Closing Editorial: Advances in Computer Vision and Image Processing for Industrial Processes

1. Introduction

Computer vision and image processing have become essential technologies for the transformation of conventional industrial systems into intelligent, automated, and data-driven processes. Over the past decade, advances in deep learning, high-performance computing, sensing technologies, and industrial Internet-of-Things infrastructure have substantially expanded the role of visual information in manufacturing. Computer vision is no longer limited to offline inspection or measurement; it is increasingly being integrated into real-time monitoring, quality control, process optimization, predictive maintenance, safety management, and human–machine interaction.

The transition, however, from laboratory demonstrations to reliable industrial deployment remains challenging. Industrial environments are characterized by variations in illumination, viewpoint, surface appearance, motion, occlusion, sensor noise, environmental conditions, and production configurations. In addition, industrial datasets are often small, imbalanced, proprietary, and expensive to annotate. Consequently, an algorithm that performs well on a conventional benchmark may not necessarily provide the robustness, computational efficiency, interpretability, and reliability required in an actual production environment [1,2].

The Special Issue entitled “Advances in Computer Vision and Image Processing for Industrial Processes” was organized to provide a forum for recent developments addressing these challenges. The six articles collected in this Special Issue demonstrate the breadth of current research, ranging from image stitching and document understanding to visual safety monitoring, structural inspection, human-in-the-loop machine learning, and lightweight perception using pressure-distribution images. Together, these contributions illustrate an important trend: industrial computer vision is evolving from isolated recognition algorithms toward integrated perception systems designed around the constraints and requirements of real industrial processes.

2. From Conventional Image Processing to Deep Industrial Vision

Early industrial vision systems relied heavily on carefully designed image-processing pipelines, geometric models, feature descriptors, and task-specific heuristics. These approaches remain valuable because they can be computationally efficient and physically interpretable. However, the increasing complexity of industrial scenes has encouraged a transition toward learned representations.

Deep learning has demonstrated significant potential in manufacturing applications, particularly for classification, object detection, segmentation, anomaly detection, and process monitoring [1]. Convolutional neural networks and their successors can learn hierarchical representations directly from images and thereby reduce dependence on manually designed features. At the same time, recent developments in vision transformers have demonstrated the advantages of global attention and large-scale pretraining for visual representation learning and dense prediction [3,4].

Several papers in this Special Issue reflect this transition. Feng et al. [5] proposed an improved image-stitching framework based on SuperPoint and LightGlue for industrial monitoring in steel production. By combining learned feature extraction and matching with clustering and adaptive image fusion strategies, the work addresses practical limitations associated with restricted camera viewpoints, parallax, and ghosting. This study illustrates how modern learned local features can be integrated with geometric image-processing techniques to construct more comprehensive industrial visual representations.

Han et al. [6] addressed a different but equally important problem: extracting information from engineering documents. Engineering and contract documents contain heterogeneous structures, including text, tables, forms, and graphical elements. Conventional optical character recognition alone is insufficient for such documents. Their cost-efficient active transfer learning framework therefore highlights the importance of reducing annotation requirements while adapting visual models to specialized industrial document domains.

These studies emphasize a broader point. The future of industrial vision is unlikely to be determined by a single dominant algorithmic paradigm. Instead, effective systems will increasingly combine learned representations with geometric constraints, domain knowledge, task-specific processing, and system-level engineering.

3. Real-Time Inspection, Safety, and Structural Monitoring

Industrial computer vision is particularly valuable when visual information can be converted into immediate operational decisions. In such applications, accuracy alone is insufficient. Detection latency, computational cost, false-negative rates, robustness, and integration with existing industrial systems are equally important.

Lu et al. [7] investigated the detection of flying-steel accidents in high-speed wire-rod production using an improved YOLOv8 architecture. Their method incorporates dynamic convolution, lightweight feature processing, and multi-scale attention. The reported results demonstrate both high detection accuracy and low latency, illustrating the continuing importance of model optimization for real-time industrial safety applications.

Similarly, Liu et al. [8] developed LGM-YOLO for automated structural-defect detection in escalator steel trusses. The proposed framework combines a grayscale-image enhancement strategy with local and global contextual processing and multi-scale feature fusion. The study is particularly relevant to industrial inspection because subtle surface defects may occur at different spatial scales and under challenging imaging conditions.

