Ce truc est pas mal mais plus basé sur le fric que sur la technologie:
Why Is 16 Megapixels As Good a Resolution as Digicams Can Get?
Mr. Mead said that because of fundamental size limits in the wavelengths of light, it is unlikely that future digital sensors will gain much additional resolution. [Mr. Mead is head of Foveon 16 MP Chip Designer] N.Y.Times article
Digital Chip Surprises
The above quote from Carver Mead, the developer of the 16 megapixel Foveon CMOS chip, will probably surprise many digital camera fans. Chips can't get much smaller. At smaller than today's 0.18 micron technology, small errors or faults on the chip can result in low yields and high costs. The sensors become too small to intercept enough light to produce a noise free signal. So chips are running into physical limits.
Foveon's 16MP chip uses 0.18 micron technology, versus 0.35 to 0.5 micron features for their current 3.1 and 6 MP chip competitors. But we are near the end of what current CMOS technology can do, now and in the foreseeable future. The flip side is that at 16MP, the results will be good enough for most current disposable camera and point and shoot users. So there may also not be much market pressure for higher density chips, just like there aren't many large format users today either.
Foveon's 16MP chip is also 22 by 22 millimeters square(!), which is a problem for rectangular format fans such as 35mm SLR's familiar 2:3 ratios. Simple geometry suggests the 22x22mm chip will only image 56% of the 24x36mm format provided by conventional 35mm SLRs. Using your 35mm SLR lenses would result in a chip-based cropping of the 2x3 rectangle to a 1:1 square - surprise! Your wide angle lenses will become much less wide too after this digital cropping. Things will be worse on medium format, with the 22x22mm chip only imaging 16% of the format area. So a very wide and heavy 40mm Hasselblad lens will act like a 100mm short telephoto lens. Too bad for us wide angle fan(atics)!
Foveon's 22x22mm square chip has 16.8 million sensors, representing 4,096 x 4,096 sensors on a square grid. Given 4,096 sensors on each 22mm axis, you must have circa 186+ sensors per linear millimeter (i.e., 4096/22=186+). You need two rows of sensors to image a line (black/white). So 186 sensors per millimeter corresponds to 93 lpmm (186/2). The sensor can only achieve such maximum resolution if you were to optimally and carefully align the lines with the image grid. In practice, the alignment would be more or less random, and you would only get about half the maximum resolution or about 45 lpmm (cf. Nyquist sampling limits).
Strange as it may seem, the big problem with most optical systems for digital sensors is they are too good. Lenses have aerial resolutions that can run as high as 300 to 600+ lpmm. A low pass filter setup is used to reduce this high frequency response so it doesn't cause problems (such as aliasing). You may also have an infrared filter in front of your sensor, so it isn't adversely affected by invisible IR radiation to which many silicon based sensors are very sensitive.
Another interesting problem with using existing 35mm SLR lenses designed for film use is that digital sensors are not flat. Instead, the digital sensor's active site is inside a sensor "well" or cavity. The walls of this cavity keep light from the side from easily reaching the sensor below. The ideal lens for digital sensors would provide a parallel bundle of rays (rather than converging) from the rear of the lens onto the sensors. Light from very wide angle lenses in particular come in at a severe angle to the chip surface from many points on the lens. This light can be blocked, providing yet another problem when trying to use current 35mm SLR lenses on digital sensors.
This problem is currently masked by the small chip size of most sensors (e.g., 22 by 22 mm for the Foveon 16MP chip). The chip is only seeing the center of the image circle, often after passing through additional optics plus the front filters (IR, low pass..) on the digital chip surface. We should mention that filters behind the lens cause focus shifts (equal to 1/3rd their thickness, usually) (see Filter FAQ). Other problems like flare can be made worse by rear mounted filters too.
Medium format users would be cropping to the same 22mm x 22mm chip size. So even if we could get such a 16MP chip in a 6x6cm digital back, we would get only 1/6th of the image, and at less than optimal resolution. You can add optics to focus the 56x56mm image onto a 22x22mm square. But besides the cost, you also have to expect rather lower total resolution. The resolution of your optics onto the 22x22mm square is already marginally low, and doesn't get full benefit from the chip's potential resolution. Since the big advantage of medium format is larger film area, using a small chip defeats this expectation. You might as well use 35mm SLR optics and save the weight and cost, since the chip size and optics are the limiting factors.
My bet is that the 16MP and larger chips will be made as small as technology limits (chiefly noise) will allow, and the lenses will be sized to match. Fortunately for users, small lenses such as those used for microfilm cameras can have very high resolutions at relatively low cost (e.g., 250 to 350+ lpmm). It is much easier to improve a small lens and minimize aberrations than in a large one. On the other hand, diffraction becomes a big problem with small lenses very quickly (e.g., past f/2.8).
So I would predict very small, lightweight, and fast optics. I think fixed wider angle lenses will be popular, while "zooming" will be done digitally using interpolation. The lenses will be fast because the smaller size of the sensors will make it hard to avoid noise unless you have a lot of signal (light). Sheets of microlenses looking like bug eyes will help focus light from the entire chip surface onto the limited light sensitive area of the chips (e.g., 30% of chip area).
I am not saying 24x36mm or even 56x56mm (6x6cm) or larger chips won't ever be made. I am betting that those larger chips will be custom production runs, for a very limited (in chip maker terms) market of 35mm and larger camera users. The really low cost mass produced chips will not be aimed at the relative handful of us owning 35mm SLRs. The current Foveon 16MP CMOS chip maker (National Semiconductor) CEO is even talking about millions of cheap 16MP sensor chips added to portable videophones and other gizmos including disposable 16 MP cameras (actually, recycleable is a better description).
