Image source: http://www.lcse.umn.edu/specs/labs/images/spectrum.gif
When I was getting started in digital photography it was helpful to me to read in-depth articles about how digital sensors work, and how digital files work. This article is in an effort to educate others about these facts as I understand them. My learning has been informal and independently gathered.
In this article, I discuss light, color, digital camera sensors, digital file types, computer displays, lenses as they relate to sensors, digital noise, noise reduction, stops of light, digital sensor sensitivity to light, ISO settings, and histograms. There are several important topics that I altogether skip or about which I present limited information.
First let us consider that the mechanism of photography enables us to capture light. More accurately, we capture varying intensities of light. More accurately yet, we are capturing varying intensities of different wavelengths (colors) of light.
Our eyes limit us in that they are able to see a rather small percentage of the waves that exist in the electromagnetic spectrum. I often wonder what it would be like if I could see the radio waves, microwaves, X-rays, and gamma rays that are being broadcast all over the place. Unfortunately I am able to see only a sliver of the energy that exists: visible light.
Image source: http://www.lcse.umn.edu/specs/labs/images/spectrum.gif
Notice how red and violet are both at the extremes of our eyes' sensitivity. If you've ever noticed how difficult it can be to make out a purple illuminated merchant sign at night, or to find a crisp edge between the edges of red and another color of an illuminated sign, it is because our eyes are at the extent of their performance ability. (Plus we don't have a lot of cones that are sensitive to violet.) Equate this to discerning the sharpness in a very low frequency 20-30 hertz audible sound or a very high frequency sound 18,000 to 20,000 hertz. Our ears cannot hear everything that is in our world and our eyes cannot see everything in our world.
Within this constraint of everything we call colors, we identify three primary colors within the spectrum of light. These are Red, Green, and Blue. Color is additive which means as you add part of a 2nd color the result is a new color. (In contrast to the spectrum of paint where subtractive principles apply). We shorten Red, Green, Blue to RGB and when we communicate information about these colors, we do so in that order from left to right. If we have zero intensity of all colors, we have black. If we have maximum intensity of all colors, we have white. Be advised, however, that white is a very interesting color. I just said that it is the maximum intensity of all colors. Remember for the future that "all colors" is not an absolute thing. Hint to a broader topic: White balance
By coincidence there are basically three kinds of light sensitive receptors of our retina. Some are sensitive to red, others are sensitive to green, and still others are sensitive to blue. (It's not quite that simple, but most seem to agree that this seems sums it up pretty well.) Perhaps this is the reason that our digital cameras, at a low level, have been built to be sensitive to this same set of three colors. Likewise, most image file formats store image data in this format.
At the genesis of a digital picture is the sensor. At the most primal level a sensor has no idea about color. When asked by the camera, all the sensor can say is zero light, very little light, halfway light, fairly bright, really bright, all the way bright. This is a physical limitation of how light sensors work. This limitation comes from the electronic components that engineers have been able to design.
By placing a filter in front of the sensor, we can then use logic to deduce what color our colorblind "eye" is seeing. For example, if we place a filter that only permits red light, then we can append to the sensor's reply and deduce the following kind of information: zero red, very little red, halfway red, fairly red, really red, all the way red.
In the previous example I created, the sensor can tell you 5 different levels of bright. In reality, sensors are much more elaborate. They are able to report many different levels. For now, I'll say it can report 4096 levels from black to white, or from none to all. Where black = 0 and white = 4095. However, this isn't the totality. Also, many file formats "throw away" much this information. (Or round it, or truncate it.)
On to the next step: If we break this sensor down into a grid of independent light sensitive squares and then we populate each square with its own color filter (Red, Blue, Green) we can then deduce the exact amount of a single wavelength of color recorded at every single site. This is what cameras do.
This is known as a Bayer pattern, named after Bryce Bayer. Notice that there are twice as many green sites as there are red sites or blue sites. This is because green is a very important color. Our eyes have a higher sensitivity to green than to red or blue. Each of the three types of retinal cones has some degree of sensitivity to green hues; it is the color range with the most overlap. If green is wrong, we really notice it.
