| I accept that there is a place for computational/algorithmic photography but I remain deeply sceptical of its actual benefits (in its current incarnation), moreover my recent bad experiences with it have only strengthened my conviction. I have previously discussed having taken photos with a smartphone where certain objects within some images have been so modified by the processing algorithm as to be almost unrecognizable so I won't repeat those various scenarios here. Instead, I'd like to dwell on the implications algorithmic image processing for a moment. Let's briefly look at the issues: 1. Despite a recent announcement by Canon about a large increase in dynamic range in imaging, (https://news.ycombinator.com/item?id=34527687), I'm unaware of any current imaging sensor breakthrough that would vastly improve both resolution and dynamic range. Thus, essentially, we have to live with what we're already capable of physically squeezing into our present smartphones. 2. Manufacturers are improving both image sensors and optics but only incrementally. Thus, with current tech and absence of truly significant breakthroughs, we have to live with the limitations as outlined in the article (aberrations, lens flare, sensor insensitivity etc.). 3. Essentially, we're stymied both by the limitations of current tech and physical (smartphone) size. Usually, to overcome such limitations, we'd fall back on the old truism 'there's no substitute for capacity' and just make things bigger as we did with photographic emulsions, past camera lenses, loudspeakers, pipe organs, etc. but that's not possible here. 4. Outside incremental improvements in hardware—the Law of Diminishing (hardware) Returns having arrived—manufacturers have had to resort to computational methods. The trouble is that it seems with the present algorithms that the Law of Diminishing (computational) Returns is also already upon us, so what does this mean? Quo vadis? 5. Clearly, in its current form computational/algorithmic processing has hit a stumbling block or at least a major hiatus. Here, further incremental improvements are likely using current methods and there's little doubt that they'll be applied to recreational photography (smartphones and such), however, unfortunately, we now have a serious (and very obvious) problem with the authenticity of images taken by these cameras. Simply, when software starts guessing what's within images then we've not only lost visual authentication but we have serious downstream issues. It raises questions about whether or not photographic evidence based on computational imaging can be relied upon—or even submitted—as evidence in a court of law (I'd reckon, without ancillary cooperating/conjunctive evidence, such images would not muster if the Rules of Evidence tests were applied. How serious is this? Clearly, it depends on circumstance but long before 'guessing-what's-in-the-image' became in vogue simple compression was 'suspect' in, for example, serious surveillance work—because compression artifacts in an image raised doubts as to what objects actually were—simply, could objects be identified with 100% certainty, if not then what figure could be placed on such measurements/identifications. (Such matters are not hypotheticals or idle speculation, I recall in nuclear safeguards a debate over compression artifacts in remote monitoring equipment. Here, authenticating and identifying objects must meet strict criteria and a failure to authenticate (fully identify) them means a failure of surveillance which is a big deal! For example, the failure to distinguish between, say, round cables and pipes with 100% certainty could be a serious problem, as the latter could be used to transport nuclear materials—thus it'd be deemed a failure of surveillance. That's not out of the bounds of possability in a reprocessing plant.) Obviously, the need to authenticate what's in an image with 100% certainty isn't a daily occurrence for most of us but as these tiny cameras become more and more important and ubiquitous then we'll start seeing them used in areas where their images must be able to be authenticated. Post haste, we need rules and standards about how these computational algorithms process images and how they should be applied. 6. What's the future. On the hardware side we need better sensors with higher resolution and more sensitivity and improved optics (that, say, use metamaterials etc.). Such developments are on their way but don't hold your breath. Computational/algorithmic processing has the potential to do much, much better, but again don't hold your breath. There's considerable potential to correct focus and aberration problems etc. using both front-end and back-end computational methods ('front-end correcting lenses etc. on-the-fly and back-end as post-image processing) but much work still has to be done. Note: such methods also don't rely on guessing. What people often forget is that when a lens cannot fully focus or suffers aberrations, etc. information in the incoming light is not lost—it's just jumbled up (remember your quantum information theory). In the past untangling this mess has been seen as an almost insurmountable problem and it's still a very, very difficult one to resolve. Nevertheless, I'd wager that eventually computational processing of this order will be commonplace, moreover, it'll likely provide some of the most significant advances in imaging we're ever likely to witness. |
Two interesting developments here are the pixels in Starvis 2 sensors, which as a first afaik use a 2.5D structure to increase full-well capacity by a lot. And another, non-production sensor by Sony where they developed a self-aligning process and pixels are actually split in two layers, with the top layer only carrying the photodiode and the bottom layer entirely dedicated to the readout transistors. That's promising for lower readout noise and also for increasing full-well capacity.