical details of small objects are often lost after multiple downsampling operations in deep networks. Therefore, fusing high-resolution shallow features with semantically rich deep features has become a mainstream approach. Liu et al. proposed the SSD (Single Shot MultiBox Detector), which first utilized feature maps at multiple scales to detect objects of different sizes, with shallow layers responsible for small objects and deep layers for large ones [10]. Subsequently, building on SSD, Lin introduced the FPN (Feature Pyramid Network) structure, constructing a top-down feature pyramid with lateral connections to effectively fuse shallow details and deep semantics [11].

These methods effectively mitigate the issue of information loss for small objects during downsampling by constructing semantically enhanced feature pyramids. However, multi-scale fusion models often suffer from high computational complexity and semantic inconsistencies between different feature layers. Shallow features are prone to noise, while deep features are highly abstract; direct concatenation or fusion can lead to redundancy or false detections. Consequently, subsequent research has gradually introduced lightweight structures, attention mechanisms, and local super-resolution enhancement modules to achieve more efficient collaboration between shallow and deep information, providing a solid structural foundation for subsequent context modeling and image enhancement strategies. (2)

Context Information-Based Methods. The limited texture information of small objects makes precise recognition based solely on their intrinsic features difficult. Thus, researchers have introduced context information modeling mechanisms to capture the semantic relationships between a target and its surrounding regions, thereby improving discriminative power. Contextual information includes not only neighboring pixels but also semantic cues from the target’s region, such as scene structure, spatial location, and class priors. Ref. [12] proposed an Orthogonal Context Attention Module, which simultaneously enhances contextual details from vertical and horizontal directions, strengthening the feature representation of tiny defects in steel strips.

Context modeling methods are particularly effective in scenarios where objects are surrounded by regular structures (e.g., weld seams, steel plate scratches) and are valuable for small object localization in complex textured backgrounds. However, their performance hinges on the accuracy of context modeling; if the background is complex or the contextual structure is ambiguous, they may introduce redundant or even misleading information. Additionally, the computational overhead of context networks requires careful consideration. Therefore, lightweight context modules and adaptive relationship modeling under Transformer architectures (e.g., global attention, dynamic receptive field regulation) will be key future development focuses. (3)

Super-Resolution-Based Methods. Low resolution directly limits the pixel-level information available for small objects, complicating feature extraction. In recent years, a substantial body of research has incorporated super-resolution reconstruction techniques into small object detection. These methods enhance critical details like edges and textures by improving resolution at either the image or feature level. Dwivedi et al. proposed a low-cost, ESRGAN-based driving method for enhancing the spatial resolution of luminescence images, significantly boosting the detection performance of microscopic defects in solar cells [13].

Super-resolution methods are especially effective in low-quality image scenarios, significantly compensating for detail loss in small objects caused by compression, viewing angles, or sampling limitations. However, challenges are evident: on one hand, Generative Adversarial Network (GAN)-based super-resolution methods are computationally intensive and difficult to deploy in real-time; on the other hand, generated images may contain artificial textures or erroneous boundaries, leading to increased false detection rates. Future directions include dynamic sparse computation (activating high-resolution processing only in suspicious regions), end-to-end differentiable upsampling mechanisms, and a paradigm shift from dense prediction towards sparse representations driven by points or contours.

Summary and Comparative Analysis. To facilitate a comparative understanding of the methods discussed above, Table 1 summarizes the key characteristics, advantages, and limitations of representative approaches for small object detection.

Real-World Implementation Considerations. In practical industrial deployment, small object detection faces additional challenges beyond algorithmic design. The resolution and optical setup of imaging systems directly determine the upper bound of detectable defect sizes. For applications such as PCB inspection or wafer defect detection, telecentric lenses and high-resolution line-scan cameras are often required to ensure that micron-level defects occupy sufficient pixels. Furthermore, motion blur in conveyor belt scenarios can severely degrade small defect features, necessitating integration with motion compensation hardware or deblurring preprocessing modules. These engineering considerations highlight that effective small defect detection requires a holistic approach combining algorithm design with appropriate imaging system configuration.