These works also reveal a persistent gap between benchmark performance and industrial reliability. Industrial defects can be rare, visually heterogeneous, and highly consequential. Industrial anomaly detection research has therefore increasingly investigated learning paradigms that can operate with predominantly defect-free training data [9,10,11]. The MVTec AD benchmark, for example, was introduced specifically to support the systematic evaluation of industrial anomaly detection and contains more than 5000 high-resolution images across 15 object and texture categories [11]. More recent approaches such as PatchCore demonstrate that representative feature memory banks can provide highly effective anomaly detection and localization without requiring large collections of defective samples [10].

The remaining challenge is to move beyond static image-level evaluation toward continuous, process-aware monitoring. An industrial system should ideally recognize not only that an image is anomalous but also whether the anomaly represents a transient disturbance, an emerging equipment failure, a process deviation, or a safety-critical event.

4. Human-Centered and Multimodal Industrial Perception

Another important direction is the integration of human factors into industrial visual intelligence. Industrial systems operate in environments where humans and machines increasingly coexist. Consequently, perception technologies must account not only for products and equipment but also for workers, ergonomic conditions, safety, and human behavior.

Li et al. [12] investigated human sitting-posture recognition using pressure-distribution maps obtained from a thin-film pressure-array sensor. Rather than relying on a computationally intensive modern architecture, the authors improved the lightweight LeNet-5 architecture through increased convolutional capacity, batch normalization, dropout, and data augmentation. The resulting system achieved 95.5% classification accuracy for five sitting-posture classes while retaining a relatively lightweight architecture.

This work illustrates an important principle for industrial and embedded AI: the most sophisticated model is not necessarily the most appropriate model. When computation, memory, power consumption, latency, and data availability are constrained, carefully designed lightweight models may provide a more favorable engineering trade-off.

More generally, future industrial perception systems will increasingly become multimodal. RGB images may be combined with depth, thermal information, acoustic signals, pressure distributions, vibration, electrical measurements, and process parameters. Such integration can provide complementary information when any individual sensing modality is ambiguous.

5. Toward Trustworthy and Human-in-the-Loop Industrial AI

A central obstacle to industrial AI deployment is trust. A model may achieve excellent average accuracy while still producing unacceptable failures under rare or previously unseen conditions. This issue is particularly serious in safety-critical visual inspection.

Wiggerthale and Reich [13] addressed this problem through an operational implementation of the R4VR framework for safe human-in-the-loop machine learning in visual inspection. Their approach combines uncertainty detection, explainability, and model diversity, thereby creating mechanisms for identifying potentially unreliable predictions and involving human operators when appropriate. In their steel-surface-inspection example, the resulting system demonstrated a very low rate of undetected misclassification while reducing the amount of manual effort required over time.

This direction is highly significant because industrial AI should not necessarily aim for complete removal of human involvement. Instead, an effective architecture may allocate decisions dynamically between automated perception and human expertise. Such systems can be viewed as adaptive human–machine decision systems in which the machine handles routine and high-volume cases, while uncertainty triggers additional inspection or human intervention.

This perspective also suggests that future evaluation protocols should extend beyond conventional accuracy, precision, recall, and mAP. Industrial systems should additionally be evaluated according to uncertainty calibration, robustness to domain shifts, failure detection, interpretability, latency, energy consumption, maintainability, and the consequences of false decisions.

6. Contributions of the Special Issue

Taken together, the six contributions in this Special Issue can be viewed as addressing several complementary layers of industrial visual intelligence:

• Visual representation and scene integration: improved image stitching enables multiple viewpoints to be combined into more comprehensive industrial scenes [5].
• Industrial information extraction: active transfer learning reduces the annotation and adaptation burden for engineering documents [6].
• Real-time safety monitoring: optimized object detection supports rapid identification of hazardous events in high-speed production environments [7].
• Automated structural inspection: context-aware multi-scale detection improves recognition of structural defects under challenging visual conditions [8].
• Trustworthy visual inspection: human-in-the-loop mechanisms address uncertainty and safety in machine learning-based inspection [13].
• Lightweight multimodal perception: pressure-distribution sensing combined with a compact neural architecture demonstrates the potential of resource-efficient industrial perception [12].