Now we are closer to having enough information to build an image. For each site on our grid, we know exactly how much there is of the designated color, but the problem is we have zero idea of the amount of the other two colors. For example, on a green site, we know how much green there is, but we don't know how much blue there is or how much red there is. The solution? Let's ask our neighbors and then do our best to make an educated guess.
Suppose I am the red site in the middle and I just recorded an arbitrary amount of red, let's say 75 units of red. I am going to ask my neighbors on each of the four sides how much green they saw. I get answers of 12, 10, 15, and 8. I average these answers and get 11.25 which I must round down to 11 because only integers are possible. Finally, I ask my blue neighbors on the corners and I get answers of 200, 240, 254, and 195. I average these numbers and I get 222.25 which I have to round down to 222.
So when that red site in the middle is asked by the camera what all colors of light were measured, its answer is (75,11,222).
Here is the color (75,11,222) calculated in this pretend example:
The camera has to do this for every single square, or pixel on the grid. (Pixel is derived from picture element, it is the smallest unit of a raster image.) When we consider that cameras contain 5 million, 6 million, 12 million and more sites, a camera is rather busy during that half second before the final image is written and presented to us.
This is a fairly effective method, but it does come with its problems. Since we know only 1 value for sure and we are guessing at the others, there is never a site that accurately records how much of each color light was really there. Most of the time we are not likely to notice the difference, but in smooth and even areas, such as the sky, we may see funky splotches going across. Some of this effect may be due to digital noise (I'll elaborate on that later) but I tend to think that most of sky problems are due to side effects of the Bayer pattern.
The Bayer system isn't the only method of measurement. Another fairly popular idea is the Foveon sensor. Sigma is known to employ these in their cameras. Rather than chopping the sensor into a grid, it employs 3 layers. Each layer is sensitive to a wavelength of light. The benefit is that every photo site reads its own level of each color. The drawback is that the light reaching the rear layers has already received interference and light "fall off" from having passed through previous layers. This may result in color shifting, especially in low light environments.
Incidentally, it is my understanding that film was built in a similar fashion, using color sensitive layers. Many films are known to color shift in low light (long exposure) situations.
If we skip beyond all of this process of collecting light and doing math, we arrive at the point of a digital image file. A widely popular format is JPG, so I'll start there. Since so many of us will elect JPG as our ultimate destination, it is important to know its strengths, weaknesses, and inner workings as we are out shooting.
I mentioned earlier that many sensors can give 4096 levels per color. (Disclaimer: This depends on the sensor, some do more, some less. It happens that the sensor in my camera is 12-bit) I also said that many file formats throw away or round off much of this information. JPG is an example of that.
A JPG can store 256 levels for each color. That is 256 for red, 256 for green, and 256 for blue. It turns out this isn't very much! A JPG stores 8-bit information, or 8-bits times 3 colors. Therefore they call them a 24-bit file. Zero means none of a color and 255 means all of a color. 255 is the limit. Reaching this limit can cause problems for us. (More about that later)Personally, I can't respect JPG as a 24-bit file. For me it seems like an 8-bit file stacked 3 deep vertically rather than end-to-end horizontally, but this is a personal peeve.
I hope to avoid too much involvement with binary numbers, but we need to visit the topic a bit. In the world where computers do math, a JPG file uses 8 bits (per color). This means that each color can be shown at 1 of 256 different levels and mixed with any of the other two colors at any of 1 of their 256 levels. If you multiply 256*256*256 the result is 16,777,216. This means there are exactly 16.7 million explicit colors that can be used to display your digital picture as a JPG.
I said that for each color, RGB, there are exactly 256 levels, or increments. Here are all 756 of those constituent colors:
Note: Click on the thumbnail to see the palette. Be sure to view actual size, your browser may try to fit it to screen.
rgb_pallette.png (25 KB)
Notice how each color has black in common. The left most region of color palette offers less significant influence or less diversity to bring to the image than the right most region of the color palette.
