Machine vision in industrial production: perspectives and trends

According to Mr Evgeny Vesnin – Technical Director at Mallenom Systems LLC, the use of machine vision technology becomes increasingly popular today. The development in the areas of Deep Learning have led to the spread of machine vision technologies into multiple areas of daily life. At the same time, however, the industrial quality control is still the main area of application for this technology.

Mr Vesnin continues by indicating that metallurgy, food, household chemicals, automotive and mechanical engineering, as well as pharmaceutical production are the main areas that implement the machine vision. These are the areas that require exceptionally high quality control for the manufactured goods.

Camera as a foundation

The cameras are the main part of machine vision system. They allow to receive high quality images that are crucial for an optimal for processing, analysis, taking measurements, recognition and control performed by a computer.  

Transfer of the generated uncompressed images. The advantage of this approach is the absence of information loss which is crucial for the surface visual quality control systems.

Compatibility with the real-time systems. A transfer of uncompressed data allows to receive an image for its further processing and analysis at a fraction of a second, which is necessary requirement for a successful production line operation.

Wide resolution spectre. Depending on the technology of installed sensor, industrial cameras can be either of line scan or area scan variety. For area scan cameras, the resolution is between VGA and 155 Megapixels, while for line scan cameras it could be up to 32K.

High frame rate (up to few hundred thousand frames per second). This allows machine vision cameras to monitor the high-speed production processes. 

Generation of input/output signals. Almost all machine vision cameras have an input/output interface to connect with any external devices.

Should the machine vision camera be smart?

Usually, the machine vision cameras are viewed as the sensors for image generation. Today, on the market, there is a wide variety of camera models that differ in interface, dimensions, utilised sensors for image capturing, etc. There are multitude of tasks that suit the machine vision cameras. It is obvious, however, that development of a specialised video camera for each and every task is not financially sound, since it would require to spread significant development costs among relatively small number of niche cameras.

But at the same time, for some tasks that are in high demand – for instance, 1D and 2D code reading, number plate recognition – it is reasonable to integrate the video signal processing directly into a camera. This is supported by a relatively high supply of these cameras that exists on the market. Also, this integration is reasonable for the solutions that must be lightweight and have an efficient energy consumption – areal drones for instance. For the majority of industrial applications, however, the machine vision camera must be versatile and be reasonably priced.

Taking into a consideration that different tasks require different amount of processing power and they are performed across multitude of computational platforms, it may be reasonable to have an open-source computational platform for developers that support connection of machine vision cameras via low-level interfaces. This would enable to have a financially reasonable solution for each unique task.

Machine vision is in high demand in metallurgy, mechanical engineering, pharmaceutical and food industries, and other areas

New technological solutions

It worth to note that both hardware and software determine the efficiency of video signal processing. In the past few years we have observed a considerable development in both. From a software side, it worth to mention OpenCV open-source library. While, for a hardware, such products as Nvidia Jetson Intel® Movidius™ Myriad™ X Vision Processing Unit deserve a special recognition.

Not so long ago, Basler has created a developer kit which includes computing boards with interfaces for receiving signals form cameras and SDK for different languages. This kit allows to create a solution in form of an open source-source computational platform. Basler also supply industrial cameras that support low-level connectivity to these boards.

At the end of 2019, Flir has produced the first line of Firefly DL machine vision cameras with specialised neuro processor which allows to initiate some Neural Network configurations directly inside of a camera.

For a long time, Cognex develops and produces all-in-one machine vision systems that consist of smart video camera and processing unit with the outer industrial body of a unit to withstand a variety of tasks. Also, these units come with multitude of instruments for analysis, including on-bard utilisation of Neural Networks.

In the last couple of years, we also have seen an inflow of solutions that combine a video camera and processing unit in a single industrial body from such producers as ADLink, XIMEA, Hikvision and others. Among the companies that develop a budget-friendly open-source platform of industrial vision, one should note OpenMV. This producer develops inexpensive but expandable solutions for machine vision.

Line scan cameras: do they have a future?

Majority of tasks for industrial control utilise the area scan cameras. However, some instances may require the use of line scan cameras. Let us consider an example of defect control of rolled metal or other continuous materials. With the use of common area scan cameras, the problem of sufficient even light for a large area of moving material occurs. Moreover, the illumination of a such large area would result in a substantial electrical bill and in a need for more space on a production line. Reliable quality control would additionally require to eliminate vibrations of a material in the inspection area, which is often quite difficult to accomplish.

The utilisation of line scan cameras allows to inspect produced materials directly on a shaft of the production line which eliminates the problem of material vibration. Additionally, specialised focused lights for line cameras provide an efficient way for illumination of controlled area. If the task requires the control of material transparency, such examination and elimination of vibrations is also easier to accomplish with the use of line scan cameras. 

The main advantage of the line cameras is their compact depiction of data. For instance, with the use of area scan camera for material movement control on a production line, the problem of excess video information may arise due to some frames having the exact same area of the production line which is slightly offset in time. The line scan cameras could be set in a way to generate scans for a movement of a conveyer belt by 1 millimeter for instance. In this case, it would be possible to construct an image from the scans which covers entire surface area of inspected material or production line in relation to the position and without unnecessary overlap. 

Machine vision for industrial safety tasks

A combination of machine vision cameras with surveillance cameras may be very desirable. For example, we have had an experience with some projects where the utilisation of common IP-cameras was impossible due to the lag that occur during formation and compression of a video stream. The delay between the time the frame is formed and the time when this frame is received by the processing device can reach 5 seconds, which is unacceptable for the personal and production processes safety control tasks. 

At the same time, the customers of such systems require to archive the video data. This, in turn, requires a development of specialised services for coding of video streams form machine vision cameras, which makes it possible to use commercially available video recorders for archiving and accessing the video footage.

Mr Evgeny Vesnin

Technical Director at Mallenom Systems LLC

Published in October-November issue of Safety Systems journal