Canon XSi Imatest Results

Canon XSi Imatest Results

We routinely use Norman Koren's excellent "Imatest" analysis program for quantitative, thoroughly objective analysis of digicam test images. I highly recommend it to our technically-oriented readers, as it's far and away the best, most comprehensive analysis program I've found to date.

My comments below are just brief observations of what we see in the Imatest results. A full discussion of all the data Imatest produces is really beyond the scope of this review: Visit the Imatest web site for a full discussion of what the program measures, how it performs its computations, and how to interpret its output.

Here's some of the results produced by Imatest for the Canon XSi:

The Canon XSi showed excellent color accuracy, as well as fairly accurate saturation levels. Hue accuracy was very good, with only minor hue shifts occurring mainly in the cyans, oranges and reds. Average saturation was 105.2% (5.2% oversaturated) and average "delta-C" color error was only 3.93 after correction for saturation, which is very low. (Delta-C is the same as the more commonly referred to delta-E, but delta-C takes into account only color differences, ignoring luminance variation.) Overall, a very good color response for an SLR, especially a consumer oriented model. Mouse over the links below the illustration above to compare results with other recent models.

As is true of most SLRs, when using the Adobe RGB color space (which provides a much wider gamut, or range of colors that can be expressed), the Canon XSi delivers more highly saturated color, with an average saturation of 111.2% and average saturation-corrected hue error of 5.14 "delta-C" units. Again, mouse over the links below the illustration above to compare results with other recent SLRs.

This image shows how the Canon Rebel XSi actually rendered the colors of the MacBeth chart, compared to a numerically ideal treatment. In each color swatch, the outer perimeter shows the color as actually captured by the camera, the inner square shows the numerically correct color after correcting for the luminance of the photographed chart (as determined by a second-order curve fit to the values of the gray swatches), and the small rectangle inside the inner square shows the numerically correct color, without the luminance correction. This image shows the excellent hue accuracy, as well as a gamma curve that results in an overexposure of highly saturated swatches. (This seems to be a fairly common tactic, to produce "bright" color that's appealing to consumers, without further overdoing the saturation.)

There's a lot in this particular graph, a lot more than we have room to go into here. (This set of plots has also changed a little in the more recent versions of Imatest. Some of the plots that were shown here previously are now shown in other Imatest output. Since we largely focus on the Noise Spectrum plot, we'll only show the graphic above, which includes that plot.)

In comparing these graphs with those from competing cameras, I've found that the Noise Spectrum graph at lower right is the most important. Cameras that manage to shift their noise spectrum to higher frequencies have much finer-grained noise structures, making their noise less visually objectionable. In the graph above, this would show up as a noise spectrum curve that remained higher on the right side, representing higher noise frequencies. The champion at this was the Canon EOS-1Ds Mark II, which produced remarkably fine-grained image noise, even at very high ISOs.

The Canon XSi does an excellent job of keeping plenty of the luminance noise energy (indicated by the black line) at high frequencies and is better than most cameras in this respect. What little low-ISO image noise that's there is more fine-grained as a result. Chroma noise is another matter though: The blue and especially red channels are quite a bit noisier than the others, but that's not at all unusual, and only slightly visible when inspecting shadow areas of low ISO shots very closely. Canon says their High ISO noise reduction (Off by default) is also effective at reducing shadow noise at low ISOs, but we didn't verify that claim.

Here's the same set of noise data at ISO 1,600. Here, the Noise Spectrum graph is shifted quite a bit toward the left-hand, lower-frequency side than it was at ISO 100, coarsening the "grain" of the image noise patterns quite a bit. The red and blue channels still exhibit more noise at the lower end of the frequency spectrum, which can be seen as slighter larger blotches of chroma noise. Enabling the Canon XSI's High ISO Noise Reduction (which is Off by default), eliminates much of the chroma noise, resulting in very pleasing results at high ISOs.

This chart compares the Canon XSi's noise performance over a range of ISOs against that of other cameras. While we continue to show noise plots of this sort because readers ask for them, we each time point out that the noise magnitude is only a small part of the story, the grain pattern being much more important. Here, we can see that the Canon XSi's luminance noise starts out very low compared to the competition, and increases at a fairly constant rate as ISO climes to 1,600, where it ends up close to the level of the Nikon D60. and slightly higher than some of its competition. However, the grain pattern is quite fine and tight, resulting in less visibly objectional noise as well as better detail retention. A very good result, especially considering the smaller pixels of the 12-megapixel XSi compared the the 10-megapixels models in most of its peer group. Do keep in mind these measurements are taken with each camera set at default settings, so the shape or position of the curve could be influenced by the settings you choose to use. The Pentax K200D's plot is a good example of this. Its noise magnitude is higher than the others, but that's partly because the K200D's default contrast and sharpness settings are a bit on the high side.

Canon XSi Dynamic Range Analysis

A key parameter in a digital camera is its Dynamic Range, the range of brightness that can be faithfully recorded. At the upper end of the tonal scale, dynamic range is dictated by the point at which the RGB data "saturates" at values of 255, 255, 255. At the lower end of the tonal scale, dynamic range is determined by the point at which there ceases to be any useful difference between adjacent tonal steps. Note the use of the qualifier "useful" in there: While it's tempting to evaluate dynamic range as the maximum number of tonal steps that can be discerned at all, that measure of dynamic range has very little relevance to real-world photography. What we care about as photographers is how much detail we can pull out of the shadows before image noise becomes too objectionable. This, of course, is a very subjective matter, and will vary with the application and even the subject matter in question. (Noise will be much more visible in subjects with large areas of flat tints and subtle shading than it would in subjects with strong, highly contrasting surface texture.)

