Up&Down Sizing

Up Size Down:
Scaling Bitmap Pics Saves a Bundle on Your Digital Camera


One of the distinct advantages of vector drawings and graphics produced in programs like Adobe Illustrator, Corel Draw, and Macromedia Freehand is that they are nearly infinitely scalable - 1/100x to 100x with little or no distortions. Trying to do the same with bitmap images in Adobe Photoshop or Corel PhotoPaint and any scaling by a factors greater than 1/2x to 2x produces some rather disturbing results. Figure 1 shows the obvious drop in image quality when the factors are 3x and 1/3x respectively. Obviously the 1/3 downsizing (bottom most image in Figure 1) has the most severe distortions. But upsizing by a factor of three has also resulted in a loss of sharpness and perhaps color saturation in the image (top is original, second is 3 times upsized image in Figure 1). In fact, there is a rather prosperous cottage industry for image-upsizing using RIP-Rasterizing Image Processors and/or advanced DSP - Digital Signal processing hardware and fractal or wavelet software that can scale 8x10 images into the advertising poster sizes with relative impunity but at a rich cost.


Figure 1 - Bitmap Image Distortion with Upsizing and Downsizing

So scaling bitmap images is not a trivial problem. In fact programs like Adobe Photoshop and Live Motion 2, Deneba Canvas and Macromedia Flash provide hybrid image processing capabilities which allow for combining vector draw with bitmap images in an attempt to provide the best of both worlds. In these programs users are able to transform vector drawings with many of the familiar bitmap filters and color corrections. Also bitmap and vector layers can be combined together with some stunning effects. But despite these advances problems of scaling bitmap images persist. And because bitmap graphics comprise over 2/3rds of all images used on the Web, upsizing and downsizing of bitmap images is a major problem encountered by Web and graphic designers everywhere.

Fortunately, the introduction of some new image processing technologies have changed the picture, so to speak, and now users of bitmap images have a fair bit more leeway to play with in upsizing and downsizing their images. In addition it appears that the popular paint editor vendors like Adobe, Corel, and Jasc have quietly improved their image resizing/scaling tools adding to or refining their algorithms. So not to lose step, this survey examines carefully to the point of looking upsize down at scaling and resizing of bitmap images.

The Improvements in Resizing

Two vendors, Celartem Technologies and LizardTech have over the past few years, developed software that promises to both up-size and down-size well past the traditional 2x or 1/2x barriers. These programs use advanced compression technologies derived from fractal and wavelet methods which have proven successful already. In fact wavelets are being used extensively in the new JPEG2000 standard for

their superior results. As well the resizing algorithms used by the major bitmap editors have also improved based on our recent tests. Specifically for the upsizing we have taken a 300 x 200 pixel image and enlarged it by factors of 5x and 7x (1500 x 1000 pixels or 25x larger filesize and 2100 x 1400 pixels or 49x larger filesize respectively). For many of the methods we found this was the practical upper limit for enlargement as edge sharpness and color bleeding became very notable in the 6x or greater enlargements. Only VFZoom and Genuine Fractals could push beyond this point. However, as can be seen from Table 1, the fidelity in rendering colors was very favorable by all the methods - unfortunately, this accuracy fell of notably in downsizing.

Color Fidelity Tests

After upsizing and downsizing we took readings use Adobe PhotoShop 7's Color Sampler Tool at many points in the re-sized images and compared them to equivalent readings from the original image. The results were quite encouraging on upsizing.

All upsizing methods produced excellent color fidelity. Table 1 shows just 4 of the 16 points sampled in each image. In general many results were within 3 to 10 points of the true value - it is almost impossible to discern any difference with the naked eye. However, there is a tendency to make the image slightly brighter which is even more pronounced when the original hue is already bright.

