To generate thematic or social clusters, clustering is another pivotal bibliometric technique that is utilized. Creating network clusters, while observing their development could help in understanding how research fields effectively develop and manifest. For example, the social cluster that was created, using a bibliographic coupling, along with co-citation analysis, shows some of the core themes that have emphasized the intellectual structure vis a vis their evolution within the research field. The significant difference between both analyses is that the co-citation analysis uses references from the primary document to illuminate the foundational themes in research fields. On the other hand, Bibliographic coupling uses primary documents of the given sample as a unit of analysis, commonly citing a secondary document to identify the periodical or current thematic development in a research domain (Donthu et al., 2021). In short, bibliographic coupling attempts to determine the emerging themes and potential future research avenues (Vogel et al., 2021). Therefore, in this study, we implemented a bibliographic coupling technique to identify the current research trend within the scholarly literature on computer vision and provide an appropriate research direction for future work.
6.2 Thematic assessment of the clusters
Table 3
List of articles within the clusters through bibliographic coupling measure
Cluster No | Labels | References |
Cluster 1 (Red) | Automated visual inspection | (Liu and MacGregor, 2006; Cheng and Jafari, 2008; Medina et al., 2011; Zhang, Prajapati and Peden, 2011; Barua et al., 2014; Bhat et al., 2015; Feng et al., 2015; Tsang et al., 2016; Grasso et al., 2017; Liu et al., 2017; Aminzadeh and Kurfess, 2019; He et al., 2019) |
Cluster 2 (Green) | Object tracking and process controlling | (Zhang, Lin and Wang, 1999; Guerra and Villalobos, 2001; Chen, 2002; Edinbarough, Balderas and Bose, 2005; Kim, B., et al., 2006; Xie et al., 2008; Taylor, Sivakumar and Dhanalaxmi, 2015; Tsarouchi et al., 2016; Chen et al., 2018; Wang et al., 2018;) |
Cluster 3 (Blue) | Real-time monitoring | (Smith, 1999; Smith and Smith, 2005; Tsai, 2011; Li et al., 2014; Jian, Gao and Ao, 2017; Xuan et al., 2018; Iglesias, Martínez and Taboada, 2018; Badmos et al., 2019; Jin et al., 2019; Lian et al., 2019; Huang et al., 2019; Penumuru, Muthuswamy and karumbu, 2020; Chen et al., 2020) |
Cluster 4 (Yellow) | Roughness inspection | (Ho et al., 2002; Tsai and Wu, 2000; Tsai, Chang and Chao, 2010; Krishnan et al., 2019) |
Cluster 5 (Purple) | Profile projection | (Carbone et al., 2001; Chan, Bradley and Vickers, 2001; Bi and Kang, 2014; Molleda et al., 2013; 2016; Martinez et al., 2019) |
Cluster 1: Automated Visual inspection
Timely identification of defects within a product is pivotal for the manufacturing industry to ensure a streamlined production process (Cheng and Jafari, 2008), product quality (Aminzadeh and Kurfess, 2019), customer satisfaction (Medina et al., 2011) and minimizing the production cost (Liu and MacGregor, 2006). As a result, computer vision-based systems have been advancing to support manufacturing defect detection and quality control. Human visual inspection can be laborious, time-consuming, expensive, and unreliable (Tsang et al., 2016). In fact, human defect inspection of 100 per cent production has become impossible due to the production needs, and the subsequently, the repetitive task effectively led to bias or error-proneness (Medina et al., 2011). Thus, when we adopt a computer vision-based framework, it would help in replacing manual inspection, and thereby reduce human errors.
The content analysis of published articles in this cluster showcases that the main focus of scholars was on developing and implementing an efficient visual inspection system for a manufacturing process that can automate multiple tasks such as defect detection, process monitoring and quality control. Aminzadeh and Kurfess (2019) used computer vision components and a Bayesian classification algorithm to identify the defective and unacceptable layers that occur during the laser powder-bed fusion process. Output of the system can be used for real-time process control, and appropriate action for the manufacturing process. Other scholars worked in defect detection related to the solid modeling process (He et al., 2019), welding process (Bhat et al., 2015) and 3-D modelling process (Liu et al., 2017).
As a result, proposed automated visual inspected systems are largely optimized, and algorithmic developments are less computational. Thus, it is capable of providing operational excellence in the production task. However, manufacturing processes generally have a mixed environment that essentially makes selections of distinctive features more difficult, compelling the researchers to select the features on a case-by-case basis (Feng et al., 2015).