Now do you think that you are going to lug around that medium format camera and lenses, or your ten pound bag of 35mm SLR bodies and heavy zoom lenses and tripods, or a six ounce $100 16 Megapixel recycle-able camera? Remember, both will deliver the same 16MP resolution. The tiny lens on the disposable camera may even have lower distortion than those oldie Zeiss or Nikon optics which weigh much more. After all, you can use digital technology to map the distortions on the lens and then correct for them in software (but not in film). Do you really think you will carry around all that obsolete glass, or just use the 16MP sensor in your video digicam or videophone and upload directly to your server and home printer?
My bet is that another ten years will have most current 35mm and larger film format cameras seem as heavy and unappealing as a wooden Kodak 5x7" view camera.
But the bright lining in this digital cloud is that film is likely to remain the high quality choice in the future as it is today. The 16MP chip developer, Carver Mead, is quoted as saying that it is unlikely that digital chips will gain much additional resolution and may already be pushing the limits. So to get more resolution, they will have to make bigger chips, but that will cost much more due to lower chip yields and limited market demands. Digital camera chips are really only affordable when they are mass produced, and that requires a mass market. Few users today have a need for quality beyond 35mm, as lagging medium format and large format market shares show. So folks who want a high quality image and larger prints will need to turn to film to supply that quality for at least the foreseeable future.
Chip Benefits
Some digital camera sensors may have extended light sensitivity ranges over some films, including into the infrared and ultraviolet range. However, the glass (or plastic) lenses usually used with most digital cameras will often limit this IR and especially UV response range. You can select films which have extended IR (or UV) film sensitivity too. Most digital cameras block this capability, to prevent the camera from being fooled by infrared light. So if you want to do IR photography you may be lots better off just using IR film in a regular camera.
Most films are limited to a dynamic range of 7 or 8 stops in practice, for a light range of 1:128 or 1:256. Silicon sensors are capable of much greater dynamic ranges. However, most prosumer digital cameras limit the range of response to a rather narrow range for best picture quality. Under challenging conditions of lighting or subject matter, you may have to reshoot after deleting the bad shot seen on the tiny camera back mounted LCD screen. Many cameras have only a limited ASA or film speed rating range in which this response range can be shifted.
By contrast, you can select film speeds from ASA 0.6 to 32,000. Film reciprocity makes it possible to adjust exposures (and filtering) for longer exposures, up to some hours long for moonlit landscape photos. You can't do this with digital sensors, unless you intend to cool them in liquid nitrogen. The sensors build up noise quickly, so good images can only be achieved for a limited range of short time durations. So certain kinds of long exposure time photography can not be done with digital cameras due to these sensor noise accumulation problems.
Silicon sensors are very much more efficient with low level light than film (however, a 10X or 1,000% faster fine grain film is in the offing, see below). Here again, most digital cameras limit the range of film speed equivalents to only one or two film speeds. Most modest cost consumer digital cameras have a fixed film speed. So what could be a benefit of digital cameras is lost against the ability of film users to pick a range of film sensitivities up to 6,400 ASA and beyond. Making this worse are the new print films with multiple light sensitivity ranges from 100 ASA to 1,000 ASA/ISO ratings in the same film.
Because of their dynamic range, silicon sensors should have a wider latitude of exposure in theory than film. Unfortunately, in practice most digital consumer cameras have an optimal lighting range for getting good pictures that is not very large. By contrast, color print and black and white films have a rather wide range of exposure latitude over which it is possible to get an acceptable print out of the film. Even slide films have a one or more stop range of over or under exposure.
Most consumer digital cameras respond poorly to under or overexposure. Most models seem to work best with a limited range of light (e.g., sunny f/16 daylight conditions). The flash units on lower end digicams are often automatically triggered to supplement even relatively bright light levels. The higher end and more costly digital cameras provide more of the potential range of silicon sensor benefits to the user. But even here, the film user can choose from a wider range of ASA film speeds and often greater latitude in under or over-exposures too.
Digital camera users often proclaim that they can make as many exact digital copies as they wish. That's important, since the rapid obsolescence of storage and computer and digital camera technology means a new generation of each is out every 12 to 18 months. Many digital users ignore the costs of these upgrades. They also ignore the time and labor it takes to organize, backup and convert their images to the latest formats and storage systems. Backups are particularly critical since destructive viruses can potentially destroy your entire online digital photo collection.
If you maintained your digital photos online, this practice would help backup your images automatically via the file server backups. But most web sites are limited to 10 megabytes storage, or 100 megabytes at best, with charges often related to the number of stored megabytes. For high resolution image scans (e.g., TIFF), it won't take many images to use up 10 or 100 megabytes of storage. So your costs for maintaining these files online represent yet another hidden cost of going digital for many users.
You can use gigabytes of local storage instead. But you have to have a way to backup those gigabytes and do so often enough not to lose data in system or virus related crashes. Digital files are subject to the various risks of film and prints (fires, floods..) but also add their own risks. Having seen lots of virus related crashes this last year alone, I suspect that digital photo files are much more at risk than traditional film and print media, especially in the home (non-professional backups) environment.
Aliasing - Or Why Film Grain is Superior to Grainless Digital Cameras
Many digital camera advocates believe that the lack of grain in digital sensors makes them superior to film emulsions due to film grain. The opposite is true; film is superior to digital sensors because it has a random pattern of varying sized film grains.
Digital sensors are subject to a rather troubling problem called "aliasing". Aliasing can occur because digital sensors are precisely deposited arrays of sensors and silicon features in a regular grid structure. This grid pattern can interact with regularities or patterns in the image to produce a series of artifacts on the digital imaging. The most familiar example are the Moire patterns and color fringing artifacts.
