snapsy wrote:
There are a lot of articles about stacking - it's been around as long as digital imaging. I imagine there's a lot of recent interest from the E-M5 II's built-in superresolution mode.
That's what go me interested. As you mentioned this has been around for years. I first used this technique to gather more DOF in macro shots, by stacking photos with different focus points. That was maybe 7 or 8 years ago. When the EM5II came out it got me thinking about the technique again.
Stacking, and of course more "intelligent" ways to manage stacking, is actually one of the few ways forward in image quality except for the ever elusive organic "full color sensor" that actually measures real human-related color (not Foveon then) in all pixels all over the image surface - and does so with reasonable read noise levels (Foveon excluded once again). Bayer filtering takes away at least 70-75% of the available light to work - really suboptimal, but the best we currently have available for mass production.
The only really interesting solution I've seen for the last three years since the Fujinon/Pana organic sensor is actually software based.
You basically start from a high-speed filming device with very low read noise figures, and then you stack in as many images as you can fit within a certain threshold of trackability.
The first problem is having an effective image registration algorithm that matches up image detail spatially between consecutive images and compensates for internal detail movement between the frames, preferably in hardware. No HW solution for this exists outside really esoteric surveillance equipment ATM.
The second problem is how to handle what we now see as "movement blur" in a still image. This effect is actually desirable from an artistic PoV in many cases, since it provides movement clues within a single still image frame. When should it be negated (by auto-registering image detail) and when should it be allowed to show up as movement blur?
The third problem is the viewfinder principle to use... real-time processed data, or an optical zoomable VF? RT data has considerable lag due to both the processing and the needed stacking depth/time.
But the principle is really interesting. We've been looking at implementing a very simplified version of the principle for number plate registration in tollbooths and so on. Five-six consecutive noisy and under-resolved images can actually stack to a readable number-plate in many cases, and we've seen improvements in identification accuracy by up to 300% (at the cost of one server-class processing station for every additional ten feeds or so). For clients, this means that they can keep existing camera equipment, increase fps slightly, and "just" buy more processing power. This may sound expensive, but compared to a complete change of capture hardware on-site, it's peanuts. Not even coffee-break handout costs.
A lot of nay sayers, but as a climber / photog I know exactly where you're coming from. I've tried many other systems including the A7 series, but even that system isn't small enough or light enough when it comes to climbing. This technique is perfect for those rare moments when you need maximum iq in the moment, while climbing. Thanks for sharing!
