Better Than USM

This tech page is Photoshop specific, in search of better sharpening than the standard UnSharpMask filter. The test image is shown on the right (Jpeg compressed of course) please download the original PSD image for accurate evaluation. So, how to sharpen without using the USM? The Filter/Other/Custom is a rarely used filter, it surely does not look very attractive. However it is a very basic image enhancement operator, the convolution kernel. Most of the image enhancement filters do it by convolution but they usually have somewhat more friendly interface. The Custom filters can be saved to a file and the dialog allows both image preview and provides the small zoom box. When the filters are named meaningfully and saved to a directory of their own  the Custom filter becomes rather useful, and one can always record the Custom Filter operation as Actions, this makes them even more easy to use.

Jpeg copy of the Original  (do not download this)

For sharpening both Blur and Sharpen kernels are needed, please download them here.  Unzip them into a directory of their own and they can be easily loaded directly from the Filter/Other/Custom filter. 

The Blur convolution filters are based on the first order Gaussian low pass filter and they provide blurring at 1%, 2%, 3%, 4%, 5%, 10%, 15%, 20%, 25%, 30%, 35%, 40%, 45%, 50%, 65%, 80%, 100%. These percentages relate to the standard Gaussian Blur function in Photoshop in such way that 100% give about the same amount of blurring than radius 0.6 in the Photoshop Gaussian Blur filter. The radius parameter in the Gaussian Blur filter in Photoshop works rather oddly, radius 0.1 does not affect at all, radius 0.2 has only a small effect and then there is a large perceptual change from 0.2 to 0.3 and so on. Also the Photoshop Gaussian Blur is not actually Gaussian. So the convolution filters gives much finer control over the blurring.

The Sharpen convolution filters are based on the Laplacian at various gain and they provide sharpening from 50% to 750%. These percentages are approximately the same as the amount percentage value in Photoshop USM when radius is set to 0.5 and threshold to 0.  These sharpening convolution filters do not give as fine control as the Photoshop USM that (seemingly) allows 1% stepping. However the 50% steps are rather adequate. Most importantly the Photoshop USM has features that makes it somewhat inefficient when compared to by-the-book sharpening.

Evaluation

At this time I provide only one example. I strongly recommend the reader to evaluate themselves, it is truly beneficial.

What is considered to be good, better or best image appearance is of course in the eye of the beholder, even if there are some quality factors that can also be measured. 

In the below example, in my opinion, the quality difference is very apparent. You may want to evaluate at other USM settings, (since I have spent a quite a while in reverse engineering the Photoshop USM parameters I can say that there is no USM settings that gives anything near to the  the Blur 30% + Sharpen 500% appearance).

Please evaluate under a dim or dark ambient lighting. 

	Convolution in Photoshop by: Blur 30% followed by Sharpen 500%	Photoshop USM at: Amount=500, Radius=0.5, Threshold=4
	Black (e.g. bottom-right) is better. Sharpening is better in the shadows. Colors are the same as in the original. More details are seen.	Smudge general appearance & colors change & high noise in the shadows & fine detail is lost.

A closer look at 200% zoom from the right eye (on the left on the monitor) is shown below:

So, there is a difference clearly. Which one do you prefer?

Photoshop Convolution is Buggy

During the above exercise I noticed that Photoshop has a bug in the Filter/Other/Custom -filter (the convolution filter), it simply calculates incorrectly. It can be seen by setting the center box value to 9 or higher (and all other to zero or empty) and setting the scale to the same value. Image should not be altered at all by such a kernel but it changes.  Only if the center box and the scale are set to value 8 or below the image image does not change. The image should not change at all since it is just the calculation of 
9 * channel_value / 9.  Now only kernels that have the divisor smaller than 9 operate properly, and there are only very few such usable kernels.

In order to experiment how the bug affects I made exactly the same filtration as above also in Corel Photo-Paint v.8 that calculates the convolution correctly. The side by side comparison is below.

You probably can easily see that Photoshop version has higher overall noise, has much less fine details and it is less sharp. The noise in the deep shadows is slightly smaller in Photoshop version, this is due to the fact that the bug in Photoshop usually makes the RGB level to be less than it should be.

Convolution using Photoshop by: Blur 30% followed by Sharpen 500%

Convolution using Photo-Paint by: Blur 30% followed by Sharpen 500%

It is possible that the same bug is in other Photoshop filters/operations also, this would explain the poor performance of e.g. the Gaussian Blur and USM.

I have also experimented the convolution sharpening using the FilterFactory, this way the Custom filter bug in Photoshop can be by-passed.