To try and prevent this troubling problem, most higher end digital cameras have extensive image processing and aliasing detection software. This software tries to guess when you have an aliasing problem (versus when you are using a diffraction grid filter, say). Then the software tries to guess how best to process the image to remove the aliasing artifacts in the image. Depending on the software and aliasing image, the result can be very good to very poor.
The solution to an aliasing problem is simple. You need to convert the regular array of identically sized sensors into an irregular, randomized array of sensors of different size. This solution to digital aliasing is precisely what we have with film grain - an irregular array of different sized grains of light sensitive elements.
Dynamic Range - Film versus Digital Sensors
A typical film grain may contain circa 20 billion silver atoms. A photon of light may interact with some of these silver atoms to cause a micro-clump of silver atoms (e.g., 4 to 10+ atoms) in the individual film grain. So you can have a range of density variation in the film grain ranging from no exposed silver atoms (zero clumps), to one clump, two clumps, on up to 4 billion or so clumps of silver in a totally exposed grain of silver. Now 4,294,967,295 is 2 raised to the 32nd power, so our 4 billion clumps corresponds roughly a 32 bit dynamic range - per film grain.
How about digital camera sensors? A top digital camera will produce a 2 raised to the 24th power or 16,777,215 colors (plus null or black). This outcome is due to an 8 bit level of information for each of 3 colors (Red, Green, and Blue), yielding 3 x 8 or 24 bits of color information. But each sensor has only 8 bits or 0-255 range of values. You could increase the color depth by higher resolution analog to digital converters on the chips. But this raises the problems with noise, as well as higher costs and complexity. By contrast, color scanners can have more color depth and billions of colors by going to 30 or 32 bits or more of color data. But if you want the full range of possible colors, you will have to use film or prints and scan it using a scanner rather than a digital camera.
What if you compare a single film grain (2 raised to 32nd power or 4 billion for its dynamic range) against a single film sensor (2 raised to 8th power, or 256). You would conclude that the average film grain has the potential to store 2 raised to the 24th power or 16+ million times more data.
If 24 bit color were enough, then why are color scanners made with 30 and even 32 or more bits of color depth? The answer is the higher the color bit depth, the more accurate the color imagery as a faithful reflection of the full range of colors of the original. The same is true of color printers. So film is capable of far greater information density per sensor - roughly sixteen million+ in favor of film against today's digital cameras.
Sensor Area - Another Advantage of Film
Most digital sensors use four silicon sensors to create a single picture element or pixel of color image data. Today's consumer digital camera sensors are such that each is typically a 3 micron square element (area of 9 microns).
The size of film grain is a randomized bell shaped curve distribution around some average grain size which is different for each speed and type of emulsion. A typical film grain for mid-speed film might be circa one micron in area.
The first thing you might conclude here is that you would get, on average, nine film grains with a total area of 9 microns (one micron area each) into the same size space as one silicon sensor, namely 9 square microns based on a 3 micron square sensor feature size.
So even if you could build a silicon sensor array the same size as film, it would have much lower resolution than film because of the large size of most silicon sensors versus film grains.
Stacking the Deck - Layered Emulsions versus Silicon Sensors
The analysis gets worse if you consider color images, which are what most digital camera users generate exclusively. For a typical prosumer digital camera color image, we need a block of 4 sensors, each 9 microns in area, yielding a total color sensor equivalent size of 36 square microns (4 sensors x 9 microns).
Film cheats by layering the different color filters and emulsion layers one on top of the other in the typical color film emulsion. This trick is impossible in making a silicon sensor, since they are a planar grid in two dimensions. You can't stack silicon sensors vertically, because the top layer would block light from reaching the sensors underneath it.
The layered nature of color film emulsion lets us achieve a remarkably small area for our color sensors. If the average grain size is our typical one micron area, it will be the same in each color layer (red, green, blue..). These color grains can be superimposed on top of each other to create the individual color elements in the final color Kodachrome slide.
Why did I say Kodachrome slide above? Because Kodachrome slide films replace the individual film grains exposed in each layer with a colored dye of essentially the same size. Other E-6 color slide films such as Ektachromes and Fujichromes are simpler to process, but the resulting color dye blobs are somewhat larger than the original exposed film grains. Now you know why Kodachrome slides are so sharp. The Kodachrome chemistry is different, and the smaller color blobs result in higher resolution and sharper slides for Kodachrome slides. Newer color print films have finally equalled and in some cases exceeded non-Kodachrome slides in potential resolution recently.
As a result of this film processing effect, we have to de-rate the advantage of the small one-micron sized film grain area somewhat. So we won't claim that our one micron film grains are 36 smaller than the average digital color sensor area (a composite of 4 color filter masked sensors of 9 microns area each). Instead, we'll suggest that color film enjoys a 20-fold or better advantage in smaller color sensor area against digital camera sensors.
Resolution Limits
If you have 20+ times as many color sensors (film grains) in the same square area as a digital sensor, you would expect to have circa 4.5 times the number of linear sensors (as square root of 20 is 4.47..). So you would expect to have a potential resolution advantage of 4.5 times greater for the smaller film grains in color film emulsions. Some other factors like the distribution of film grain sizes impact this issue (in favor of film), but we will ignore them here. So you would expect that for the same area, film would be capable of resolving more detail. In the same area, color film will have 20+ times as many film grains as a 36 micron square silicon sensor.