When constructing an 8-bit pixel in an image, exactly one and only one color from each of the three rows you see here is chosen. Remember that color in this space is additive not subtractive, so adding only 75 of blue to an existing 255 of red doesn't make the red blacker or darker (even though blue 75 is rather dark) as this would be subtracting. Instead, it adds a small amount of blue. The way to make red darker is to add less of it to begin with. 100 red would be darker than 255 red. If this were paint, the results would be quite different.
If we are to do an image in strictly black and white, then we are limited to a mere 256 shades. (Support for the reason I personally decree a JPG as an 8-bit file that is 3-deep rather than a 24-bit (8+8+8) file. Aha!) These 256 increments are a tragedy because 1.) a human eye is sensitive to far more increments than this, 2.) there were more increments recorded by the sensor but they have now been rounded up or down, and 3.) there were more levels in the first place than what our sensor could see. When applying adjustments to the image, the restricted mathematical slot space available in these measly 8-bits of depth create a lot of problems for us, especially for images which contain a lot of blacker pixels or those images which are dark and we wish to push to a brighter appearance.
The full topics of bits, color depth, and the binary numbering system is beyond the scope of this tutorial. There are many resources on these topics.
The things that JPG changes with color data in order to achieve a smaller file size is yet another weakness of the format. I am going to skip this topic as it has been relentlessly discussed on the internet already. A query for JPG, artifacting, compression, lossy should produce numerous articles.
In spite of its weaknesses, JPG remains a very convenient format, and I therefore consider it useful enough that I shoot JPG much of the time.
As a side note, a JPG's file size can be made smaller by adding a slight blur to your image. Sharpening will increase file size. A white-weighted JPG consumes more space than a black-weighted JPG.
Many cameras can also output in a TIFF format. This has some benefits, but we would still have a drawback because of the fact that the camera already made a decision about white balance (and it could be wrong). There are many instances where this is a great choice, if your camera offers it.
For those other times, there is RAW format. Not every camera offers this option. Again there is a lot of information about this and I am not going to try to offer a full explanation. The short story is that the camera does not do the previously described math calculations. Instead, it just does a raw dump of what each sensor site measured and it is left to your favorite photo processing software to figure it out.
There are several benefits to this. You, the user can decide the proper white balance. The cameras decisions about white balance are frequently incorrect, especially during night photos. (White balance is a fairly involved topic that I am also going to skip, but you should look it up!) With new enough photo software, you can elect to import the raw file into a 16-bit color space. This is great news. Now instead of those insufficient 256 levels of color, we have 65,536 levels for every single color. Woohoo!
Now we can pick from 281,462,092,005,375 colors as compared to the previous limit of 16.7 million colors.
What was limited to 256 levels is now increased to a much more comfortable 65,535 levels.
Keep in mind that the camera's sensor may or may not have contributed that many levels of color. Even if it didn't, a 12-bit image piped into a 16-bit file still offers the editor a better input to begin with and then a larger quantity of levels to which the image can be adjusted and tweaked.
RAW comes with a few expenses. Now instead of a photograph requiring 1/60th a second of our time to create, it can require 30 seconds, a minute, 3 minutes as the user fiddles around deciding how it looks best. This is the greatest drawback for me. Having shot only 30,000+ images; if I were to assume that every picture took only 1/60th of a second (which they haven't) then I would have spent 500 minutes with the camera shutter open. That's about 2 whole days. If we suppose I'm reasonably efficient with RAW processing & tweaking and every image were to take one minute, that's 3 months of messing around.
The RAW files, even with compression, and the associated resultant saved files consume 4x to 8x the disk space of JPG in my experience. I'm a big fan of RAW, but for me, I try to use it only when I need to.