Although the six studies address different application domains, a common theme emerges: industrial computer vision must be designed as part of an operational system rather than treated solely as an image-recognition problem.

7. Future Research Directions

Several research directions deserve particular attention.

First, industrial foundation models are likely to become increasingly important. Vision transformers and large-scale visual pretraining have demonstrated that representations learned from broad datasets can transfer across tasks and domains [3,4,14]. More recently, foundation models such as Segment Anything have demonstrated promptable segmentation with impressive zero-shot transfer capability [15], while DINOv2 has demonstrated the potential of self-supervised visual representations for general-purpose downstream applications [16]. The challenge for industrial vision is to adapt such models to specialized domains without requiring enormous quantities of proprietary industrial data.

Second, data-efficient and anomaly-centric learning will remain essential [17]. Industrial defects are often rare, while normal operating data are abundant. Consequently, self-supervised learning, few-shot learning, one-class learning, synthetic-data generation, active learning, and anomaly detection should receive continued attention [9,10,11].

Third, physics- and process-informed vision deserves greater emphasis. Purely data-driven models can struggle when production conditions change outside the training distribution. Incorporating geometric constraints, optical models, physical process models, temporal dynamics, and engineering knowledge could improve both robustness and interpretability. In this context, classical computer vision and image-processing techniques should not be regarded as obsolete; rather, they can provide valuable inductive biases for modern learning systems.

Our own previous research in structured-light range imaging provides an example of this complementary perspective. Color and binary structured-light techniques can provide rapid three-dimensional measurements from projected patterns, including one-shot or single-frame acquisition [18,19,20]. More recently, value-probability analysis was used to improve linear phase estimation in sinusoidal structured-light range imaging [21]. These studies illustrate how optical modeling, image formation, signal processing, and algorithmic design can be combined to obtain reliable visual measurements. Such approaches may become increasingly relevant as industrial AI moves from passive image recognition toward quantitative three-dimensional and physics-aware perception.

Fourth, edge intelligence and resource-efficient vision will become increasingly important. Industrial cameras and sensors are being deployed closer to production equipment, where bandwidth, latency, power, and privacy constraints favor local inference. Lightweight neural networks, model compression, knowledge distillation, efficient transformers, event-based sensing, and hardware-aware neural architecture design, therefore, represent promising research directions.

Fifth, multimodal and temporal industrial intelligence should be developed beyond independent processing of individual sensor streams. A future industrial perception system may jointly reason over video, depth, acoustic signals, vibration, pressure, temperature, and process-control data. Temporal reasoning will also be critical because many industrial failures are better characterized as evolving processes rather than isolated anomalous frames.

Finally, trustworthy and continuously adaptive industrial AI should become a central research objective. Models deployed in factories must cope with domain shifts caused by changes in machines, materials, lighting, cameras, production recipes, and environmental conditions. Continuous monitoring of model confidence, data distributions, and failure modes should, therefore, accompany model deployment. Human-in-the-loop mechanisms, uncertainty estimation, explainability, and continual learning may provide the foundation for safer long-term operation [13].

8. Conclusions

The six papers collected in this Special Issue demonstrate the expanding role of computer vision and image processing in industrial processes. From image stitching and engineering-document analysis to high-speed safety monitoring, structural inspection, human-centered perception, and trustworthy visual inspection, the contributions represent different aspects of the transition toward intelligent industrial systems.

At the same time, the field is entering a new stage. The primary challenge is no longer simply to achieve higher recognition accuracy on a fixed benchmark. Instead, the central question is how to build visual intelligence that is accurate, efficient, robust, explainable, adaptive, and economically useful under real industrial conditions.

The convergence of computer vision, foundation models, multimodal sensing, edge computing, physics-based modeling, and human–machine collaboration provides an exciting opportunity to address this challenge. Future research should therefore increasingly move from isolated algorithms toward integrated perception-and-decision systems that can understand industrial environments, quantify uncertainty, learn efficiently from limited data, and support real-time operational decisions.

We hope that the papers in this Special Issue will stimulate further research and collaboration among researchers in computer vision, image processing, artificial intelligence, sensing, robotics, and industrial engineering. The continued integration of these fields will be essential for realizing the next generation of intelligent and resilient industrial processes.