 

 Table 1 - Color Fidelity Upsizing
Measured Using Photoshop 7 Color Sampler Tool

Scaling Method

 RGB Pt.1

 RGB Pt.2

 RGB Pt.3

RGB Pt.4

Original Upsize image

 227.73.9

246.228.250

 214.71.149

 204.32.106

Genuine Fractals 2.5 7x upscale

 227.71.9 -2

 248.230.251 +5

 214.72.150 +2

 204.32.105 -1

VFZoom 2.2 5x upscale

 228.75.9 +3

 248.231.252 +7

 214.72.151 +3

 204.32.104 -2

S-Spline Pro 2 228.74.11 4 248.233.247 4 213.72.149 +1  

Photoshop 7 Bicubic 5x upscale

 227.73.9 0

 249.233.252 +10

 213.71.151 +1

 204.33.107 +2

Photoshop 7 Bilinear 5x upscale

226.73.8 -2

249.234.252 +11 

 215.74.153 +8

203.31.105 -3

Photoshop 7 Nearest Nbor. 5x up

228.73.10 +2

249.234.252 +11

214.73.152 +5

204.32.106  0

PhotoPaint 10 Aliased 5x upscale

 229.73.9 +2

 248.234.252 +10

214.74.151 +5

 204.32.106 0

PaintShop Pro 7 Bicubic 5x up

226.73.8 -2

249.233.252 +10

214.72.151 +3

203.32.104 -2

PaintShop Pro 7 Bilinear 5x up

226.73.8 -2

249.234.252 +11

214.74.153 +7

203.29.103 -6

PaintShop Pro 7 Smart 5x upscale

228.72.8 -1

249.232.252 +9

215.73.152 +6

204.32.106 0

PaintShop Pro 7 PixelResize 5x up

229.73.8 +1

249.234.252 +11

215.75.153 +9

205.30.104 -3

Note:Net Change Calculations=Observed-Original= 227.71.9 - 227.73.9 = 0 -2 +0 = -2

 Table 2 - Color Fidelity Downsizing
Measured Using Photoshop 7 Color Sampler Tool

Scaling Method

 RGB Pt.1

 RGB Pt.2

 RGB Pt.3

RGB Pt.4

Original Downsize image

255.255.255

9.208.31

186.34.46

235.14.23

Genuine Fractals 2.5 7x downscale

254.255.254 -2

4.217.32 +5

199.31.38 +2

228.7.16 -21

VFZoom 2.2 5x downscale

254.255.254 -1

 9.213.33 +7

179.28.36 -23

225.8.18 -21

Photoshop 7 Bicubic 5x downscale

255.255.255 0

 7.215.32 +6

178.21.32 -35

222.6.15 -29

Photoshop 7 Bilinear 5x downscale

255.255.255 0

10.215.36 +13

173.24.32 -37

222.5.13 -32

Photoshop 7 Nearest Nbor.5x down

255.255.255 0

10.215.36 +13 

182.22.32 -30

222.7.15 -28

PhotoPaint 10 Aliased 5x downscale

255.255.255 0

9.217.36 +14

170.18.37 -41

220.5.12 -35

PaintShop Pro 7 Bicubic 5x down

255.255.255 0

6.216.33 +7

180.28.36 -22

221.1.10 -40

PaintShop Pro 7 Bilinear 5x down

255.255.255 0

6.215.32 +5

180.28.35 -23

228.10.14 -20

PaintShop Pro 7 Smart 5x down

255.255.255 0

8.216.34 +10

179.29.38 -20

222.7.15 -28

PaintShop Pro 7 PixelResize 5x down

255.255.255 0

11.216.39 +18

180.28.36 -22

221.6.13 -32

Note:Net Change Calculations=Observed-Original= 254.255.254 - 255.255.255 = -1 +0 -1 = -2

However, about 25% of the time colors were rendered darker. Rarer still were the occasions when color readings jittered - brighter in one color hue, darker in another.

Table 2 summarizes the color fidelity results when downsizing. Here measurements were a bit more problematical as the downsizing was by a factor of nearly 1/5 (1/25 the filesize) from the original - so many features got "imploded". Still the results were fairly consistent across about 20 readings. As in upsizing, solid areas of color were preserved very well as seen in the first column pure white of Table 2. But edge colors tended to shift more as different hues were forced to mix in compression. Thus color fidelity was notably worse for downsizing as images took on a slight grayish pallor - particularly photo images having many detail colors as in our sample on the "pebble" background.

In sum, upsizing color fidelity is uniformly very good while downsizing loses fidelity at the edges and color mixing areas to the degree of just noticeable shifts either brighter or darker depending on the contrast in the local region. Genuine Fractals, VFZoom were first and second consistently in both up and down sizing color fidelity. The third best for upsizing were very close with Bicubic methods nosing out the others. In contrast, for downsizing color fidelity, the Jasc Bilinear method seemed to hone in better than the other methods for third spot.