In fact, you can calculate the maximum resolution of today's digital camera sensors based on the sensor size. When you do, you get a value of circa 55 lpmm as the maximum resolution of most sensors. Simply realize that a lpmm has to have a line of black dots and a line of white dots to tell it is a distinct line. Color sensors (with support and control lines) are typically spaced 9 microns apart, so two lines of sensors take 18 microns to make a line of black and white dots. From the math, 1/18 microns is 0.0555 lines per micron, times 1,000 microns per millimeter, yielding 55.5 lpmm. So the best case resolution of current day consumer digital camera sensors is typically around 55 lpmm.
In the real world, you need to consider sampling issues, as image data will practically never align with the sensor array in the theoretically maximum resolution best case. Generally, you only get about half the maximum value, or under 30 lpmm in our example. This resolution limit would be easily achieved with 800 ASA print film in the $8.95 disposable Kodak Max HQ cameras with two element plastic lenses (rated at over 30 lpmm in Popular Photography tests).
Image Quality - 6 vs 8 lpmm in the Print?
The original Leica standard for a quality 8x10" print viewed at 10" distance is 8 lpmm on the print. A younger human eye can see differences between 4 and 6 and 8 lpmm on the print. Most of us can readily detect 4 versus 6 lpmm resolution on the print. But few adult eyes can resolve or see quality differences from more than 8 lpmm on the print at the specified viewing distance.
Many photographers are happy with less than 4 lpmm on prints which are viewed at a longer distance (e.g., 20 inches or more). Larger prints are often viewed from afar. So you can't detect the lower print resolution without getting up close (e.g., 16x20" at 20"...). If you masked off an 8x10" section of these larger prints, and looked at it from ten inches, you would see that the print quality is less than optimal (with experience). The effects of low quality minilab prints has led to a further erosion in the level of acceptable print quality too.
Most digital printers use a relaxed print quality standard as a way of expanding the size and area of an "acceptable" quality digital print. The math is again simple arithmetic. For 300 dpi printers, you divide 300 dots per inch by 25.4 millimeters per inch. The result is 11.8 dots per millimeter. Unfortunately, it takes a row of black and a row of white to make a line. So we have to divide that 11.8 dots per millimeter by 2 to produce circa 6 lpmm on the final print. So a 300 dpi printer is capable of producing nearly 6 lpmm quality prints.
To reach the Leica standard of 8 lpmm on the final print, you would need over 400 dpi (25.4mm/inch x 8 lpmm x 2 dots/line).
What does a typical 3 megapixel camera deliver? First, start by ignoring that many 3 megapixel cameras have really just 2.7 million usable pixels. We will also ignore that most digital sensors are not 8:10 aspect ratios. An 8x10" print printed full frame has 8" x 10" or 80 square inches of area. Dividing 3 million pixels from the digital camera by 80 sq. in. yields 37,500 pixels per square inch. The square root of 37,500 is 193 pixels per linear inch. In other words, a 3 megapixel camera can only produce 193 pixels of true color data when making an 8x10" print. But many color printers print at 300 dpi, or 600 dpi, 1200 dpi, or even 2400 dpi. So where are all those millions of extra colored dots coming from?
In practice, software is used to interpolate or project a smoothed set of data for the printer even when printing at a modest 300 dpi. As we calculated above for a 3 MP digicam, we have circa 37,500 pixels/sq. inch. For 300 pixels/inch, we need 300x300 or 90,000 pixels/sq. inch. We have only 37,500 pixels/sq. inch. In other words, the software is interpolating roughly 2 out of every 3 pixels in a typical 8x10" print at 300 dpi.
Now you know why digital prints have such a smooth and "creamy" texture to them. The vast majority of printed color dots are interpolated between the relative handful of actual or real color data points from the digital camera. The higher the printer dpi, the more dots and the more smoothing that goes on.
Given that 400 dpi corresponds to 8 lpmm, 193 dpi corresponds to less than 4 lpmm. So an optimally sized true 3 megapixel camera is delivering at best less than half the 400 dpi needed for a Leica quality standard print. Stated another way, a Leica quality print (at 8 lpmm) will have four times the resolvable image data on the same size print. That is quite a quality difference!
To get a Leica quality standard (8 lpmm) 8x10" print out of an optimized aspect ratio (4:5) sensor digital camera, you need not a 3 megapixel camera but more like a 12 megapixel camera (4 times more sensors). Assuming future 2:3 aspect ratio (corresponding to 35mm film's 24x36mm) digital cameras, a 16 megapixel camera will just about produce an 8 lpmm quality standard 8x10" full print on a 300 dpi color printer.
Squandered Silicon
A related issue is that some digital sensors are "in-line" arrays, while others are not. In some sensors, you use up a large fraction of the silicon sensor area in support and data storage functions, with the actual sensor being only a small fraction of the overall surface area (e.g., 30%). So only a small part of the image data is falling on a light sensitive sensor, with the rest falling on data lines and other chip features.
Imagine a window screen in which each tiny square blocked 2/3rds of the light. You will still have an image, but it might be somewhat different from one in which 100% of the light is used to generate data. The actual point where a dark area and light area change over in the image may be incorrectly guessed by the software. Moreover, smaller light sensitive areas mean fewer photons are captured, and eventually the noise levels kill the quality of your digital image.
Some chips (such as RCA) use a matrix of microlenses over these smaller chip sensors. This trick helps the smaller sensor act more like their more efficient (70% or so) cousins, especially improving noise performance. But the flip side is that the smaller sensor size can't be used to improve the potential resolution of these chips.
Chips and Resolution Limits
Chip makers face some daunting challenges. If you make the chip sensors smaller, you can get higher densities and more image data from the same sized chip. But the sensors become so small that you get less light intercepted by each sensor. The effects of noise become more problematic with smaller sensors, producing problems in the images. So even if you could make a chip with very tiny sensors packed very closely together, problems with noise would be hard to overcome.
