A common request for those of us who work in video analysis (whether privately or for a law enforcement or government agency) is to enhance a license plate or compare the person (face, article of clothing, etc.) in a video to a photograph of a person. When is there adequate information to expect that various methods of enhancement may generate value? I develop rules of thumb for these situations.
I think it makes sense for analysts to experiment with files to develop their own rules of thumb as well. In this post I’ll share some of the methods I’ve used to develop my own rules.
License plates. The first thing I did regarding license plates was to determine how many digits may be in a plate, and what the average, minimum number of pixels required are to display the plate. If the state or country uses seven digits, and we consider that most digits require a minimum of three pixels horizontally and one pixel between digits, then a minimum of 27 pixels can, theoretically display most letter and number combinations. If the plate has a couple of W’s and M’s then 27 may not be enough, and if it has fewer digits, or if all the digits are the Letter ‘I’ then fewer pixels would suffice. So, that’s a starting point, but there is more to this.
Do real world examples match this theoretical approach, and are there other things to look for that may be of value? How do compression, lighting, and recorded tonal values affect this? This is where one needs to experiment. You can either take high quality photographs of license plates, or download them from the internet. Then, make multiple versions of these at a variety of pixel resolutions, at a variety of compression settings, at varying degrees of reduced contrast. Save these files and begin analyzing them (starting with the worst) to determine what observations you can make at what settings, sizes, etc. For instance, can you determine the colors in the plate at 20 pixels wide? What resolution enables you to determine the difference between a dealer plate and a standard plate? How many tonal values are required before you can distinguish various characters? As you do this analysis, take notes and see if you find consistency from one plate to another. Start to make other observations – such as when can you differentiate between a ‘5’ and an ‘S’ or a ‘1’ and an ‘I’?
Follow this same approach for other items. For faces, I do a few things. I download mug shots of celebrities, save photos of myself, family, colleagues, and friends, and download random photographs of strangers. As with the license plates, I save multiple copies at varying resolutions, degrees of contrast, compression settings, etc. Then I make my observations. How small of an image shows me an accurate rendition of hair and skin color? When can I determine eye shape or nose shape? When do blemishes begin to show? How about tattoos? How much contrast is needed for a tattoo or blemish to show?
By doing these experiments, we can each have a fairly good idea of what level of resolution, quality of lighting, degree of compression, etc. is necessary to get what level of value from the images we work with daily. We can have some rules of thumb, and we can provide realistic expectations to our colleagues or clients when they bring us images. We can easily explain not only what those realistic expectations might be, but also why, and show them examples to help them understand what limitations may exist.