The most likely file format choice to handle this kind of image is a TIFF file, although I am a big fan of PNG files. It seems that PNG is a little-known file format even though it is an open standard (JPG isn't open, it is proprietary) and like TIFF, PNG is a lossless format too. However, it is compressed and occupies approximately 50% of the disk space in my experience, and it does take longer to save. Both TIFF and PNG formats are available in either 8-bit color depth or 16-bit color depth.
A TIFF file is great because it does not do anything to destroy the image data, it writes exactly what the pixel is. You can always predict the file size of a TIFF right down to the byte. (Number of available colors * width * height) = file size in bytes.
There is endless discussion about how successive saves to JPG files will create more and more artifacts. A tiff file (because it is loss less) does not do this. It has been alleged that good quality, high end photo applications don't re-calculate the whole JPG with every single save. In other words, you don't necessarily damage the photo again in places that were not edited. Based on my experiments, I tend to believe this is true. Therefore, I don't worry much about using JPG for many of situations.
Regardless of all these numbers and the amount of depth in our files, our ability to output stunning imagery is going to be limited to what we can see when we are editing it. (Of course many images might be great right out of the camera, but I'm not one to think that the list of preprogrammed assumptions and lookup tables inside my camera are so perfect.)
I lament the days of the CRT. It used to be that a Sony Trinitron or a high end Princeton Graphics (or countless others) could be a great monitor for everything. Good for photo editing, good for office applications, good for gaming. Now LCDs are the rave. It was with the greatest of resistance that I have moved into this realm.
I'm in favor of lower energy consumption, but I've generally disliked the color of LCD panels. When my good old Sony's picture started to shake and rattle around to the point I could no longer read text, I did some serious research into LCD technology in order to establish more technically what it was I wasn't liking about them.
In the world of LCD panels, there is *always* a tradeoff. Whereas a good CRT was always good, a good LCD excels in one category and not as much in other. An LCD, (TFT technology, which is thin film transistor) works by having a bunch of rear-illuminated color sites (kind of the opposite of a digital sensor) and between the site and your eye is a sort of micro miniblind. Depending on the desired color of an area, these "blinds" are either closed, open partially, or open fully; thus mixing together the colors of Red, Green, and Blue reaching your eye.
The vast majority of LCDs out there, (I'm convinced that almost all of consumer models) are TN based LCDs. If one is to go to Sam's Club for an LCD, then chances are good that one gets a TN based display. Likewise with Best Buy, or even your local computer store. (Your friendly neighborhood store could order something nice.) While these are good general purpose displays, I personally dislike them greatly. My greatest protest is that fact that they advertise 16.7 million color ability (remember that number?) BUT, the panel itself is actually capable of only 63 levels per color! That's 250,047 total colors. This is a far cry from the 16.7 million which we previously established wasn't enough in the first place. They still display from none to full and they interpret the full 256 color levels, but what a TN based display does is fake (emulate) the rest of the colors.
"For example, the monitor has to output the color RGB:{154;154;154},
and the matrix doesn't’t physically support it, but it supports the two
neighboring colors, i.e. RGB{152;152;152} and RGB{156;156;156}. If we
were outputting these two colors alternately with the frequency of the
refresh rate, the similarity of these colors and the inertia of the
human eye (which doesn’t perceive flickering at a frequency of 60
hertz) as well as of the matrix itself (which is “smoothing” the moment
when the colors are being switched) would give us what our eyes would
perceive as some in-between color, i.e. the required
RGB:{154;154;154}."
Source: http://www.xbitlabs.com/articles/other/display/lcd-guide_11.html
This is admittedly a great workaround for the engineers to have devised, but I am not really satisfied with the result.
For the photographer it is worth looking into S-PVA based (and related predecessors or derivatives) as well as S-IPS based (and related predecessors or derivatives).
The very unfortunate part is that these are very expensive. When I went shopping, I could barely find anything below $800. This was very bad for me, because I was upgrading to a dual monitor setup. Eventually I settled for a pair of good used LCDs for a lucky price on ebay (Below $200 each I think it was). These are based on S-IPS and I am very happy with them.