In tasks such as PCB inspection, road crack detection, and pipeline corrosion assessment, defects are often characterized by their small size, variable morphology, and high degree of blending with complex, cluttered background textures, resulting in an extremely low signal-to-noise ratio. This poses significant challenges for precise localization and identification. Existing methodologies have largely evolved through two iterative phases: the traditional model-driven stage and the deep learning data-driven stage. Traditional methods, based on handcrafted features and machine learning classifiers, are theoretically sound but susceptible to noise interference [14]. With the rise of CNNs, architectures like U-Net, FPN, DeepLab, and BiFPN have significantly improved accuracy by leveraging large-scale annotated data. However, these models are primarily designed for natural scenes with clear boundaries, whereas industrial defect images feature intricate textures and minuscule targets, creating a “domain shift” [15]. To suppress background noise and accentuate defect regions, researchers have extensively integrated attention mechanisms, which can be broadly categorized into three types: channel attention mechanisms, channel-spatial joint attention mechanisms, and global attention mechanisms. (1)

Channel Attention Mechanisms. This mechanism enhances the response strength of discriminative features by learning the importance of each channel. Zhang et al. integrated SENet into YOLOv5, significantly improving the localization accuracy of insulators in foggy conditions [16]. Hao et al. embedded Residual Split Attention into YOLOv4, strengthening insulator defect recognition in complex scenes [17]. Zhou et al. proposed IFIFusion, employing a dual-attention Transformer backbone combined with an LWNet and an IFIF multi-scale coupling module to enhance defect perception [18]. Ma et al. designed ELA-YOLO, introducing linear attention within the backbone and constructing a selective feature pyramid to compress computational costs while maintaining high precision [19]. Channel attention can be integrated into existing detectors without significantly altering the network topology, making it a common solution for rapid upgrades in industrial production lines.

Table 2 provides a comparative summary of attention mechanism-based methods for complex background suppression.

In summary, in industrial visual inspection, defect regions often exhibit low contrast due to high similarity with background textures or limited imaging conditions, making their saliency difficult to express effectively. Beyond the primary focus on introducing attention mechanisms, current research also employs multi-scale feature fusion strategies (e.g., FPN, U-Net++), frequency-domain analysis techniques (e.g., wavelet or Fourier transforms) to enhance the response to weak abnormal signals, as well as self-supervised or contrastive learning methods for pre-training models on unlabeled data to improve sensitivity to subtle anomalies. Although these methods have improved defect discernibility to some extent, challenges remain, such as attention being easily distracted by strong textures, the high computational overhead of frequency-domain methods, and unstable transfer effects in self-supervised learning. Future trends will place greater emphasis on multimodal information fusion (e.g., combining visible light with polarization/infrared images), generative prior modeling (e.g., diffusion models reconstructing normal sample distributions), and physics-informed network design guided by imaging principles to achieve more robust saliency representation.

Addressing the issue of class imbalance caused by uneven distribution of defect samples, current research primarily focuses on two directions: the data level and the algorithm level. This problem is particularly prominent in scenarios such as industrial quality inspection (e.g., rare defects on PCBs, special stains on textiles) and public infrastructure inspection (e.g., road cracks, where the distribution of defects in shape and size is extremely uneven, and severe network cracks are far less frequent compared to numerous fine cracks) [28]. Data-level methods adjust the data distribution through defect sample augmentation (e.g., GAN-based defect generation, data augmentation) and normal sample filtering. Algorithm-level methods enhance the model’s ability to identify long-tail and hard-to-detect defect classes by designing improved loss functions (e.g., weighted cross-entropy, Focal Loss), label optimization strategies, and feature enhancement mechanisms (e.g., attention mechanism-guided feature focusing).