One reason current sensors remain relatively large is that the overall system costs are kept lower by using cheaper lenses. This factor more than offsets the slightly larger silicon real estate used in making the larger chips. With a 9 micron sensor, we had circa 55 lpmm as a resolution limit. The sensor is smaller than 35mm film, so it is easier to make a higher resolution lens for it more cheaply than for 35mm film too. Most third party zoom lenses can easily deliver this level of optical resolution for the small size sensors of most current chips. So we are at a "sweet spot" where the chip size and required low lens resolution makes it cheap and easy to build current digicams.
Let us say you develop a 16 MP chip with much higher density in the desired 24mm x 36mm size format. Since today's 3 to 5 MP chips are slightly smaller than 35mm film (by 40% or so on), some of the higher density will come from larger chip area to reach 24x36mm sizes. The rest will have to come from higher sensor density. But there is a problem with denser and smaller sensors due to the cost of higher resolution optics.
Assuming you have roughly halved the sensor size, you need twice the lens resolution you needed before. Our old 9 micron sensors delivered roughly 55 lpmm. Twice that 55 lpmm value is 110 lpmm. Ooops! Now you can't use cheapy 35mm optics on your digital camera. Many 35mm zooms typically deliver only 50 or 60 lpmm as it is. The best 35mm SLR zooms deliver only about 70 or 80 lpmm. Even fixed lenses rarely achieve ratings above 90lpmm. To deliver 100+ lpmm on a 35mm sized silicon sensor, you need lenses covering 35mm areas at least as good as the best Leica or Zeiss can make today. I doubt this is going to happen for a sub-$1,000 mass produced digital camera.
The flip side of this argument is that 35mm sized chip sensors can only deliver resolutions slightly better than today's 3MP to 5MP cameras using current 35mm lenses. Do you see the problem here? If you get denser chips (e.g., 16 MP), then the current design 35mm lenses will be the resolution limiting factor. You can make the chips as dense as you want, but the lenses will only permit an overall system resolution under 70 or 80 lpmm. I doubt that we can improve the resolution of lenses aimed at 35mm sized sensor chips enough to reach 100 lpmm anytime soon.
How about a medium format sized chip? With a larger area, you could use current medium format lenses with a larger area chip to yield higher quality digital images (e.g., 16MP). The problem here is few folks have medium format rigs, and they are big and heavy. The larger chips would have higher rejection rates. The bigger chips would have more chances to have a defect on them due to their larger area. That spells higher costs too. To me, these observations suggest that the cost of custom digital backs for medium format will remain high for some time to come.
New Lenses for 16 MP Cameras
The obvious solution is to make smaller and higher density chips in smaller formats. It is relatively cheap to make microfilm lenses which resolve in the 200 lpmm range, and even 300 lpmm is readily possible. A smaller high density chip would have higher yields, and hence lower costs too. The microfilm format lenses would be lighter and cheaper to make. Most of the 40+ million digital cameras projected to be sold this year won't use existing 35mm SLR lenses. So I suspect that the smaller chip sizes of the 16MP high density chips will obsolete the use of 35mm sized camera bodies and lenses for all the reasons stated here.
The other side of this issue is whether it will be worthwhile to have a digital back or digital film insert for existing 35mm SLRs and medium format rigs. None of the current 35mm film based SLRs nor medium format rigs are optimal platforms for a digital system. Only a few medium format cameras even have data links to their backs and lenses (e.g., Rollei).
Now suppose a 16 MP digital camera with a super high resolution microfilm format zoom lens weighs less than a pound with batteries and gigabytes of removable data elements. Thanks to mass production, it costs under $1,000. Do you really think you would lug around, let alone buy, medium format or 35mm SLR bodies to get much lower resolution images from the lower resolution 35mm or 6x6cm lenses? No, huh? Conversely, given you can have a 200 lpmm zoom microfilm format lens for $100 cost on your 16 MP digicam, will you really be bummed out by not having to carry around all those 35mm or 6x6cm heavy lenses? Hmmm? The only thing you are getting by using your 35mm or medium format SLR as a base for a digital camera is a poorly designed and heavy case and access to very sub-optimal resolution lenses.
In short, I think we will see an interim design using the existing base of 35mm SLR lenses at or near their resolution limits on a digicam body for the 5 to 10 MP resolution chips. Medium format backs will continue to be specialty items at high cost, due to the small size of the market and its fractured nature (hasselblad vs. rollei vs. mamiya..). The future 16 MP digicams will probably use smaller high density chips mandating smaller high resolution optics which will obsolete 35mm SLR lens based cameras.
Financial Sanity
The average USA household shoots under 100 photos per year, or roughly 4 rolls of 24 exposure film - one roll per season. Now you know why disposable cameras are so popular. Only a relative handful of amateur photographers shoot more than a roll or two of film per week. At a typical minilab cost around $10 for film and processing, that $20 per week adds up to around $1,000 per year in running costs. Most casual photographers shoot more like a roll of film per month, or circa $100 per year. If you shoot slides (as I mostly do), then your costs are more like half this figure.
Digital camera users would have you believe that since you don't need film or minilabs to do digital prints, the cost of digital photos is effectively zero. Maybe so, but they must be stealing those 2CR23 batteries from somewhere. If you are used to replacing a mercury or silver cell in your light meter every 3 to 5 years, the battery costs for a digital camera can be a suprise. Even worse, if you need to use flash for many of your photos, you will be shocked at how fast even a small flash eats up lithium batteries. On one of my web cameras, I can take 250 shots per set of batteries (3 cents per photo), or 60 shots with flash. A typical mix yields circa 8 cents per photo for digital camera battery costs alone.