Earlier we discussed how sensors work, but how do they work, really?
Two widely used sensor types are CMOS sensors and CCD sensors. I am going to discuss CCDs.
A CCD has a photo diode. This is somewhat like a transistor and somewhat like an LED in reverse. Whereas an LED emits light after you apply a voltage, a photo diode creates a voltage when light hits it. This part is analog. When your camera begins to take a picture this voltage is reset and then the sensor is exposed to the world (aka, your picture). The more photons that reach this photo diode, the greater the voltage. More photons occur either because the scene is very bright, dumping in a bunch of photons quickly, or because the photograph is very lengthy, allowing photons to slowly accumulate. Thinking of a photographic basic of reciprocity (shutter versus aperture), if we could choose between a short photo and a long photo, both of which yield the same overall lighting; the short photo is better off as far as the sensor data is concerned. I'll explain this soon.
Once the exposure is complete, the sensor uses a circuit to read the voltage at the site and this is sent to an amplifier. Eventually this result is converted to a number which, depending on the actual sensor, can range from 0-255 (remember these familiar numbers?) or 0-4095 or perhaps 0-65535
If we consider than any electronic circuit by nature is going to have some degree of electronic noise, then we can understand that there are at least 3 different areas through a CCD where digital noise can be introduced and interfere with our photograph.
For reasons that I don't yet fully understand, a CCD at rest will slowly build up a charge. Since we desire any and all charge to be generated solely from photons, this natural voltage buildup is noise to us. At the start of a picture, the CCD voltage is first dumped, however, it will immediately begin building this charge once again even while it is taking a picture for us. The rate at which this charge builds is contingent on the temperature of the sensor. A warmer sensor will build a charge more quickly than a cold one. I am much more active in my night time shooting during the winter. Astor-photographers may use a cooled CCD. A shorter exposure offers less time for this to occur.
Some cameras offer in-camera noise reduction. Immediately after taking a photograph, the camera will take a duplicate exposure, but only with the shutter closed. It makes note of its readings while everything was black (thus any reading is assumed noise) and this is subtracted from the RAW data of the previously shot image. Because the noise factor is dynamic, this reduction algorithm cannot be hard wired into the camera or done with software; it has to be done when the photograph is done.
The craziness of photons yields some photon noise. Try as we may, we can't make them all go where they are supposed to go. Some light will end up in the wrong "square" and make our picture lie.
When the sensor reads the voltages that were generated by the photo diodes, some read noise is introduced. I suspect some of this is related to the amplifier, as every amp always introduces some trouble.
We've established that a sensor introduces some noise along the way. Remember, this noise is all taking place before the existence of the computer file which stores the resulting numbers. As discussed earlier, our assumptions because of the Bayer pattern are going to introduce some problems and "grainy" appearances. After all of this, so many of our images meet their fate as a JPG file, which introduces a yet another set of lies and distortions of the truth.
Let's go backward a step. Before the light got into the camera it had to go through the lens. The lens is round, therefore it projects a circular image into the camera. In the days of film the film plane was typically rectangular or square.
The above diagram is specific to SLR cameras. When SLR cameras came out there was a lot of discussion about "magnification factor" and a lot of corrections to this term which is more properly referred to as crop factor. One of the reasons we like SLR cameras is because of their large sensors as compared to smaller point and shoot cameras. (More about the reasons to follow.)
Better as they were, these SLR cameras introduced something new. Because the SLR sensor was smaller than the frame for which all of the lenses had been designed, we found that our images looked different, more zoomed. This was disliked by wide angle shooters and appreciated by telephoto shooters.
Because the smaller sensor is utilizing only the center portion of the projection, this we are using only the center portion of our lens glass elements and ignoring the rest of the light. This is all the more reason for good glass. (Canon L glass, Nikon ED glass) Imperfections such as distortion, chromatic aberration, etc., are magnified.
