Following the identification of a window of interest, analysis is carried out to target certain features in the lab. Threshold analysis is often described as transcribing an image to black and white pixels. This not a new approach by any means but it is still considered one of the most affective forms of analyzing images. The main reason why this was so popular in the infant days of machine visioning was due to its ability to screen out the sensing systems operational variability.
The threshold approach is used in many binary systems, though these systems are not limited to a singular threshold method. If you consider the problem of identifying a snake which is grey in colour, on a black and white tiled floor for example. This approach would simply portray the image as being a black snake on a white floor, with no tiles visible at all. Maybe the automated system in question does not need to worry about this but other systems will and therefor a solution was needed. In a laboratory setting, two thresholds are set select the gray part of an image and allow the device to pick-up just that part. For this very reason, threshold systems are designed to accept two distinct thresholds, allowing the system to get around the problems outlined above. If only one threshold is required, the other threshold can be tuned to one extreme of the brightness scale, leaving that threshold completely void.
The act of finding the perfect threshold value for a binary visioning system is a key process in creating a successful vision automation system. This problem is much trickier than you may think, with many external variables affecting the outcome. Selecting a threshold that lays exactly half-way between the two light extremes can lead to a totally dark image, if both thresholds are below mid-range. Some sort of intelligent approach must be adopted to select the required threshold. Where this intelligence comes from is entirely up to the designer of the project but some sort of manual input from a human operator may be required. Although computer programs have improved greatly in how they calculate these ranges, they are still not up to the task completely. This does, however leave a system that is not fully automated and therefore cannot be classed as an automated device as such.
As you can see, there are many issues with this approach but most of these can be overcome with a little planning and design.