Cluster 2: Object tracking and process controlling
I4.0 automation or smart manufacturing calls for a paradigm shift in manufacturing styles, i.e., moving from massive production to small quantities or massive customization. In the production process, these modes create the need of utilizing robotics systems with intelligence, agility, reliability, and cost-efficiency features for process automatization (Chen et al., 2018). However, the primary goal of using such systems is to accomplish dexterous manipulation tasks, for which, the robotic system does need to detect and track the objects appropriately, to determine manipulation position and thereby adjust its position for the task at hand (Xie et al., 2009). Earlier, robotic systems required manual data inputs or human collaboration to perform the tasks, which made the system semi-autonomous.
With the advancement in model training and deep learning processing, computer vision-based systems are being utilized today for detecting and tracking objects from a real-world scenario. As a result, robotic systems in manufacturing industries are now integrating with computer vision to provide visual data insights too. Besides, published articles within this cluster also concentrate on creating and advancing computer vision-based systems to object tracking for the manufacturing robots. Wang et al. (2018) discussed a real-time 3D human tracking system, whereby they used a monocular camera with an ultrasonic sensor by the extended Kalman filter (EKF).
Tsarouchi et al. (2016) presented a 3D recognition system that integrates the data of an image-based system and CAD system to track and estimate the co-ordinate of the object. In contrast, some scholars focused on algorithmic improvement for robotic systems to process control in manufacturing. For instance, Chen et al. (2018) developed a stereo vision algorithm using point cloud data generated from a visual system to select and adjust the pose of a robotic mobile manipulator for the manufacturing process.
Cluster 3: Real-time monitoring
In a highly competitive global marketplace, maintaining optimum performance and streamlining the production process is always a top priority for manufacturers to ensure the organization’s success. With the introduction of I4.0, the manufacturing industry has now heavily focused on automation, interconnectivity, cyber-physical system, and machine learning to attain operational excellence in their manufacturing. However, the technology and methods utilized for such advanced system heavily depend upon the mathematical calculation or pre-judgment decision, which require extensive human experience and reliable processing time. Therefore, such systems need to stop or pause during the production process to make appropriate decisions for further procedures, and it makes the systems incapable of real-time processing (Jin et al., 2019).
In order to find a suitable solution for this problem, industry experts and academic scholars have tried to develop a computer vision-based system capable of performing real-time monitoring. Articles within this cluster focused on building a real-time monitoring system, integrating computer vision technology for making the manufacturing task more robust. The most representative publication within this cluster by Huang et al. (2019), who presented a rapid recognition model using the (You Look Only Once) YOLO-V3 algorithm for real-time inspection of electronic components in the production process. Other scholars utilized the convolutional neural network (CNN) algorithm to develop real-time monitoring systems; such as: Jin et al. (2019) for fused deposition modeling, Chen et al. (2020) for solar cell manufacturing and Badmos et al., (2019) for lithium-ion battery manufacturing. The primary reason for utilizing the CNN model for real-time monitoring is its capability of features parameter sharing and dimensionality reduction.
Cluster 4: Roughness inspection
Roughness is an essential component of surface texture, which plays a vital role in determining how an object interacts with its environment. The roughness value decides the quality of the end products and mechanical components' performance in the machining process (Krishnan et al., 2019), as irregular surfaces' roughness may act as nucleation sites for corrosion or cracks. It is crucial for industries to measure the surface roughness accurately to adhere to the required quality standard of production. However, the traditional stylus method is the most widely used technique for roughness measurement within the industry. But a significant problem associated with such approaches is that they require direct physical contact and line sampling that cannot represent the real characteristics of the surface.
Visual systems (data acquisition platform) are becoming an essential tool for both inspection and measure of applications in manufacturing automation. Most publications in this cluster included Lee et al. (2004), who developed a model using an abductive network. In fact, they measured the surface roughness of the turned part. A system capable of measuring surface roughness at a lower cost and time. Ho et al. (2002) proposed a method using an adaptive neuro-fuzzy inference system (ANFIS) to efficiently predict the surface roughness using cutting parameters. In contrast, some scholars utilized artificial Neural networks (ANN) to predict and measure roughness (Krishnan et al., 2019).
