Digital Imaging Notes

Sometime I find published descriptions of digital printing techniques a little too vague/obscure for my needs. I then sometimes look a little more deeply  into how  these procedure actually work. When I find out something interesting I write a little note about it. I am going to post some of these notes here.

My first completed write up is on the NEED FOR RE-SAMPLING WITH EPSON PRINTERS. My results show that letting the Epson driver automatically do the re-sampling is not always a good idea. This write up is posted below.

The next write up will deal with the ADVANCED BLACK AND WHITE printing procedure now available on Epson photo printers.

 If you have comments on these notes, particularly about possible errors, please contact me ( eskingerry@aol.com ).  

_________________________________________________________________________

 

 

The issue discussed here is - at what pixels per inch (PPI) a photographic image should be submitted to an Epson Inkjet printer. At first one might think that any value will do, in that the printer will be happy to do the necessary conversion for you. Will it do the best possible job for you? The answer is frequently “no”.

Others have recognized that Epson uses the nearest neighbor (NN) algorithm to convert images prior to printing and this may result in undesirable artifacts creeping into the image. This recognition has led some to recommend that one should submit images to the printer only at resolutions that divide evenly into the dots per inch (DPI) setting of the printer. For example, because Epson printers often print at 1440x720dpi (and sometimes 2880 x 1440) the recommendation is make to print at settings such as 180, 240,360 or 720.

It is true that non evenly dividable PPI can cause problems. One way to see this is to produce a target of parallel black lines spaced one pixel apart and then see how the printer handles this at various settings. Here is such a target-image.

This particular target contains 290 pixels vertically (alternating black and white). When sized at 290ppi a print from this file will be one inch in height. The target, as presented here, has been enlarged some 14 time so that its detail can be seen clearly

 

To see what the printer does with this file we pick a PPI value that does not divide evenly into the printers, submit the image to print (in black and white) and observe the results obtained.

290ppi

As an example of a ppi value that does not divide evenly into 720dpi or 1440dpi we select 290ppi. As mentioned earlier, at this ppi (density of pixels) the target will produce a print that is 1 inch high. Below is a likeness of the print resulting from this setting (printer set at 1440dpi).

Rather than being uniformly grey, as would be desireable, there are distinct BANDS.

To see more clearly what is going on we enlarge a section of the print with a microscope and compare it to particular re-sampling scheme of interest, namely, we re-sample the file (in Photoshop) from 290ppi to 360ppi using the NN (Nearest Neighbor) method.

We note that the enlarged print show the same patterns as the Photoshop conversion. The lighter area is made up of alternating black and white line of  mostly one pixel width. Every other white line is of two pixel width. In the dark area, every other black line is of two pixel width.

We conclude from this example and some 20 other similar cases studied that:

• The printer always re-samples to 360 ppi before converting to dots (except when "finest detail" is selected in which case it re-samples to 720 ppi )
• It uses the NN (nearest neighbor) algorithm to do this.
• This is not the best way to re-sample photographic material.

To illustrate the NN problem, in a more realistic context, we re-sample a file containing a cartoon image.  We re-sample this 290ppi image to 360ppi by Bi-cubic (BC) and by NN and display facsimiles of the resulting print appearing as below:

Facsimile of Cartoon Image, scanned at  290 ppi and printed

NN has added a diagonal cross hatch pattern (moire pattern) to the image where BC has not.

If NN has problems and there are better approaches, why would Epson use the method? Only Epson knows, but my conjecture is that it is related to the primary use of most Epson printers- that is, to print text.

The above discussion deals with the situation where the PPI is not evenly dividable into the printer DPI.  Now we consider dividable cases, Some of the dividable settings can be dismissed out of hand. Generally 45ppi and 90ppi are too low to make high quality prints so we do not consider these cases.

240 ppi

This is an often recommended setting but it still does have problems. A slight modification of our target will help illustrate this. Recall that the target is made up of one pixel wide black lines spaced one pixel apart. We now add a second set of lines identical to the first but which are offset by one extra black line between the two sets. The file containing this target now has (290x2) +1 pixels vertically.

To study the 240ppi case we resize (but do not re-sample) to 240ppi.Submitting the file directly to the printer result in a print similar to below.

Something is still wrong, although this is one of the divisible cases.  The flaw in the argument about divisible cases is that it assumes the printer converts from the PPI directly to DPI. But this not the way the printer works, it first converts to a “native” PPI (namely 360ppi) and then uses a dither to represent each of these pixels with a set of dots.

720 ppi

 Next we consider the case of 720ppi. Again this case can be studied by resizing our target to 720ppi and submitting it to print. This results in a print similar to the following:

.

Here, we have the surprising result that half the printable area turns black and the rest is pure white. At first one would think that that a 720ppi would produce very high quality image but we must remember that with NN the printer is not using all of this information. It is in fact selecting one out of each 2x2 pixel areas to “represent” the area in the conversion from 720 to 360ppi. Our particular target has the characteristic that the “representative” pixels in the upper part of the target are all white and are black in the lower section.

180 and 360 ppi

So what”s left?. Are there any settings where it is safe to let the printer do the required resampling. The obvious case is 360ppi. This works because no resampling is done. This is the printer’s native resolution.

180 is also pretty good. This ppi setting is also free of artifacts even though the printer re-samples from 180 ppi to 360 ppi via NN. What the printer does in this case is replace every 180 pixel with four identical 360 pixels. The image will retain all of its original detail but will not be as smooth appearing as an image converted using bi cubic. Bi-cubic conversion, although smoother will lose some detail.

Conclusion

The safest way to avoid possible NN anomalies is to not depend on the printer to do re-sampling but to always submit images to the printer at 360 ppi. Either do the re-sampling in Photoshop using bi-cubic or use a plug in such as Genuine Fractals. Submitting at 180 ppi is also an option. 

 My Personal Preferences

1. Whenever I can, I avoid all re-sampling. I mostly shoot with a Canon 20D which naturally produces a   6.5” x 9.7” print at 360ppi or a 13.0” x19.5” print at 180ppi. Whenever possible I print at these sizes.
2. If an image needs re-sampling, I do it myself rather than let the Epson printer do it.

xxx