Data-level strategies directly adjust the sample composition before training to mitigate distribution bias inherent in the input data. Among these, normal sample filtering methods mainly include undersampling and oversampling. Random undersampling balances classes by randomly removing samples from the majority class but risks discarding useful information and degrading the model’s ability to discriminate the majority class [29]. Random oversampling augments the dataset by replicating minority class samples, effectively mitigating information loss caused by undersampling but potentially introducing redundancy and increasing training burden [30]. In contrast, the SMOTE algorithm and its variants generate synthetic samples via k-nearest neighbor interpolation, balancing class ratios while avoiding redundancy from simple replication [31]. However, they may produce unrealistic samples near feature boundaries, reducing model generalization.

Defect sample augmentation is another common data-level solution. Transformation-based data augmentation methods [32], such as image cropping, flipping, rotation, and affine transformations, can generate diverse image variations, thereby improving model robustness. While these methods can alleviate global sample imbalance, they do not directly address intra-image class imbalance (e.g., coexistence of multiple defect types). Synthesis-based augmentation methods like Mixup generate new samples through image interpolation or concatenation, enhancing the distributional diversity and generalization capability of minority class samples to some extent [33]. However, controlling the authenticity of synthetic samples is challenging, and some generated images may deviate from the real distribution. To further enhance sample quality, methods based on generative models (e.g., GANs, diffusion models) are emerging. For instance, the Trans-GAN-Cla developed by Ma et al. is suitable for synthesizing drainage pipe defect images [34]. However, these methods heavily rely on high-quality training data and often struggle to generate sufficiently accurate samples in few-shot, multi-class tasks. Generated images also frequently suffer from issues like content blurring and edge distortion.

Algorithm-level strategies compensate for performance degradation caused by data imbalance by optimizing the training process or model architecture, primarily including label optimization and loss function optimization. (1)

Label Optimization Strategies. Traditional methods often rely on a fixed Intersection over Union (IoU) threshold for positive/negative sample assignment. For example, Faster R-CNN [35] and the YOLO series use 0.5 as the positive sample threshold with an ignore region. Such static labeling methods are simple and straightforward but cannot adapt to targets of varying sizes and complexities, limiting model performance improvement. Consequently, researchers have introduced dynamic label mechanisms. For example, ATSS dynamically sets thresholds based on statistical distributions [36], and OTA even models label assignment as an optimal transport problem, balancing matching quality and training efficiency [37]. Correspondingly, the Noisy Anchor approach guides the model to focus on high-confidence samples by weakening the influence of low-quality samples, thereby enhancing learning capability for minority classes [38].

Table 3 summarizes the key strategies for addressing class imbalance at both data and algorithm levels.

In summary, research on class imbalance has gradually evolved from single-dimensional techniques toward multi-dimensional integration. Data augmentation, label mechanisms, loss design, and feature modeling now complement and synergistically optimize each other, forming a systematic strategy to address challenges such as difficult learning of minority class samples, sparse distribution, and high false detection rates. This provides more robust technical support for intelligent surface defect detection in complex industrial scenarios.

In industrial production, defect detection solutions must be flexibly adjusted according to the characteristics of the application scenario. For static or low-speed moving objects, such as PCBs, circuit boards, and glass panels, single-frame image detection is typically employed. This strategy offers three main advantages: First, complete information can be acquired in a single imaging instance, resulting in high processing efficiency. Second, it eliminates the need for object tracking, thereby reducing system complexity. Third, the model structure is simple and suitable for deployment on edge devices, such as portable inspection systems. However, in dynamic scenarios like high-speed conveyor belts, rotating components, or moving assembly lines, video stream detection technology demonstrates significant advantages [43]. This type of technology addresses four key issues: (1) Mitigating blurring and distortion caused by rapid object motion through motion compensation algorithms; (2) Utilizing temporal information to remove background interference such as metal reflections and dynamic shadows; (3) Employing multi-frame fusion to enhance the recognition capability for minute defects (e.g., micron-level cracks); (4) Enabling defect evolution tracking and monitoring through time series analysis. Typical applications include: full-perimeter detection of surface cracks on rotating equipment, identification of intermittent defects under vibration conditions, multi-angle defect modeling on complex curved surfaces (e.g., blades, gears, aerospace components), and quality tracking of long-running structures such as pipelines and bridges.