Naturally, you could use an external battery pack with rechargeable batteries and a charger, and carry spare batteries. Some digital pro cameras use AA rechargeable batteries, although many prefer the higher energy and much higher cost NiMH rechargeables. My homebrew external battery pack for one of my memory card cameras weighs more than the digital camera. But most folks just put more batteries in while arguing to themselves that they are really saving much more on film and processing.
Note that I am not counting the cost of external strobe batteries here either, since that would be the same for film too. But on many digital consumer cameras, there isn't a provision for triggering an external strobe except by using the internal flash to trigger a slave photosensor on the bigger external flash. Most digital camera strobes have such low flash power that they can't really do much for lighting even at 6 to 10 feet. So you may end up with a much larger kit using a real strobe to light the eyes of subjects even ten feet away.
What if you elect to make digital prints of all your photos to 4x6" or 5x7" or whatever your photolab provides? You have to pay for the costs of paper and ink. If you want the highest quality photos, the proper papers cost around $1 per sheet. For four 4x6" prints, that's about 25 cents each - for high quality photo-grade papers. So if you are using the high quality paper, plus ink, plus factor in battery costs, the cost per 4x6" print is higher for digital prints than the cost of film and processing for mini-lab prints.
One nice feature of digital printing is that you can make pretty good quality 8x10" prints, even 11x14" and panoramics up to 11x48" on some color printers. You can, that is, if you start with a film image that is scanned into a digital file. Today's current 3 megapixel digital cameras may produce an acceptable 8x10" print on some shots, but few can do a full print (to the borders) on 11x14". You can find mail-order places offering 8x10" prints at modest costs (often just over $1 per print in bulk). I find it a bit paradoxical that the big savings with digital image processing and printing isn't feasible with current digital cameras, but rather only to those of us using both film and digital technology.
You can use regular paper or lower cost photo paper in many printers, but the quality will be less (and possibly less archival). Another significant cost is ink, which can run anywhere from a few cents per print on up (some printers require 3 or 4 ink cartridges at up to $30 retail a pop). So while it is possible to spend less on printer paper with some printers, you still have significant on-going costs for ink, paper, batteries, and other supplies.
What other supplies? How about all those CDROMs or zip disks that are storing your digital images? A 3 megapixel camera (at 16 million colors is 24 bits or 3 bytes per color pixel) works out to circa 9 megabytes of raw data per image. Even a 1 megapixel 24-bit color image is 3 megabytes. You can use lossy compression (e.g., JPEG) and greatly reduce these file sizes, but at the loss of already marginal image data and fidelity of the image. You will also need to backup your files on other media (e.g., negatives and prints are backups of each other). What I am suggesting here is that you have a significant cost in storage media which is often ignored by those claiming that digital camera costs are nearly zero.
Obsolescence - Digital's Hidden Cost Iceberg
An iceberg has 90% of its bulk hidden underwater. The same is true of digital camera costs. The big dollar cost in digital photography is not the cost of batteries, ink, photo-quality papers, or storage media cited above. The big cost is buying a $2,000 digital camera plus $1,200+ computer and monitor and color printer ($200+) setup with software ($?). Two or three years later, the computer will be worth a few hundred dollars, and the digital camera about the same.
Nobody wants a 640x480 web camera, or even a 1.2 megapixel digital camera, when you can get a 3 megapixel (usually 2.7 MP on chip) camera. As with older computers, the price drops precipitiously. That economic loss when upgrading to a new digital camera model every few years - new printer and new computer and new storage system and new software version of Photoshop... - hey, it all adds up.
Let's assume that your depreciation losses on upgrading your setup to a new digital camera and new printer are as suggested above. That works out to a $2,000 depreciation loss on hardware, software, and peripherals in two years, or roughly $1,000 per year. This figure is roughly the same as the running costs for the more active film amateur photographers shooting a few rolls of film per week. Our casual shooters burning a roll a month or $100 a year are spending rather less on their photography. So the vast majority of amateur photographers would be out less money if they are shooting film with film cameras with a much longer obsolescence period (e.g., ten years or $100 per year in camera obsolescence).
Speaking of obsolescence, don't forget that high power computer, disk drives, monitor, CDROM, backup tape drive, and color scanner. Now you need a CDROM burner, no make that a DVD reader, no, you really need this DVD burner and buggy software that goes with it. The syquest tape backup drive is out, you need a DAT backup tape system. The old 15 inch monitor is too small, you really need this 19 or 21 inch monitor. And your old color scanner is only 24 bits, don't you really need 30 or 32 bits? Don't forget to get the light table for it to scan film in too. You could use an Intel 486 for your internet email and office projects, but to run Photoshop with 128 megabytes you really need a Pentium II, or is it III or IV? If all this sounds familiar, it is probably because you too are on the digital express. As Alice in Wonderland said, you have to keep running to stay in place. Only with digital photography, you have to keep paying and paying!
Learning Curves
How much is your time worth? Maybe you would be better off working at one of those film minilabs for $5 per hour and doing your prints for free? When you pay for minilab prints, you are also paying for the labor to do the job, and re-do the job if it isn't done right. Even after learning and doing digital photography, you will often find yourself reprinting prints and spending hours "tweaking" your images. If you don't match your color monitor to your output device, I can guarantee you will be doing a lot of reprinting if you are finicky about your photos!
Speaking from my teaching experience, I can assure you that there is a steep learning curve for non-geeks learning digital photography and computer technology. The cost of books and courses is also never mentioned by advocates of digital photography. You can often buy a nice coffee table sized photobook by a favorite pro photographer for the price of a thick and boring software book with CDROM.