Manufacturers began making lenses especially for the smaller APS-C sized sensors. I can't comment too much about them because I have never seriously considered them. Because it was predicted that a day would come when a full frame sensor would emerge, I considered the "digital" lenses a product with a finite life cycle.
We can define signal as anything that we wish to measure, record, or collect. We can define noise as everything else that obstructs this reading, confuses the reading, or otherwise contributes to false and incorrect readings of our signal. We discussed earlier that there are several opportunities for noise to enter our digital picture through both noisy photons and electrical noises. Our signal is light.
All else being equal, if we compare two different sensors of the same resolution where one is large and the other is small, we would choose the larger one because the resulting images will have lower noise. I am unable to determine if the larger sensor actually produces less noise, but I am certain that it receives a stronger signal. Therefore the larger sensors have a better signal to noise ratio.
Because a larger sensor has more area, it is exposed to a larger volume of light, therefore a stronger signal. Compare it to being in a crowded room and trying to make out a quiet song, versus a loud song. The larger sensor produces an image whose signal information comprises a larger percentage of the overall resulting numbers. Here are some examples of the difference.
These cameras were manufactured in approximately the same era. It is worthy to note that if the same S50 were manufactured now using today's sensor technology, we would expect it to have less noise than this one did several years ago. However, within the same generation, we will always expect a larger sensor to yield a better picture with regard to noise.
In the lower right, notice how the noise starts to give hints of that Bayer pattern. You can see hot dots of red, blue, and green in there. The sensor voltage readings in this section was amplified quite a lot, thereby magnifying those problems we get from estimating the colors from neighboring pixels as described previously.
It isn't fair to protest about the color differences, since each camera made its own decision for white balance, tint, exposure compensation, contrast, brightness and such.
It was mentioned earlier that a CCD is worse about generating noise when it its temperature is greater. I don't know to what degree it is affected and what temperature range this means. Nor do I know how warm a CCD gets during exposure either. Whether this is noticeable at 0 degrees F or 105 degrees, I am not sure. I have done long exposure night shots (remember that long exposures produce more noise) in 7 degree weather and in 85 degree weather. I haven't noticed the differences in my images.
CMOS sensors work in a different way and they have their own tendencies with noise. I am going to skip CMOS sensors.
In photography, we measure light intensity based on units of stops. 2-stops of light means there is twice as much light as 1-stop of light. When we move from f/2.8 to f/4 we allow twice as much light through our lens. [Full stops increment by the multiple of the square root of 2. F/2.8 * sqrt(2) = f/4.0]
These numbers are a little rough. The human eye at any given instant is sensitive to approximately 6 or 7 stops of light, but with the changes in our pupil (our natural aperture) it is estimated that we can see around 10-11 stops in a scene. Our eye can also recalibrate itself to adjust to overall dark or overall light as we know from having our eyes get used to the dark. With this added feature, we can make visual use of about 20 to 24 stops, depending on whom you ask.
A digital sensor, however, is limited in sensitivity to about 5 or 6 stops of light. Many scenes we photograph contain more stops than this, therefore we have to make decisions about what we are going to capture and what we are going to let get away. We stand a chance of manipulating everything that is not 0 and is not 255. Those two pixel intensities have special considerations.
Example:
Here is an image, as shot and desaturated. This image contains pixels at 0 intensity down at the lower left and pixels of 255 intensity on the tops of the saw blades and on the spark guard up by the sharpener. It also has about everything in between.
Suppose I want to make this image darker, so I apply some aggressive downward curves to the image:
In this example, the pixels are lowered in intensity by a percentage of their starting value. (A non-linear percentage, at that.) This is true, EXCEPT for those at 0 and those at 255. If you study the image with an editor, you would find that what started at 0 is still 0 and what started at 255 is still 255. Notice how the bright spots are now disproportionately bright compared to the rest of the image. It is impossible for me to change those areas in any meaningful way. Those areas are called blown highlights. In general, we are best served to try to avoid blown highlights on our subject. On people, the nose, forehead, or cheeks can be prone to this, especially in sunlight.