Traditional inspection methods and fixed-camera solutions often suffer from response lag and detection blind spots, making it difficult to meet the requirements for fine-grained, real-time monitoring. Traditional image processing methods, such as edge detection and threshold segmentation, rely on handcrafted features and parameter tuning, resulting in poor robustness and susceptibility to lighting variations and noise. To improve performance, some systems have attempted to introduce high-precision industrial cameras, structured light, or multi-spectral equipment. However, stringent environmental adaptability requirements (e.g., explosion-proof, high-temperature resistance) and on-site constraints often lead to high costs and deployment difficulties for such solutions [44].

The rapid development of deep learning has driven breakthroughs in detection capabilities for dynamic scenes. End-to-end detection networks have improved recognition accuracy and adaptability through automatic feature extraction [45,46]. Liu et al. proposed a domain fluctuation suppression strategy based on deep ensemble learning, which mitigates dynamic data imbalance and cross-domain shifts, making it suitable for real pipeline inspection [47]. Tsung et al. developed a high-performance road health recognition system that combines simulation and measured data to achieve intelligent identification of various road damages, such as longitudinal/transverse cracks and potholes. Deployed on vehicle-mounted edge devices, it effectively improves inspection efficiency [48]. Guo et al. optimized YOLOv5 by incorporating an attention mechanism for complex road defect detection on mobile devices, successfully deploying the system on a portable platform [49].

Beyond video analysis methods, some research has also explored the integration of robotic and Unmanned Aerial Vehicle (UAV) platforms. However, Yang et al. point out that such systems are limited to providing only local positioning and are suited for small-scale scenes. To address this, they proposed the Det-ReconReg framework, which integrates UAV and machine learning technologies to achieve defect detection and localization in large-scale infrastructure, significantly enhancing the system’s scalability [50].

To address motion blur and artifacts in dynamic images, Cui et al. proposed a two-stage weld seam detection method called TRDM. It first uses a lightweight network (LSN) for target localization, followed by an SRD network for defect recognition, effectively improving detection accuracy [51]. Liu et al. designed a belt damage detection model combining an attention mechanism with a Temporal Convolutional Network (TCN), processing complex backgrounds and dynamic shadow interference from both spatial and temporal dimensions, significantly improving the robustness of dynamic detection [52].

Regarding industrial implementation, Xu Hao et al. built a cable defect identification system based on YOLOv5, achieving efficient detection and tracking through adaptive parameters and multi-threshold strategies [53]. Yan Hexiang et al. combined YOLOv7-seg with DeepSORT to create a pipeline defect identification framework capable of simultaneous detection and tracking. The model’s performance was optimized using a dual-training dataset to meet the demands of pipeline network inspection [54]. These solutions highlight the significant value of the “detection + tracking” technical approach for dynamic defect detection.

It is worth noting that despite considerable progress in existing research, most deep learning models are still based on static distribution assumptions. Real industrial data exhibits dynamic changing characteristics, which easily triggers domain shift problems and leads to performance degradation when models are deployed on production lines. The emergence of these issues reveals that current dynamic scenario defect detection still faces significant challenges in terms of model generalization and cross-domain adaptability, urgently requiring further research and solutions.