The cost of image processing software, add-in packages for special effects, and other software packages is also not trivial. It is not unusual to pay more for computer software than for computer hardware, especially with some programs like Photoshop and the Adobe suite costing over $500 for a commercial copy. Don't forget to factor in all those digital photography and computer magazine subscriptions that you will be reading to learn the inside tips too.
And finally, every hour behind the computer monitor or reading a computer software book or manual is another hour you won't be spending taking pictures.
Color Reference
If you take a color slide or color print, you have an inherent color reference against which you can check your final print. But what do you use in the digital world? A simple example of this problem would be to view the same image on a Mac and on a PC monitor, with obvious color differences between monitors. But the only solution is to preserve the color fidelity throughout the entire chain, including the final output device. But CMYK inks and printing processes may vary as significantly as monitors, as there is no defined standard on the output, let alone for each element in the chain. Efforts such as Epson's Print Image Management system attempt to fill in some of these gaps. But digital photographers are basically adrift without such start to finish color fidelity management capability. By contrast, it is easy to compare a slide or print with the resulting ad copy or magazine or book pages. [see Professional Photographers Newsletter (email from British Journal of Photography) of 27 March 2001].
Beyond 16 Megasensor Chips?
Today, only 1% of all minilab prints are 8x10" prints or larger. Most computer color printers are limited to 11 inch widths, implying a standard 11x14" print size. My guess is that prints larger than this standard width will be farmed out to minilabs with larger printers. The flip side of this observation is that the optimal digicam size is likely to be around the 16 MP chips which can support a decent 11x14" print with some cropping allowed.
I am suggesting that it may be hard to justify a costly 64 MP chip camera if you are just doing 11x14" prints. Many of us will be happy with 8x10" or 11x14" prints from a 16 MP digicam, just as we are today with minilab prints in this size. A square chip would provide 8192 pixels on each axis, or 4096 lines, yielding 500 mm of print with circa 8 lpmm print quality, or one meter (circa 39" ) of 4 lpmm print quality. But if you are just printing 8x10" or 11x14" prints, the higher density of the 64 MP chips may be overkill that won't show up in the prints. The human eye can't see or resolve the data past circa 8 lpmm, so the extra information might not be readily discernible?
Data Interpolation
Digital data is currently interpolated to a large degree to supply a larger image when printing from current 1 to 3 MP digicams. As noted above, a 3 MP camera provides about 37,500 pixels per square inch in an 8x10" print (optimally mapped at the 4:5 aspect ratio). The typical 3 MP camera (2.7 MP true) actually delivers closer to 150 pixels per linear inch in an 8x10" print. Most printers generate 300 dpi or better output. That means we require 300x300 or 90,000 dots per square inch to make the print. Where do those extra dots or data come from? Interpolation!
Interpolation happens at a number of levels. The sensors on a typical chip are only able to measure levels of light, typically 8 bits or 0 to 255 levels of greyscale data. To generate a color picture element (pixel), we have to use at least three sensors, each of which is masked with a color filter to respond mainly to red, green, or blue light. The 8 bits of red, 8 bits of green, and 8 bits of blue data are used to create a 24 bit color value. In practice, we use a four element Bayer pattern of RGGB, partly because such a power of two array is easier to design, access, and build. Since the human eye is most sensitive to green light, the averaged green information provides the best and most pleasing image results.
However, some high resolution digital cameras are made using three separate chips. Each of these chips is masked with a different color filter, resulting in the required red, green, and blue color data.
While we currently use 24 bits of color data depth, other higher values are possible with lower noise and higher analog to digital converter bit depth (e.g., 30 bits of color data with a 10 bit A/D converter). We currently use 24 bits as the best compromise of cost and complexity against acceptable quality of the resulting millions of colors provided by 24 bit color depth. Color scanners have improved and increased their color bit depth from 24 bits to 30 and 32 bits and beyond, so digital cameras may follow suit in the future too.
Imagine a bathroom floor made of patterns of red, green, and blue tiles. The pattern is RGGB in a square or diamond shape. The software takes the observed 8 bits of red, blue, and (averaged) green data and generates a 24 bit color value for that square. That data point can be considered to be at the center of that square, and is a dimensionless point. But you can also realize that it represents the average intensity and color of the light falling on the light sensitive sensors in that grid of four sensors.
If it takes four sensors in the RGGB Bayer pattern to produce one color pixel, how does a 3 megasensor camera deliver 3 million pixels of color data? This process varys with different cameras, but in general, the camera uses the nearest available blocks of the required colors to interpolate a color data value at each point. So the four nearest red sensor cells to a blue block might be averaged to get an 8 bit estimated value for red at that point, and similarly for green. Now move on to the next sensor, say a red masked sensor, and repeat the averaging for the four nearest blue and green sensors around it. Keep going, stepping through the matrix of sensors.
One minor problem is that when you get close to the edges of the chip pattern, you don't have the required color data to project estimates for these edge sensors. For this reason, many chips are unable to provide quite as many pixels of color data as they have actual on-chip pixels. The larger the chip, the smaller the percentage of these lost data points. In some cases, the software tries to "mirror" or guess an interpolated color pixel value for the edges too, but the guess may not be very good.
A more interesting problem is that the average and maximum dimensions of these interpolated color pixels may be different from those computed using the close Bayer block pattern (RGGB). Blue sensors on a grid will have a pattern of four green and four red sensors in a box around them. Red sensors will have a pattern of four green and four blue sensors around them too. But what about the two green sensors next to each other? Ooops! The patterns must now be different. Depending on your approach, you will be using data from sensors farther away (e.g., to get four red and four blue values for averaging).
