Artificial Intelligence (AI) is the science and engineering of making intelligent machines or systems that can learn, at least somewhat as human beings do. Machine learning is one of the areas of AI that focuses on developing systems that can learn to make decisions or predictions and improve from experience without being explicitly programmed. Machine learning and, precisely computer vision, have demonstrated performance near human-level in many real-world applications owing to the transformative developments in the field of AI. As a result, computer vision algorithms are replacing human supervision in domains such as manufacturing process control Computer vision enables computers and systems to process, analyse, and derive meaningful information from visual data such as images and video.
Machine Vision
One practical application of computer vision in the machining industry is machine vision. Machine vision is a practical realisation of computer vision techniques in solving practical industrial problems involving a significant visual component such as measuring the size and position of objects, assessing their surface quality, identifying defects or contamination, sorting, checking the integrity of safety-critical components and assemblies, or guiding industrial robots [1]. According to a study by McKinsey & Company, machine quality inspection can increase productivity by up to 50% and defect detection rates by up to 90% compared to manual inspection. For instance, on a production line, machine vision systems can inspect hundreds or thousands of parts per minute reliably and repeatedly, far exceeding the inspection capabilities of humans[1].
Machine vision systems can be categorised as conventional (traditional) or modern (deep learning). Traditional machine vision requires well-established logical steps (rules) to realise industrial machine vision applications. They operate via step-by-step pre-processing and rule-based algorithms that are more cost-effective than human inspection. Unlike traditional machine vision, deep learning automatically discovers and extracts observed patterns from data with minimal domain knowledge and human effort engineering. It combines the flexibility of human visual inspection with the speed and robustness of a computerised system.
Yet, both traditional and deep learning-based machine vision are complementary technologies with overlapping capabilities. The choice between them will depend upon the use-cases, amount of data being processed and processing capabilities.
Consider the example of automated vision inspection in a manufacturing scenario where a CAD model of the object detected, measured, or inspected is immediately available. In most cases, the type and range of defects are known in advance of the inspection system. In addition, considering the manufacturing process environment, the position and pose of the objects to be inspected can also be established in advance and readily fixed when viewed by a camera. All these reduce the need to accommodate change and uncertainty within the inspection process, and thus a traditional machine vision is a perfect choice for this scenario[1][2]. Consequently, conventional machine-vision technology has remained popular in the manufacturing factory due to its proven repeatability, reliability, and stability[3].
For other machine vision applications such as assembly verification, it is notably challenging to program a rule-based machine vision system due to multiple variables that can be hard for a machine to isolate, e.g. lighting, changes in colour, curvature, and field of view. Due to scale, rotation, pose distortion, complex surface textures, and image quality issues, some variability in a part's appearance makes the task inherently tricky for the hand-crafted rule-based conventional machine vision system[1][2]. As a result, the emergence of deep-learning technologies allows for expanded capabilities and flexibility to such challenging use-cases. The recent report by Landing AI, argue that “Deep learning is likely to offer so much potential that deep learning- based, machine vision techniques in smart manufacturing will see an annual growth rate of 20% between 2017 and 2023, with a revenue that will reach US $34 billion by 2023.”[3]
Components of Machine Vision System
A typical Machine vision systems consist of three main components: image capturing, image pre-processing, and actuation [4].
First, the image is captured using vision sensors such as digital, ultraviolet, or infrared cameras to capture the image and transform it into the desired format for the processing step. Then the pre-processing components perform the desired image analysis and processing to produce a more suitable form for subsequent operations. Other necessary pre-processing steps include segmentation, feature extraction and classification[2], [4]–[6]. Segmentation seeks to partition an image into meaningful regions corresponding to part or whole objects within the scene. In contrast, feature extraction, in general, attempts to identify the inherent characteristics, or features, of objects found within an object. On the other hand, pattern classification identifies an object within an image. These four pre-processing steps are necessary for conventional or classical machine vision systems to realise almost any industrial machine vision application task. On the contrary, the pre-processing steps is greatly simplified for deep learning-based machine vision systems where steps like segmentation, feature extraction, and classification could be combined as one step.
The actuation provides machining control and decision trigger mechanisms for the received output from the vision processing and analysis stage. It also includes user interfaces for the integration of multi-component systems and automated data interchange.
Challenges for Adopting Machine Vision
Despite the potential of machine vision systems in machining companies, several challenges, such as difficulty to collect data, scalability, integration and supporting continuous learning, limit the wide adoption. For most use-cases, such as visual inspection, it is challenging to collect data to train the machine-vision algorithms. As a result, most machining vision projects end up with insufficient data or low-quality data. To this end, it is advisable to focus on getting quality data while exploring other data-centric approaches such as data augmentation, transfer-learning for improving the performance of the machining vision systems. According to the recent talk by Andrew Ng from Landing AI, having a data labelling pipeline that ensures consistent data labelling and analysis is a key for a successful machining vision project. Other learning strategies such as semi-supervised or self-supervised learning could be leveraged to address the challenge of limited data.
Scaling machine vision projects from proof-of-concept to initial deployment while maintaining the pilot performance in the production processes is another challenge. Consequently, it is advised to have a system to provide continuous monitoring and improvement of the deployed machine vision systems while supporting continuous learning for AI-based vison system
The complexity of integrating AI within existing infrastructure is also another challenge that stalls down the adoption of machine vision in machining companies. This is mainly due to the need for suitable software and hardware infrastructures to power the machining vision solution. Thus it is advised to align the hardware and software investments needed to support the machine vision project.
Summary
The transformative developments in AI, particularly computer vision, have demonstrated performance near the human level in many real-world. For instance, computer vision applications are replacing human supervision in several machining processes from visual inspection and quality control to guiding industrial robots. This post presents an overview of machine vision, the application of computer vision in manufacturing, the component of the machine vision system and challenges for adopting this technology in the manufacturing space.
References
[1] Cognex Corporation, “DEEP LEARNING FOR FACTORY AUTOMATION Combining artificial intelligence with machine vision,” 2020, [Online]. Available: https://www.cognex.com/resources/white-papers-articles/deep-learning-for-factory-automation.
[2] M. L. Smith, L. N. Smith, and M. F. Hansen, “The quiet revolution in machine vision - a state-of-the-art survey paper, including historical review, perspectives, and future directions,” Comput. Ind., vol. 130, p. 103472, 2021, doi: 10.1016/j.compind.2021.103472.
[3] “2020 State of AI-Based Machine Vision,” 2020.
[4] H. Golnabi and A. Asadpour, “Design and application of industrial machine vision systems,” Robot. Comput. Integr. Manuf., vol. 23, no. 6, pp. 630–637, 2007, doi: 10.1016/j.rcim.2007.02.005.
[5] N. Chen, J. Sun, X. Wang, Y. Huang, Y. Li, and C. Guo, “Research on surface defect detection and grinding path planning of steel plate based on machine vision,” Proc. 14th IEEE Conf. Ind. Electron. Appl. ICIEA 2019, pp. 1748–1753, 2019, doi: 10.1109/ICIEA.2019.8834354.
[6] D. Frank, J. Chhor, and R. Schmitt, “Stereo-vision for autonomous industrial inspection robots,” 2017 IEEE Int. Conf. Robot. Biomimetics, ROBIO 2017, vol. 2018-Janua, pp. 2555–2561, 2018, doi: 10.1109/ROBIO.2017.8324804.
