Today was the big technical day of the ISO photography standards meeting that I’m attending virtually. We had presentations and discussions on the topics of standardisation of measurements of camera imaging noise, resolution, autofocus repeatability, depth metrology, image flare, as well as standardisation of Adobe’s DNG file format, and a presentation on new work by JPEG.
Much of it was very technical and probably not very interesting to most people. However the autofocus presentation had some fascinating experimental results. The presenter had at first assumed we could do image statistics to determine the best focused image from a series of photos taken by a camera. Defocus blur smooths out the image, so the variance in the pixel counts is lower, which means that if you measure the variance in a photo (of the same subject, at the same light level, taken by the same camera), then the image with the highest variance should have the best focus.
However, doing an experiment in which he measured hundreds of images, he found that sometimes when the autofocus failed and the image came out blurry, it actually had a higher variance than in-focus images. The reason was that the camera added artificial image noise as an image processing step. The reason it might do this is because it’s known that slightly blurry images look sharper to human eyes if a little bit of image noise is added. So the camera has been designed to add some noise, to fool human users into thinking the photo is sharper than it really is. The result of this is that when a photo is truly out-of-focus, it adds so much noise that the variance ends up higher than an in-focus image. (This was a phone camera that was being tested, by the way, not a DSLR.)
So to make our standardisation of a method to measure autofocus workable, we have to deal with this artificial image noise that some cameras add to the image, and we can’t rely on the image statistics being sensible and based merely on the physics.
This sort of thing is becoming more and more of a problem for us in this work. Measuring the performance of a camera is getting more complicated because of all the post-processing that modern cameras (particularly phone cameras) do to make the image look “nicer”. Even a conceptually simple thing like defining the exposure time of a photo is riddled with complications caused by cameras that take multiple exposures when you press the shutter button, and then combine different parts of different images to produce a composite final image. For example: some areas of the resulting photo might have pixels taken from an exposure with one exposure time, while another area has pixels from an exposure with a different exposure time, while another area has pixels that are an average of two or more different exposures, and then the brightness levels might be adjusted in different ways. At one extreme, there is no single “exposure time” that physically describes what is represented by the pixels across the whole photo, and at the other extreme to fully describe the “exposure” you need to list an array of different exposure times and their blending coefficients for every pixel in the image. While that would be physically correct, it’s obviously impractical. We still haven’t figured out how to address this issue.
Another interesting thing came from the JPEG presentation. JPEG is not just an image format, it’s a large technical committee (separate from the ISO Photography committee), working on a lot of new stuff related to image encoding. Their representative was giving us a report on recent work they’re doing. One thing I thought was interesting is a new project to add privacy controls to images. Say you want to share a photo of yourself on social media, but you don’t want random strangers seeing your face. This JPEG project is working on a way to select a region of a photo (e.g. your face), and encrypt the image data for that region, so that a person without the key can see the background but where your face is it just displays a blurred/pixelated version, but a friend who has your encryption password can see the original photo with your face. (I described this to a friend of mine and he criticised the idea as unnecessary complexity, as there are already ways to achieve basically the same effect without building encryption into JPEG. I’m no expert in file encoding, and I suspect there’s more to it than that, but *shrug*.)
Anyway, this is kind of all I did today – this sort of highly technical stuff. One more day of the meeting tomorrow. There’ll be a bit more technical discussion, followed by administrative stuff. (And I’m not getting paid for any of this…)
Oh, the other thing I did today was go to teach my Ethics class this morning. I had time to do this because the virtual meeting is running on Tokyo time, so it started at 11 am Sydney time. So I had enough time to go teach my class. However, when I was set up and ready to go, and the school bell rang… no students showed up! I had to go find a teacher, and they told me that Year 6 was away on camp this week! So I packed up and headed home. Oh well… next week!
New content today: