Introduction
Users of that strange piece of transparent and flexible celluloid (with a layer of sensitive silver-halogenid grain suspended in gelatin attached to it) as a means of definitive recording of a slice of reality have no choices: the captured image is sharp or it is not. Some manipulation in the darkroom or at the printing stage might save a slightly unsharp picture, but basically the sharpness problem is a matter of good optics, accurate focusing and the characteristics of the film emulsion/developer interaction. The film emulsion is a faithful medium: what the lens projects onto the surface is recorded and fixed. It is then logical that lens designers take great care to design the best possible lenses: the lens is at the head of the imaging chain and the most important transmitter of the shapes and textural details of the scene in front of the camera. The film can be inspected visually immediately after development, but the negative (or positive in the case of transparency film) cannot be altered. The image is fixed in the arrangement of the clumps of grain in the gelatin layer. Printing the negative is a simple matter of projecting the negative image onto a sensitive paper surface.
Basics of image quality
For a long time, it has been noted that the image quality is based on a few parameters, tonality or gradation of the differences in light intensity (for the plasticity of the image), the crisp delineation of the subject shapes (also known as the acutance of the image) and the clean recording of the fine textural details of the subject surfaces (also known as the definition of the image).
Optical and scientific analysis has indicated that the contrast transfer function (MTF) is a reliable measure of the optical capabilities of a lens or lens/film combo to record these aspects of the image. Since then we are all familiar with the MTF graph, which essentially has the same universal shape: contrast is high when few image details (linepairs per mm) are recorded and contrast is low when many lp/mm are recorded. Between both extremes the curve has a downward sloping shape.
Characteristics of CCD/CMOS capture media
In current times the medium for recording an image is the CCD, an array of light sensitive pixels. CCD’s store images in the form of electrons that are sensed as an analog voltage at the output stage. These voltages are digitized by discrete sampling methods (10, 12 0r 14 bit) and processed to create a digital representation of the image, where every color layer has its own image and each image is stored with 8 or 16 bits of information. The image processing then has to accomplish spatial interpolation and color matrixing to produce an image that humans can see and interpret. This reconstruction of the original image is both the strength and weakness of the current so-called digital imaging chain.
In the domain of digital imagery an image is defined by rows of numbers in a digital file. We know that any number in a computer file can be changed and regrouped. The whole optical image processing science is based on this principle and it even possible in Matlab or Mathcad to recreate images form a matrix of numbers.
It is also well known that the mathematical manipulation of the numbers can enhance the basic features of an image: detection of a subject contour line is easy and it is simple arithmetic to change the values of individual pixels. This is how sharpening algorithms work: the original image is blurred slightly, then the blurred version is compared with the original version on a pixel for pixel basis. If a pixel is brighter (has a higher value) than the blurred version it is lightened further and if a pixel is darker than the blurred version it is darkened more. Thus the contrast between pixels (image details) is enhanced and we can see the difference more clearly. Note that no additional detail is reconstructed: only the basic information is exaggerated. Here we can see the validity of the Leica claim that the lens still has to very good otherwise the details can not be recorded and afterwards not reconstructed.
Sharpening as defined by Photoshop and its mathematical base.
On the other hand do we have to admit that the reliance on high contrast lenses with excellent MTF values is less important as the post processing (sharpening) algorithms can enhance edge contrast with the unsharp mask option in as example Photoshop. The three parameters (amount, radius and threshold) define how many pixels have to taken into account when calculating the value of the individual (source) pixel (radius), how the differences are exaggerated (amount) and at what value the sharpening should start (threshold). Film users know about this when they refer to the acutance quality of film/developer combinations.
All raw developer programs offer sharpness algorithms with more or less sophistication and more or less options. To add to the confusion, the numbers are not always referring to the same values (a radius of ‘1’ in program A is not identical to the same value in program B. And when the numbers refer to the same aspects, the results may differ: an amount of ‘200’ in program A and B might produce different results. As was the norm in the true darkroom days: long experience and daily exposure to the intricacies of the art are necessary to master the craft. It is no different in the era of the mathematical post-processing programs.
All programs, however address the sharpness issue from the same base: the numerical value of an individual pixel is being used as a comparison with values of surrounding pixels and the difference between these values is input for the Fourier transforms and other signal processing functions. Every program has its individual character, not unlike what we are accustomed to in the darkroom with different chemicals and films. Basically the unsharp mask is a kind of edge detection program or a signal enhancement program.
The main questions I am interested in can be stated as follows:
(1) what is the relation between the optical performance of the lens and the post processing algorithms.
(2) what are the differences between these algorithms as executed in the raw converters.
The test setup.
A test chart has been photographed with the Leica M8, with a Summilux-M 1.4/50 Asph. attached to it (camera and lens are owned by me). The test chart has been aligned to the vertical plane with great exactitude. The camera was set in front of this test chart and with the help of an infrared device precisely aligned to the test chart. The chart consists of a number of the Siemens star patterns, with sinusoidal grating. The camera has been adjusted to a zero tolerance for the rangefinder, but even so a series of test images was made with focusing bracketing to be sure that the plane of maximum sharpness was focused.
The images were then 'processed' by a number of Raw processors (Aperture, Photoshop CS2, Lightroom, Picture Window Pro, LightZone, Bibble Pro, Capture One 4.00 and in addition the original JPG file from the M8 was also analyzed).
The sharpness parameters of the Raw processors were adjusted to produce several levels of sharpness, using all possible settings in advanced mode. As I am interested in this first test in the MTF values and maximum definition that can be generated by the processors, all images were processed with a specially designed program that can create MTF graphs from Siemens star patterns. The resultant graphs were cut off just beyond the Nyquist frequency limit: beyond this value Moire starts to mess up al graphs. See below the test chart set up. All analysis is done on the centre spot of the central star pattern in full picture area.
siemensster
Base-line situation
The lens/optical system has its inherent MTF performance (this as been explored in the companion article:
To understand what the post processing does, we need to find a base line as a reference. The program Picture Window Pro is the only one I know of where the input parameters for raw processing can be adjusted. This one was used to read in the basic M8 DNG image with all parameters set to zero. The result is the MTF Graph reproduced below. Left top the image as recorded by the lens, right, the image as processed by Photoshop CS2, USM filter and left bottom the original M8 file in JPEG mode.
The shape of the line does follow the shape we expect form the optical MTF from the lens alone: maximum contrast is below 100% and the line gently drops to the cut-off frequency, where about 10% contrast is recorded. As a comparison the the straight JPG image from the M8 is ssen below the Basic Raw file. . Note that there is a heavy amount of sharpening for an MTF boost) in the low and medium frequencies, but the very fine detail is lost. The MTF graph as developed by Photoshop (CS2 version and USM) shows a very aggressive boost of edge contrast for the whole image. Specifically the mid frequencies get a true turbo boost. This boost is accompanied by a loss of subtle grey tonal values and an increase in More at the border of the Nyquist frequency.
In most reports you will find a great interest in the calculated Nyquist values as they are thought to represent the maximum resolution of the sensor. Theoretically this is correct. Based on practical tests one should use a more realistic approach: 70% of the Nyquist frequency is the best one can hope to get in most situations. The mid frequencies are indeed more important than the maximum resolution. This conclusion does not differ from the experience with film emulsions.The maximum resolution is possible when the image details are precisely aligned to the pixels on the sensor. But most often the image details cover more than one pixel. This phenomenon will reduce the effective resolution substantially. See below the diagrams taken from a Rodenstock paper, covering this topic. Note that skew lines will reduce definition quite substantially. And the world of imaging has abundant skew details.
It is evident that the improvement in the MTF graph as delivered by the raw processors is impossible to get with improved optics alone. The conclusion I have presented several times in previous articles that the power of the raw processing is of such a magnitude that the optical engineer will not be able to approach this level of performance without extreme efforts.
The interplay of factors between optical design and construction and mathematical image processing will change the landscape of camera optics and the evaluation/importance of the basic optical parameters. (by the way: signal processing is focused on one dimensional data (or temporal data) and image processing is focused on two dimensional data (or spatial data).
The role of optical design is still very important, especially as we noted and demonstrated that the definition of fine detail is still the exclusive realm of optics, the software is not able to ad detail definition where none exists. The very important mid frequencies are responsible for the visual impact of images.
It is also true that the common approach by most image analysts (to analyze the image quality on screen and after the raw data have been processed by some program) is not valid and does not give a correct impression of the quality of the lens. I compared the Leica MTF graph after processing in CS2 with a graph from a cheaper Canon lens, also processed by CS2 and the differences in the mid frequencies were negligible. The true differences on film and by MTF analysis were substantial. This shows that the software has become an important part of the imaging chain. Performance analysis of a lens with the help of raw developers is a non-issue: what can be looked at is the combined effect of the lens quality and the power of the raw processing engine. The raw processing engine has the most impact and all you can derive from an analysis of a lens/camera sensor /raw converter combination is the final impact of the image. You cannot discuss the influence of the individual components, unless you have the equipment to separate the effects.
The impact of the Raw processing engines.
It is evident by now that the impact of the raw processing engine is quite substantial and that there are large differences in effect when the sharpness parameters are changed. Small changes in value already have a big impact on the final image. It is also a fact of life that sharpness settings that look good on screen are not best for printing. A program like Nik sharpener or Q-Image or a book like
Below are graphs of the three important raw developers, Photoshop CS2, Aperture(top row left and right) and Lightroom (bottom row (left). The graph as made by CS2 is different from the one above in this article and the changes in amount, radius and threshold are quite small. Note the big difference in MTF graph. Aperture and Lightroom do not differ in their main characteristics, but note that Lightroom has better results for lines at different orientations.
The Siemens star has been divided in eight pie parts, representing different angles for the lines in the star pattern. I have shown here only four (the left half of the star. Because of symmetry effects these four parts represent the whole picture. The Aperture diagram shows that the quality of the image depends a bit on the orientation of the details. The CS2 graph shows the quite aggressive sharpening of the main mid frequencies and explain in part why strong sharpening will give this artificial edge enhancement effect of many digitally manipulated images. Aperture and Lightroom are more photographic in their approach.
The specialist raw developers
Bibble Pro, Lightzone, capture One and Picture Window Pro have a more limited scope than the mainstream colleagues.
Their appeal is none the less quite high. Every program has its strengths and its own ergonomics and work flow. Capture One has an intriguing behavior, that is focused on a smaller bandwidth of mid frequencies than does Photoshop. The net result is a picture with strong visual impact and still a classical photographic fingerprint. It is not the best program money can buy, but it offers a smooth workflow. The detail definition of the very high frequencies lags slightly behind the best developers, but as noted above, one should look at the overall performance and not pay too much attention to the small region around the Nyquist limit.
Below I have presented some graphs, made with one program (PWP) to indicate the range of differences that can be generated when one changes the parameters. Clockwise: PWP no sharpness, moderate sharpness and high sharpness.
The upshot is that every raw developer and its choice of sharpness parameters have a major impact on the final result. Even more impact than can be created by the characteristics of a lens. An excellent lens can be downgraded by the wrong choice of developer and/or parameters and a modest lens can show great performance by an expert choice of developer and set of parameters.
The corollary is then that a review of a camera/lens is incomplete without mentioning the developer used and the settings chosen
Visual inspection on a screen, even at 200% or more, of the image gives a clue, but is no substitute for a true analysis. In the past when chemical development processes were used to develop a strip of celluloid, it was customary to note film, developer type, development time and other parameters. This information was needed to assess the picture quality.
The graph below summarizes the results.
graphC
Conclusion
The final quality of the printed image depends on a number of factors and in the current state of the art the post processing software is the main factor of influence. Especially the spatial processing algorithms (sharpness, detail definition) have a significant impact on the quality. The role of the lens (the basic optical quality) is not diminished, bacause the programs cannot recover detail that has not been captured in the first place. But the edge contrast of the mid frequencies (that define the visual impact of the image) is now the province of the raw developers. The programs are not created equal and they may bring unexpected results. Just as is the case with chemical developers, you cannot claim that one developer is the best. You need to experiment and find the one that gives you the best images. The ‘best’ image on screen is not the ‘best’ image for print and it also depends on the size of the print and the viewing distance. Film based photographers will recognize this. There are no new insights in the world of digital photography. The knowledge and experience of the film based photography is not obsolete, only transformed into a new language and a new set of tools.
The current obsession with numerical results and screen analysis of images will hopefully be superceded very soon by a more promising focus on the final result of the workflow: the high quality print. The current quality of the baryta inkjet papers and the inkjet printers and inks is such that one can now produce fine art prints of excellent imagery. Computer screen projection is no substitute for a good print. The study of individual pixels on the screen is not the best approach to good photography. The raw developers are very powerful tools, but the classical relation between sharpness, graininess and tonal gradation still holds.