It is sometimes mentioned that humans do not see with their eyes, but with their brain. The eyes are just the instrument to provide just enough information for the brain to interpret. We are excellent performers of pattern recognition and post-processing. This processing is hardly ever noticed by us, except when looking at an optical illusion like the picture below.
“Using the processing of our brain, we are able to act beyond the limits of the hardware of our eye”
The image illustrates that our vision is a clear collaboration of hardware and processing. Using the processing of our brain, we are able to act beyond the limits of the hardware of our eye or in this case deceived. There’s even more: by simultaneously processing information from our other sensing systems (tactile, hearing), our vision and sensing system can easily be deceived through for example virtual reality hardware allowing us to get immersed into rollercoasters, outer space and constructing new buildings and architectures. However, it is important not getting deceived in many applications in industrial and medical imaging systems.
For machine vision, there initially is no pattern recognition or processing present. All that we want the system to observe, must be visible by a camera or other acquisition sensor (radar, magnetic field, ultrasonic etc.). When well-constructed, they are typically hard to deceive or at least deception can be detected. Some interpolation of information is possible, but introducing large amount of processing is time consuming, often unstable and prone to errors as exceptions and assumptions need to be made. Therefore, it is better to aim for the simplest, but most accurate (i.e. just enough resolution) hardware that does the job rather than to try to apply quick and dirty fixes afterwards.
There is however an area which uses explicit processing to reconstruct information. This is called computational photography and many applications exist today like 3D stereo vision, MRI and CT systems and are extremely computational intensive.
“MRI and CT systems and are extremely computational intensive.”
Correct information extraction is vital for the performance of a robotic system which doesn’t have human interpretation & feedback systems. As an example, let us consider a pick and place robot and its vision system. The robot is used to pick individual components from their bulk package. Once the robot head has picked up a component, the orientation and position with respect to itself has to be determined. The robot now aims at positioning the component in a predefined way in order to complete an assembly step.
A human cannot detect very small objects or errors. Let’s compare the pick & place task described earlier to the assembly of a USB cable in a computer. Think of what our steps would imply.
1. Grab the cable head with the USB connector.
2. Inspect its orientation.
3. Feel or visually inspect the orientation of the mounting hole.
4. Turn the cable to achieve the correct orientation.
5. Move towards the mounting hole until a fit is found, use little pressure.
6. When the hole is found, use more pressure. Adept the pressure to what you think is required but preventing damage.
These steps require our vision, motion, tactile feedback and a lot of interpretation, just to figure out how to position a USB cable in a computer. Not to mention we often are mistaken and need to rotate the USB connector. It is highly iterative, which is not wanted in an automated system. If translating this to the pick & place process, it is apparent that for a robot, step 2 & 3 (orientation inspection) are crucial in the success of the placing process. The system cannot ‘feel’ if the component is positioned correctly, iteratively positioning & orienting is strongly unwanted. Thus, it has to position the component right the first time; as fast as possible.
These challenges can also be found in imaging systems for applications in life science in the medical domain; we want an accurate picture without any artefacts. For metrology systems we need to measure features or anomalies in micrometer or nanometer scale to ensure quality. In scanning imaging systems, vision is typically in-the-loop with motion; failing to capture the image or line (e.g. TDI line scan) at the right time causes the overall system to deliver erroneous data/images. For all these applications we face high performance and throughput requirement and above all strict real-time requirements.
At Prodrive Technologies we have an R&D group for these vision and sensing systems. We are able to support from sensor (area or line scan) to processing (x86, fpga, dsp) up to motion control and actuators.
Our competencies & knowledge base includes advanced packaging for custom imagers / sensors, wire bonding and optical assemblies in our cleanroom to meet even most demanding applications. Camera assemblies can include sophisticated cooling technologies to make sure dark currents are low as possible and mechanical tolerances to meet highest optical performance possible. When no standard optics seem fit we also do design of custom optics for illuminators and lenses to make sure the resolution and costs meet the application. Last but not least we know how to integrate or support a camera in total solutions when it comes to high-speed interfaces, triggering & motion and image or signal pre-processing.
Send us your e-mail, phone number, name and company name and we will contact you.