What makes most sense then, is to specify useful dynamic range in terms of the point at which image noise reaches some agreed-upon threshold. To this end, Imatest computes a number of different dynamic range measurements, based on a variety of image noise thresholds. The noise thresholds are specified in terms of f-stops of equivalent luminance variation in the final image file, and dynamic range is computed for noise thresholds of 1.0 (low image quality), 0.5 (medium image quality), 0.25 (medium-high image quality) and 0.1 (high image quality). For most photographers and most applications, the noise thresholds of 0.5 and 0.25 f-stops are probably the most relevant to the production of acceptable-quality finished images, but many noise-sensitive shooters will insist on the 0.1 f-stop limit for their most critical work.

The image below shows the test results from Imatest for an in-camera JPEG file from the Canon XSi with a nominally-exposed density step target (Stouffer 4110), and the XSi's Contrast (0), Auto Light Optimization (On) and Highlight Tone Priority (Off) settings at their default positions.

These are really excellent numbers compared to current-model DSLRs, and in part, reflect the effectiveness of the XSi's new Auto Light Optimization feature, which is On by default. With ALO set to Off, dynamic range at the high quality level drops from 8.01 to 7.44 EV, while total dynamic range falls a full stop from 11.6 to 10.6 EV.

Processing the Canon XSi's RAW (.CR2) files through Adobe Camera Raw (ACR) version 4.4.1 actually reduced the dynamic range score by about 1/3 stop at the highest quality level, compared to the in-camera JPEG with Auto Light Optimization (ALO) set to On (the default). Compared to a JPEG with ALO Off, ACR was able to improve DR slightly, by about 1/4 f-stop from 7.44 to 7.68 EV at the highest quality level. These results were obtained by using ACR's automatic settings; slightly better results may be possible by adjusting the sliders manually. It's also worth noting here is that ACR's default noise reduction settings reduced the overall red- and blue- channel noise somewhat (see the plot in the lower left-hand corner) relative to the levels in the in-camera JPEG. ACR also shifted slightly more of the noise energy into high spatial frequencies, making its residual noise even less objectionable than that found in the camera JPEGs.

Dynamic Range, the bottom line:

The net result was that the XSi performed really well when compared against most current DSLR models, positioned near the top of the current crop of DSLRs for in-camera JPEG dynamic range results. Results from processed RAW files were also good, but not quite as impressive because of the lack of ALO for RAW files.

To get some perspective, here's a summary of the Canon XSi's dynamic range performance, and how it compares to other digital SLRs that we also have Imatest dynamic range data for. (Results are arranged in order of decreasing dynamic range at the "High" quality level.):

The results shown in the table are interesting. One of the first things that struck me when I initially looked at test data for a wide range of d-SLRs, was that here again, purely analytical measurements don't necessarily correlate all that well with actual photographic experience. There's no question that the Fuji S3 Pro deserves its place atop the list, as its unique "SR" technology does indeed deliver a very obvious improvement in tonal range in the highlight portion of the tonal scale. I was surprised to see the analytical results place the original Olympus E-300 as highly as they did, given that our sense of that camera's images was that they were in fact noisier than those of many other d-SLRs that we looked at. In the other direction, I was quite surprised to see the Nikon D2x place as low on the listings as it did, given that we found that camera's shadow detail to be little short of amazing.

One thing that's going on here though, is that we tested each camera at its base ISO setting, which should produce best-case noise levels. This is in fact what many photographers will be most interested in, but it does perhaps place some of the Nikons (like the D40) at a disadvantage, as their lowest ISO setting is 200, as compared to the ISO 100 settings available on most other models.

All the above said, it's clear that the Canon XSi delivers really excellent dynamic range, clearly among the best on the market at this time. (It's particularly interesting that cameras like the Canon XSi and Nikon D40x are now showing higher dynamic range figures for their camera JPEGs than did the legendary Fuji S3 Pro, which had a special dual-sensor CCD in it, designed specifically for high dynamic range.)

Canon XSi Resolution Chart Test Results

The chart above shows consolidated results from spatial frequency response measurements in both the horizontal and vertical axes. The "MTF 50" numbers tend to correlate best with visual perceptions of sharpness, so those are what we focus on here. The uncorrected resolution figures are 1,980 line widths per picture height in the horizontal direction (corresponding to the vertically-oriented edge), and 1,897 lines along the vertical axis (corresponding to the horizontally-oriented edge), for a combined average of 1,939 LW/PH. Correcting to a "standardized" sharpening with a one-pixel radius increased vertical resolution by quite a bit, but horizontal resolution increased only slightly, resulting in an average of 2,157 LW/PH. This is somewhat lower than you'd expect for a 12-megapixel camera.

To see what's going on, refer to the plots below, which show the actual edge profiles for both horizontal and vertical edges, in both their original and corrected forms. Here, you can see that a nearly ideal amount of in-camera sharpening is applied in the horizontal direction (undersharpened by only 2.38%, explaining why standardized sharpening wasn't able to improve on the MTF 50 numbers much), while in-camera sharpening is not as aggressive in the vertical direction (13.1% undersharpened). Despite the "undersharpening" reported by Imatest though, there is a noticeable "bump" visible in the horizontal edge profile. (A very slight one in the vertical edge profile, but probably negligible.) You should thus turn the camera's sharpening down a little for optimal results when sharpening post-exposure in Adobe Photoshop or other image editing software.