Cluster 5: Profile projection
Since the industrial revolution, manufacturing processes have usually been composed of a mixed environment of small and large parts. Consequently, it is important to dimensionally test each component against prescribed standards before incorporating it into the manufacturing systems in order to ensure high quality of production lines. However, dimension inspection of multiple parts requires 3D point measurements of the components. For which 3D data acquisition techniques are most commonly utilized; most of the 3D data acquisition techniques are based on both contact and non-contact methods (Molleda et al., 2013).
Contact methods measure the dimensionality of an object through a touch probe, which is an iterative way to acquire points of the physical model and make the process both time-consuming and inaccurate (Carbone et al., 2001). Non-contact devices on the other hand, use magnetic, optical or acoustic principles to obtain 3D data point from an object. This seems to be more efficient in terms of speed, reducing human labor too (Molleda et al., 2013).
The published articles within this cluster also primarily focus on developing and enhancing a computer vision-based system for geometrical measurement of various components used in the production line. Molleda et al. (2013) developed a novel 3-D imaging system for dimensional quality inspection of long-inflated products in the metal industry. Chan et al. (2001) integrated stereo vision and a Co-ordinate measuring machine (CMM) to speed up the process of dimension measurement. Finally, in rail manufacturing, Molleda et al. (2016) built a novel machine-vision-based non-contact profile measurement system for rail rolling mills for measuring the rail dimensions, based on geometric parameters of the transverse section.
6.3 Future research agenda
Based on review of extant literature, the study proposes six following agendas of research to advance future exploration of computer vision applications in manufacturing.
6.3.1 Widening the multi-dimensional scope of understanding- The co-word analysis suggests that prior literature has made substantial strides in understanding the extent to which computer vision has been applied in manufacturing. Extant research focal area like automation, computer-aided design, inspection, quality control, robotics and several other applications may be further explored by future studies. Moreover, analyzing customer satisfaction, I4.0, defect detection could also be explored as an emerging research area in computer vision applications. Furthermore, based on content analysis of the dynamic co-citation clusters, we recommend that the methodologies and frameworks developed by prior scholars to enable computer vision deployment in manufacturing could be applied to other contexts. This application could enable future scholars to test the multi-domain applicability of previously proposed frameworks and engender significant theoretical and practical contributions to this field.
6.3.2 Advancing knowledge for emerging nations- From our analysis of bibliographic coupling, citations, and co-authorship from a country-oriented lens, we assume that there is a presence of geographic-specificity in terms of research based in developed countries, such as U.S.A., Germany, U.K., along with a few developing countries like India and China. Further, the thematic assessment of clusters indicates that the application of computer vision to automated visual inspection, object tracking, real-time monitoring, and profile projections need further investigation, especially in the context of developing or underdeveloped countries. Thus, future research may focus on investigating the applications of computer vision in manufacturing domain in the context of more emerging and underdeveloped nations to advance the current body of knowledge.
6.3.3 Potential for object tracking and real-time monitoring- The content analysis divulged another area with a potential research interest, which relates to exploring the advancements in model training and deep learning processing for computer vision-based systems for detecting and tracking objects from a real-world scenario. In future, industry experts and academic scholars could develop a computer vision-based system, capable of performing real-time monitoring for making the manufacturing task more robust. Importantly, empirical investigations should be carried out to explore factors that may positively or negatively impact the widespread implementation of computer vision in manufacturing.
6.3.4 Cross-industry implementation – The content analysis of published articles showcases that the focus of researchers was on developing and implementing an efficient visual inspection system for a manufacturing process that can automate multiple tasks, such as defect detection, process monitoring and quality control. Some scholars focused their research on vision-based inspection algorithms for quality assurance in additive manufacturing. Manufacturing processes are usually composed of a mixed environment of large and small components, which makes selecting distinctive features even more complex. The study thereby proposes the need for research directed at empirical and analytical exploration of technological, organizational, and environmental conditions that might inspire cross-industry implementation of computer vision technology.
6.3.5 Potential for roughness inspection and profile projection- The content analysis posits the potential areas to be developed for measuring the surface roughness accurately to adhere to the required quality standard of production at a lower cost and time. In future research, various methods for dimension inspection of multiple parts can be investigated to test each component against prescribed standards to ensure the high quality of the production line.
6.3.6 Potential for integrating theoretical grounding- During the assessment of the methodologies and predominant agenda from existing literature, we noted a sparse deliberation on investigative frameworks that have been grounded in theories. We thereby recommend that in the future, scholars should appropriately infer other theories to develop robust frameworks.