With the increasing demand for high robustness and adaptability in industrial intelligent inspection, traditional surface defect detection methods relying on single models or single features face significant performance bottlenecks in multi-scenario, multi-target, and multi-modal tasks. In recent years, researchers have conducted in-depth exploration around three main research directions: structure optimization-driven generalization enhancement methods, cross-domain adaptation and transfer learning mechanisms, and multi-task fusion with lightweight deployment optimization strategies. These efforts aim to mitigate performance degradation of models across different defect types, complex environments, and heterogeneous data. (1)

Structure Optimization-Driven Generalization Enhancement Methods. Traditional surface defect detection methods often rely on rule-making based on low-level image features like thresholds, edges, and textures, or combine shallow classifiers (e.g., SVM, K-NN) to achieve defect identification. Although these methods offer strong interpretability in specific scenarios, their generalization ability heavily depends on input image quality, lighting conditions, and background interference, making it difficult for them to adapt to complex and variable industrial environments [55]. In recent years, CNN-dominated deep learning methods have rapidly risen, improving models’ ability to represent unstructured data through end-to-end feature extraction and detection strategies. Models like Faster R-CNN, the YOLO series, and SSD represent the mainstream development of two-stage and one-stage methods. Some architectures enhance robustness to scale variations and background interference by incorporating feature pyramids, attention mechanisms (CBAM, SE), and dilated convolutions.

However, most of these structurally improved models still focus on improving accuracy on a single dataset, overlooking robustness across different materials, equipment, and process conditions. To address this problem, researchers have proposed hybrid attention mechanisms (e.g., integrating Transformer self-attention with channel-spatial fusion attention) and dynamic structure regulation strategies, such as deformable convolutions [56] and conditional convolutions (CondConv) [57], thereby enhancing the model’s adaptive capability to different defect geometric structures and texture distributions. Additionally, multi-stage architectures like cascade detection heads have been introduced to alleviate instability in detector regression under high IoU thresholds, further improving boundary precision and classification discriminability [58]. These methods provide fundamental support for the generalization of surface defect detection models in typical scenarios involving complex backgrounds, small targets, and strong occlusion. (2)

Application of Cross-Domain Adaptation and Transfer Mechanisms in Generalization. In industrial applications, significant distribution differences often exist across different production batches, manufacturing equipment, and material surfaces. This leads to severe performance degradation—manifested as “domain shift” or “data drift”—when models trained on a source domain are deployed on a target domain. To address this issue, cross-domain learning has become an important means to enhance generalization in recent years, primarily including two categories: Domain Adaptation (DA) and Domain Generalization (DG). In DA methods, typical strategies like Domain-Adversarial Neural Networks (DANN) [59], Maximum Mean Discrepancy (MMD) [60], and adversarial style transfer networks (e.g., CycleGAN) [61] aim to adapt training to unlabeled target domains by minimizing the feature distribution discrepancy between source and target domains. In DG strategies, researchers enhance model performance on unseen target domains through multi-source data joint training, meta-learning optimization, or style perturbation augmentation.

Despite the good results achieved by these methods in classification and segmentation tasks, the specificity of surface defect detection tasks—such as sparse defect instances, large morphological variations, and difficulty in obtaining labels—makes it challenging to apply transfer methods directly. Consequently, some work has proposed self-training mechanisms based on pseudo-labels, dual-decoder and multi-discriminator structures to improve stability and consistency in unlabeled target domains. Other research has introduced domain attention modules to achieve dynamic weighting and adaptive aggregation in feature spaces. Furthermore, for specific defects (e.g., cracks, scratches) with extremely limited representation in the target domain, sample weighting mechanisms and hard example mining strategies are used to enhance the discriminability of feature representations. Overall, cross-domain adaptation mechanisms are gradually becoming a core direction for improving model generalization across different scenarios, particularly suitable for needs like batch updates, equipment migration, and remote deployment in defect detection contexts. (3)