In this next example, I've taken the curves the other way, making the image quite bright. Notice how the black areas still remain completely black (zero value) and are disproportionately dark as compared to the rest of the image. Depending on your scene, you may need to decide where 0 values are ok and where they are not. Due to the limitations of sensors, we will find that it is the highlights that warrant the most worry and not the shadows.
I have just said that we generally need to watch out for those highlights, but at the same time, there is a motive to try and expose pretty close to the limits of our sensor. It's like playing The Price Is Right but with light. We want to get as close as we can without going over. The first reason for this is that we optimize signal this way (assisting the signal to noise ratio) and the second reason is that there are progressively more and more levels available to us as we move toward brighter intensities.
There is already a wonderful article about this at Luminous Landscape but I will try to summarize it here.
An understanding of binary numbering is useful to understanding this.
Earlier I said that most sensors record data using 12-bits of color depth. This table represents the numbers to which this luminance data can be written. The CCD receives light in a linear fashion, but apparently, when the voltage is converted to binary numbers, that numeric space which contains this data is non-linear. As you can see in this table, each successive stop can store twice as many levels as the previous stop. This means that as you expose more and more (without going over) you are capturing more and more factual lighting detail about your image.
If we are to use an image like the example I posted, there would be a significant difference between capturing this image as a JPG (a space which converts all these into a 8-bit space) and pushing it to the brighter levels versus capturing it with a longer exposure in the first place. Not only does the brightening of the image further amplify any existing noise, but since the mathematical numbers are being shifted from the left side of the diagram which is less descriptive to the ride side of the diagram which is more descriptive (more available levels) the numbers do not always fall into place as elegantly as they would have if the sensor had captured them at that level in the first place. The result can be banding in your image. (And also bands in the histogram. I plan to cover histograms later.)
In this example of the Notre Dame cathedral, the first image is underexposed because the camera metered based on the sun behind it. When I shot again, I got the image I was seeking. In theory, the same ratio of image data should be there, but as we push up the levels with software we find that image #1 looks quite a lot different from image #2. The following factors are at work here: Image #1 saw less light, therefore the signal to noise ratio was worse. Image #1 stored the image data in the left end of the numeric system and therefore there was less detailed color level information about the image. The goal is to illustrate that a higher exposure is better, but we wish probably don't want to blow the highlights (expose all the way) of our subject. Keep in mind that the "subject" is up to you. In image #2 the top between the steeples clearly has blown highlights, but this is ok with me because I did not consider the sky as the important part of the subject. The correct answer and therefore the correct exposure comes from you. Sometimes an image is supposed to be dark.
Unfortunately there will always be situations where we can't avoid noise. When this happens and we wish to reduce it after-the-fact, I know of two useful programs. One of them is Neat Image and another is Noise Ninja Both get good reviews and I have used Neat Image somewhat extensively. This program has a few modes, but in the basic mode, it will select a portion of a noisy image and profile it for a baseline of how this image contains noise. Once that pattern is established, it subtracts this noise profile from the entire image. Since noise is usually distributed in a somewhat repeating pattern, this system works pretty well. Images can be left a little soft afterward.
The ability to change ISO sensitivity on the fly is rather convenient. It is roughly equivalent to changing film speeds at the change of a setting.
It comes with its tradeoffs. Now that we've thought about how pictures are exposed to the sensor and how these numbers are stored, we should be able to understand what is happening when we change the ISO setting.
First, understand that a sensor has a definite sensitivity to light. This sensitivity is fixed, it begins the moment the sensor was born and nothing ever changes this. One sensor may be more sensitive than another, but when it comes to your sensor, it always takes the same amount of light to equal the same voltage on the photo diode. No camera setting ever changes this.
When we change ISO sensitivity, we are telling the camera to underexpose the image by 1, 2, 3, or 4 stops and then to push the image by the same number of stops via its internal software before writing the resulting file. Remember we established that a camera is sensitive to 5 stops of light (technically probably 6 stops, but that lowest stop doesn't do us much good so we ignore it).
