Optimising machine vision systems to minimise takt
CMOS sensors in the light of recent research
Machine vision has played an unarguable role driving up the quality of in-line inspection systems - and not just a subset of processes, but for a near unending set - from car inspection systems that drive down recalls to food sorting and even in industries that could previously only rely on master craftsmen. But, as the technologies improve, the focus begins to shift to ensuring integrated systems run inspection tasks cost-effectively, without false positives being sent and without slowing the production line.
Factors in reducing takt time
Takt time is the average time between the start one unit being produced and the start of the next. Minimising it plays an increasingly key role in the overall throughput of a system and therefore is essential in cutting the costs of production per unit.
Takt time can be first determined with the formula:
Where Takt time (T) = Net time available to work (Ta) / Demand (D).
There is no one way to cut it, instead the system as a whole should be considered. The image sensor’s ability to capture frames quickly is an obvious factor, but far from the only one; lighting, the image processing algorithm used, the module’s dynamic range should also be among the methods deployed.
The natural place to start is the CMOS imager itself. Those designed for consumer applications tend to maximise frame rate by having each row begin the next frame’s exposure after completing a readout. This means each row is exposed for the same length, but each timeframe is shifted slightly, and so the row at the bottom of the sensor is almost a whole frame behind the top row. While this is allowable in consumer applications, it is problematic for fast-moving objects, with a clear spatial distortion visible so the top of the captured object appearing to be behind the bottom.
A global shutter architecture, used in virtually all CMOS imagers designed specifically for machine vision applications, allows all pixels to be blanked and activated at the same time. This prevents the shifting that can result from a conventional CMOS shutter design.
Consistency is everything in machine vision, and therefore lighting plays a key role as it maintains the consistency (and therefore repeatability) of the image captured.
For exposure time to be as short as possible, you can either set the captured area’s lighting at a high level, or increase the light sensitivity of the image sensor itself, but this is just one aspect - and not without its own challenges as high levels of lighting can lead to glare in other parts of an image.
Therefore, the direction of the illumination is a similarly vital step in creating a strong contrast between the captured component and the background.
Next, maximising the uniformity of light source will reduce takt by reducing the amount of post-processing required by the image-processing software used to detect each mounted component or product feature for alignment and quality.
Additionally, the required level of illumination balance across the subject may be exceptionally difficult to achieve over an entire image, with key regions of the product / sub-assembly lying in the shadows of other components.
Finally the system’s lighting should address issues that may arise through vignetting, where the illumination level as captured by the sensor varies across the lens, or the need to mount the lighting very close to the object being examined could result in difficulty achieving consistent coverage. In many cases, this leads to areas in some parts of the image appearing subtly lighter than others.
Image processing algorithms
Once lighting has been considered, the image processing algorithms should be next. These can be used to partially correct problems, but they also increase computational load, which slows the system and can also make it more difficult for the detection processes to ascertain if a detected defect is real or an artefact of illumination.
However, by controlling the capture at the pixel level, and through the use of high frame rates - achieved through using CMOS imagers - it is possible to overcome the problems of illumination consistency almost universally.
Dynamic range and multiple captures
A high dynamic range enables shading correction on parts of the image, enabling better recognition without the loss of effective bit depth that would normally be encountered using single-exposure images.
This is an increasingly common technique in photography and involves the capture of multiple shots, taken in sequence quickly but using a different exposure time for each.
The resulting images are then combined with the composite photograph having a much higher bit depth than a single image and enabling a system to capture areas that would otherwise be lost through high brightness or darkness.
Multiple captures also has an additional benefit of increasing the image’s overall sharpness. This counteracts the effects of, for example, heat shimmer, which can cause sharp lines and points to appear to move over time - causing blurring on long exposures and shifts in position from product to product on subsequent short exposures.
The effect of this for the software is in false positives, interpreting the visual effect as the result of a manufacturing process being out of tolerance. If this happens products and sub-assemblies are typically either sent for time-consuming re-inspection or scrapped, which either way is expensive.
Reference frames and look up tables
Modern image sensors can also be employed to further enhance the detail of a shot by altering the effective brightness of pixels that lie in areas not lit as effectively as others. This is achieved by calibrating image capture with a reference frame and - taking this further - multiple reference settings can be stored to allow for multiple light sources in successive captures, which can then be adopted to highlight different parts of the object under inspection.
Similarly, look-up table (LUT) support and pixel defect correction provide further mechanisms to deal with issues that arise from lighting and the image sensor.
By shifting the gamma of the image via the LUT, it’s possible to make full use of the bit resolution of the image sensor’s output stream and optimise the contrast within the image. Finally, defect correction can be employed, using the pixels around a failed one to complete the picture - meaning pixels stuck at one value are not interpreted as issues on the product being photographed.
Regions of interest
A further factor in streamlining overall system performance is the ability to select specific regions of interest. This effectively programmes the image sensor to send only portions of an image, and therefore the use of network bandwidth can be optimised.
This, in turn, allows more imaging subsystems to be employed so more images of a product are captured and more sensors deployed along the production line so an individual processes can be monitored more effectively.
To use a sporting analogy, the process of minimising takt can be summarised in the words of the Rugby World Cup winning coach, Sir Clive Woodward, who said “Winning the Rugby world cup was not about doing one thing 100% better, but about doing 100 things 1% better.”
There is no one way to cut takt, instead the system as a whole should be considered. The image sensor’s ability to capture frames quickly is an obvious factor, but far from the only one; lighting, the image processing algorithm used, the module’s dynamic range should these all come together to create marginal gains, which while individually small, combine to improve the takt time significantly.
By bringing together a number of techniques and technologies, image sensors designed specifically for machine vision can help integrators and end users build machine-vision systems that continue to deliver improvements in throughput and takt time in industrial systems.
Vision: Halle 1, Stand D 31