What happens to resolution if you are averaging in light from more distant sensors? In effect, you have generated an average color value from a larger area sensor, right? And a larger area sensor means lower resolution, given a fixed sensor density and chip size. So the chip resolution depends a bit on color and software algorithms used in interpolating these color values over the grid of the chip sensor array (excluding those "falling off the edge" values).
Effects of Chip Defects
In order to keep yields of high density chips higher, some chips use complex software to keep track of glitches or bad spots on the chip. Imagine a dust mote in chip processing has zapped a tiny spot on the chip's surface. While the spot is tiny, it has destroyed two sensor sites. The chip has to use special software routines to lookup and interpolate values to use in place of the bad sensor data. Some chip makers might permit only a few such defective sites in their chips before destroying them. Others might be happy with lower costs and higher profits from accepting scores such bad sites and fixing them in software. But the result is further loss of fidelity of the resulting image data against the original image projected by the lens of the subject.
Interpolation In Printing
All of the on-chip related interpolation issues pale in comparison to the issue of interpolation during printing. The typical 3 MP camera (actually closer to 2.7 MP image data) can provide circa 150 color dots per inch from actual sensor data (itself interpolated). The photo-realistic color printers typically start around 300 dots per inch, and go up to 2400 dpi and above. You have data for 150 dpi, but want to print at 300 dpi, so you need to interpolate 3 new data values for each value of image data that you have. That's for an 8x10" print using a 3 MP camera. For an 11x14" print, you have twice as much image area. So you have to interpolate 7 out of 8 data values in software. Stated another way, 3/4ths or 75% of what you see in an 8x10" print is interpolated data. For an 11x14" print, some 7/8ths or 88% of what you see is interpolated data. To do 16x20" prints, you would have to interpolate 15/16ths of the image data on the print.
In practice, few folks would find 16x20" color prints from a 3MP digicam to be of sufficient quality, and most would find 11x14" prints rather marginal.
One of the giveaways of low resolution (3 MP) digicam prints is their "creamy" texture. The texture is creamy because it is largely interpolated, with a series of interpolated values smoothing out the steps between actual data values from the 3 MP cameras (itself interpolated). Lots of people like this creamy smooth image effect, unless they have had experience with higher quality photo prints.
Sad to say, but few people nowadays have ever seen a high quality photographic print. In many minilabs, the enlargers are purposely defocused slightly to hide the effects of dust and scratches on your negatives. The low quality of many fast 800 ASA/ISO films also doesn't help. So the low quality of today's minilab prints has accustomed the public to lower quality photographic images, making the lower quality of digital prints seem as good and in many cases better than minilab photographic prints.
We are already seeing people rediscovering photography and high quality print making after seeing quality photographs such as the traveling exhibitions of Ansel Adams environmental prints or even a local camera club salon print competition. I suspect that one of the future benefits of film based print making will be precisely the ability to produce contrasty and detailed prints which extensive software interpolation and smoothing makes impossible in digital prints.
Best of Both Worlds
Fortunately, we have another way of getting digital images - scanning. By scanning film, we get the benefits of film with its high quality along with the economy of scanning over digital sensor arrays. Even modest cost scanners enable us to get a scanned image from film or prints which are much larger than the current prosumer digital cameras.
For many amateur users, modest cost scanners and color printers will make a nice combination with their existing computer systems. Personally, I think this trick will make 8x10" prints more accessible to many amateur photographers, including those who don't have access to a home darkroom.
One of the intriguing options here is to use the data from panoramic selections and print panoramic prints of any length (by setting printer software and using roll papers).
Another alternative is to have the film scanned on a drum scanner or other very high quality scanning device. The resulting digital image data files (at up to 600+ megabytes per image) dwarfs the image data provided by 3 to 5 megapixel digital cameras. Costs vary from $15 USD on up, with delivery via CDROM or over the Internet as popular options.
At a past Dallas Hasselblad University program, the discussion and interest on digital photography focused on the scanner based options. Perhaps this is just a reflection of the $50,000+ cost for a medium format digital back for Hasselblads from Dicomed and others? But many pros have found that adding a scanner and learning digital skills have opened up some new markets. But probably 3/4ths of pro photographers have yet to see enough advantages to begin to make the sizeable investments in time, money, and learning efforts to make the transition to digital. They may be wisely waiting for us amateur digital photographers to drive down the costs enough to make digital photography worth the real costs - including high depreciation rates.
Conclusions
Today's high end digital 1 to 3 megapixel digital cameras do a decent job of providing modest sized and quality images (up to 8x10" or so), with an emphasis on speed and convenience over film processing requirements. But even at 3 megapixels, today's digital cameras have limitations and deliver less quality than even modest cost film based cameras. The total costs of a fully digital camera based system is also much greater than usually claimed, largely due to rapid depreciation of cameras and equipment. Many of the benefits of digital imagery (for web use..) can be achieved by using film with an appropriate scanner or color printer. Until the costs of 16+ megapixel sensors falls significantly, film will still provide much higher potential quality than even the best prosumer digital cameras.
,en réaliter,si les reflex numeric ne coutai pas si chere,j'en aurai surement 1 (et oui,g bien dit numeric,car c le numeric ki se raproche + de mes bessoin)
) même si l'aspect extérieur à souffert 
Cet appareil, c'est vraiment du solide. (et du lourd aussi 
Quand j'ai essayé un reflex type EOS avec son zoom Sigma 200mm-tout-carbone, ça m'a fait un choc ! Plus léger que le F1 et son 50 ! 
![[:netbios]](https://forum-images.hardware.fr/images/perso/netbios.gif "[:netbios]")
![[:skylight]](https://forum-images.hardware.fr/images/perso/skylight.gif "[:skylight]")
Je sais pas pourquoi mais je doute un peu...