Advances in Generalization through Multi-Task Fusion and Lightweight Deployment Strategies. As task complexity continues to rise in smart manufacturing scenarios, detection models driven by a single task can no longer meet the comprehensive demands for multi-dimensional objectives like defect classification, localization, segmentation, and counting. Therefore, Multi-Task Learning (MTL) has become an effective means to enhance model generalization and expressive power. By sharing a bottom-level encoder and introducing task-specific branches with optimized objective functions, researchers have constructed models with joint detection-segmentation functions, such as PAD-Net [62] and DefectNet [63]. These models demonstrate stronger robustness and contextual expression capabilities in surface defect scenarios like steel, PCB, and photovoltaic modules. Additionally, designs that integrate auxiliary tasks like classification, feature matching, and defect counting help mitigate overfitting caused by scarce samples and enhance generalized feature expression.

On the other hand, in industrial settings, inspection systems often face constraints in computational power and require deployment flexibility. Therefore, lightweight model design and knowledge distillation strategies have become another core direction for improving practical usability. Researchers have developed various lightweight detectors based on architectures like MobileNet [64] and ShuffleNet [65], incorporating self-attention mechanisms and residual connections for deployment on edge or mobile platforms. Simultaneously, knowledge distillation, as a “large-teaches-small” transfer paradigm, is widely used to transfer the feature distributions, prediction scores, and attention responses from large teacher models to smaller student models. Frameworks like adversarial distillation and multi-teacher collaborative distillation, which integrate multi-task and multi-perspective knowledge from teachers, significantly enhance the generalization capability and compression effectiveness of student models in complex defect detection tasks. For example, Zhou et al. implemented stable detection of various defect types based on knowledge distillation [66], while Hu et al. achieved high-performance PCB defect detection under lightweight conditions using aligned soft-target knowledge distillation [67].

In summary, multi-task fusion and lightweight deployment methods not only improve the synergy and performance of models across multiple sub-tasks in object detection but also greatly promote the deployability of surface defect detection algorithms, establishing an integrated foundation for generalization from perception algorithms to system integration.

To systematically understand existing datasets, this paper introduces datasets from the following four dimensions, as detailed in Table 4.

In practical research, an appropriate dataset can be selected based on the specific defect challenges according to the following procedures.

Beyond quantitative characteristics, the quality of dataset annotations and the alignment with real industrial conditions significantly impact the validity of research findings. This section discusses key considerations for dataset quality assessment.

Annotation Quality. Datasets vary substantially in annotation granularity and consistency. Image-level labels (e.g., classification datasets like ELPV) are suitable for defect classification tasks but insufficient for localization or segmentation. Bounding box annotations (e.g., RDD2022, GC10-DET) enable object detection but may suffer from subjective variations in box boundaries, particularly for irregular defects like cracks. Pixel-level masks (e.g., MVTec AD, CrackForest) provide the highest annotation detail but are costly to produce and may exhibit inconsistencies across annotators. Researchers should consider these factors when selecting datasets for specific tasks; for instance, segmentation models trained on bounding box annotations may achieve inferior boundary precision.

Noise and Diversity. Real industrial environments introduce numerous confounding factors that are often absent in laboratory-collected datasets. Important diversity dimensions include:

Illumination variations: Datasets like DTU-Drone and RDD2022 capture images under diverse lighting conditions, enabling evaluation of illumination robustness.

Background complexity: While datasets like NEU-DET have relatively uniform backgrounds, AITEX and PCB datasets incorporate complex textural backgrounds that challenge detection algorithms.

Sensor variability: Datasets combining visible light, thermal (PTID), and electroluminescence (ELPV) imaging enable multimodal algorithm development.

Industrial Applicability Assessment. Based on the above considerations, Table 8 provides a qualitative assessment of dataset suitability for different industrial application contexts.