A given sensor's sensitivity is bound by physical characteristics. When a camera is set to use a higher ISO setting, the camera responds by exposing to a lower stop. Have a look at the next table.
This is the same as the table above but now I've added ISO settings. Normally a camera would meter with consideration of its rightmost ability on the chart. You'll notice that camera ISO settings almost always include provisions for 4 and not more than 4 ISO settings. As you look at this table, the reason should become more clear. A sensor that is rated ISO 200 is twice as sensitive as a sensor rated at ISO 100.
It may seem as though this could all be done in post-processing. If you are dealing with RAW files, then this is probably true, but if you are shooting TIFF or JPG it would be best to let the camera do this math because it is doing it with the original numbers rather than numbers that may have been rounded off or truncated by the file format.
One of the last topics I want to cover is the histogram. A histogram is a very useful tool and I started taking better photos as soon as I began paying attention to it.
The histogram is a graph that shows you visually how much of your picture is comprised of light at certain levels. Review of this can help you to know if you're getting the exposure you want. If the graph appears to be rolling off the right side, then chances are you are blowing a lot of highlights. If the bulk of it is in the center, you might be underexposing slightly. There is no right or wrong result of a histogram, it simply is. It is a report about your image.
Revisiting the examples of Notre Dame, let's study their histograms a bit.
In this one, we see a lot of what we already know. There are two spikes in the histogram. One corresponds to the dark area of the face of the building. The other one corresponds to the sky. There is a tiny blip at the right edge of the graph, this is the sun. Part of the sun is (255,255,255) therefore we can deduce that the brightest part of the image either met exactly (not likely) or exceeded the levels the sensor could capture at this exposure setting. Your software might give you other statistics, but on this screenshot you see a mean, median, and standard deviation.
I think most cameras can now be set to give you a histogram review. Some cameras and software can also give a break down of the histogram into the Red channel, Green channel, or Blue channel, or combine all of them in color on the graph. All of the examples I give are the plainest view of the histogram.
In this example the shutter was open for a longer period of time, thus the sensor received more light. The corresponding histogram contains more information to the right than the previous histogram. In studying the image, I can tell you that virtually all of the sky is (255,255,255) and there is a massive but thin spike at the rightmost edge of the histogram to prove it. The histogram reveals that even in the darkest areas of this image, none of it is pure dark, or 0. This histogram never touches the left side.
In this image of the Eiffel Tower, most of the sky is pure black, and the the majority of pixels in this image are fairly dark. The lit areas populate the bump on the right of the histogram.
In this histogram we can see that I left room for a significant number of levels. That empty space on the right probably leaves a whole stop. This however, is ok with me because I wanted to keep the lot of richness in the color of the sky. Unfortunately the further something is exposed to the right, the more is starts to resemble white and the less rich its color; in my experience. (Brighter colors move closer toward (255,255,255) which is white. I often struggle between good rich color and luminosity in my images, trying to decide which I want more. (I really desire both)
This is the same image as above, but with a heavy edit applied. I wanted to use it to give some kind of an example of banding. The spots you see are dust, most likely on the sensor, although sometimes I get this from dust on the lens. I change lenses frequently, and dust happens to me a lot (I blow it off with a nasal bulb that you'd use for a baby's nose). In the sky, notice the lack of a gradual transition. It goes in noticeable stripes from a pinkish hue, to turquoise, to royal blue. In reality the sky wasn't so quick to change. This image was exposed somewhere in the 4th stop of light, meaning that it only made use of about 50% of the total available light levels that were available to the sensor. Clearly this edit is a bit extreme, but this undesirable banding effect can happen from time to time and it is more prone when the image data resides in the middle or leftward part of the exposure range.
I hope this is useful. I encourage you to get out there and read up on these topics. There is much more information that I did not include.
You can think of all of these facets and facts the next time you release the shutter.
By Ryan Richardson