This paper provides a systematic review of five core challenges in machine vision-based surface defect detection: complex background interference, small object detection, class imbalance, dynamic scene modeling, and cross-scenario generalization. By combing the technical paths of traditional image processing, CNN-based deep learning methods, and the emerging Transformer architecture, the advantages and limitations of various methods in dealing with specific challenges are revealed. Research shows that attention mechanism, multi-scale feature fusion, context modeling, dynamic label assignment and loss function design, temporal reasoning and domain adaptation technologies have become key means to improve the robustness, accuracy and adaptability of detection models in complex industrial environments.

Concurrently, this paper comprehensively summarizes and analyzes the characteristics of publicly available datasets covering a wide range of industrial scenarios. While these datasets have propelled algorithmic research, they also expose limitations misaligned with real industrial needs, such as dominance by static single-frame images, heavy reliance on full supervision, and singularity in modalities and scenarios. Constructing benchmark datasets that better reflect reality (e.g., containing temporal information, supporting weak/unsupervised learning, fusing multimodal data) will be a crucial foundation for advancing technology deployment.

Overall, although deep learning methods have made significant progress, the current technological system still exhibits a fragmented, “challenge-driven, scenario-specific” characteristic. A “generalized” defect visual inspection system capable of stable operation, efficient generalization, and easy deployment in variable and demanding real industrial environments still faces dual theoretical and engineering challenges.

Looking ahead, surface defect detection technology will evolve towards greater intelligence, generalizability, lightweight design, and trustworthiness. In conjunction with advancements in cutting-edge technologies like large models, the following directions warrant in-depth exploration:

Revolution of Visual Large Models and Foundation Models: The perceptual capabilities of current mainstream models based on CNNs and ViTs remain constrained by specific tasks and datasets. The emergence of Vision Foundation Models and Vision-Language Models (VLMs) brings the potential for a paradigm shift. Their potential is evident in: 1) Few-shot generalization ability. Leveraging the powerful visual-language alignment and in-context learning capabilities of large models, it may be possible to identify unseen defect types using only a few examples or textual descriptions (e.g., “micron-level linear scratch”), alleviating detection challenges for few-shot and new defect categories. 2) Cognitive ability for complex scenes. Large models possess stronger scene understanding, relational reasoning, and commonsense knowledge. This aids in distinguishing real defects from background pseudo-textures and understanding “logical anomalies” (e.g., incorrect/missing parts), enabling semantic-level defect diagnosis closer to that of human experts. 3) Unified multimodal representation and generative augmentation. Generative large models like diffusion models can synthesize rare defect samples with higher quality or generate realistic anomalous samples based on normal samples to address data imbalance. Simultaneously, large models can serve as unified encoders to fuse multimodal information (e.g., visible light, infrared, X-ray, 3D point clouds), enhancing the comprehensiveness and reliability of defect detection.

Integrated “Detection-Tracking-Diagnosis-Decision” Intelligent Systems: Future systems should evolve beyond mere defect locators into closed-loop intelligent agents integrating real-time perception, tracking, evolution analysis, and predictive maintenance decision-making. This requires deep integration of temporal modeling, knowledge graphs, physical mechanism models, and reinforcement learning to achieve the leap from “seeing defects” to “understanding defect evolution, predicting failure risks, and guiding maintenance strategies.”

Lightweight and Adaptive Architectures for Edge Computing: To meet the real-time, low-power, and privacy-preservation demands of industrial sites, models must be extremely lightweight. Future research needs to combine Neural Architecture Search (NAS), dynamic networks, adaptive computation, and advanced model compression and distillation techniques to develop “agile” models that can dynamically adjust computational resources based on input complexity, enabling high-performance deployment on resource-constrained edge devices.

Building an Open, Collaborative Industrial Vision Ecosystem: Promote the establishment of large-scale open-source benchmark datasets encompassing richer scenarios, multiple modalities, and incorporating temporal and physical information. Simultaneously, explore distributed collaborative training frameworks based on federated learning and privacy-computing technologies to aggregate industry knowledge and jointly train more powerful, general-purpose defect detection foundation models while ensuring the data privacy of participating enterprises.