Image Fidelity

Again image fidelity seemed to be much better preserved on upsizing and most notably by

 Table 3 - Commentary on Upscale Image Fidelity

Scaling Method

Comments

Original Upsize image

Pink floral with solid gray border, very sharp

Genuine Fractals 7x upscale

soft at border, ultra thin color needles, sharp

VFZoom 5x upscale

soft at border, color bleeding, jaggy at some edges

Photoshop 7 Bicubic 5x upscale

mushy, bleeding at edges, soft jaggies

Photoshop 7 Bilinear 5x upscale

soft, fuzzy at edges

Photoshop 7 Nearest Nbor. 5x upscale

pronounced jaggies, faultless border

PhotoPaint 10 Aliased 5x upscale

soft bleed at edges, fuzzy also

PaintShop Pro 7 Bicubic 5x upscale

bleeding and jaggies at edges, soft

PaintShop Pro 7 Bilinear 5x upscale

very jaggy, soft edges but sharp border

PaintShop Pro 7 Smart 5x upscale

heavy color bleed at edges and border, soft

PaintShop Pro 7 PixelResize 5x up

very jaggy edges but sharp border

 Table 4 - Commentary on Downscale Image Fidelity

Scaling Method

Comments

Original Upsize image

Cutout of skaters in pebbly red to green gradient

Genuine Fractals 7x downscale

less jaggies, softening of pebbles

VFZoom 5x downscale

more jaggies on the edges, checker pebbles

Photoshop 7 Bicubic 5x downscale

more jaggies on the edges, fair pebbles

Photoshop 7 Bilinear 5x downscale

less jaggies, fair pebbles

Photoshop 7 Nearest Nbor. 5x down

less jaggies, mushy pebbles

PhotoPaint 10 Aliased 5x downscale

jaggies on edges, mushy pebbles

PaintShop Pro 7 Bicubic 5x downscale

jaggies on edges, mushy pebbles 2

PaintShop Pro 7 Bilinear 5x downscale

less jaggies, fair pebbles

PaintShop Pro 7 Smart 5x downscale

jaggies on edges, checker pebbles

PaintShop Pro 7 PixelResize 5x down

less jaggies. smooth pebbles

the Genuine Fractals program. This was the only method we could push confidently past the 5X limit without getting significant color bleed or edge softening. VFZoom was distinctly better than the remaining methods often pushing into the 6x to 7x range before surfacing notable distortions. But be advised, even Genuine Fractals cannot match the edge and overall image fidelity of SVG and other vector graphics. To repeat, vector graphics edges and colors are bang on 1x to 100x. But of course, with the digital camera revolution more images are coming to graphic artists in bitmap form.

For third place on upsize image fidelity, the choice is difficult but as in upsizing color fidelity the Bicubic methods appear to be better. In contrast, on downsizing image fidelity was disappointing for all the methods even at 1/5x. Every method tended to introduce softness and or jaggies at the edges. Pixel checkering was another phenomenon. For top 3 methods its your choice among the Bilinear methods, VFZoom and Genuine Fractals. In fact for both upsizing and downsizing, it is distinctly the user choice on image sharpness and fidelity. Some designers can tolerate more sedge softness but no color bleed. Others cannot tolerate jaggies. However, testing showed the upper bound to be about 1/2.5x when some of these distortions start to appear notably. In general, it is safer to upsize than downsize an image.

Conclusion

Since upsizing up to 5x is possible with most good bitmap editors (and can be pushed beyond that with the Genuine Fractals $200 or VFZoom $120 Adobe plugins), digital camera buyers looking for a bargain are advised to layoff the expensive multi-megapixel units. Instead concentrate on a digital camera in the 1-2Mpixel range with superior optics, zoom range, color fidelity and overall usability. Likewise, graphics and design shops with thousands of images and dozens of users should consider Genuine Fractals and VFZoom server versions, which can store medium scale starter images and then confidently upscale by 5-7 times (or more in the case of Genuine Fractals) and downsize by up to 40%. If lots of downsizing is required, store second starter downscale or thumbnail images. But the savings in storage space and processing time can be considerable. Remember, the color fidelity on upsizing is excellent. So designers can do most of their color and filter edits at low resolution and have confidence the colors will be maintained on the upsizing step. In fact, the only editing that should be done at full resolution would be borders and straight lines. Doing most of the edits on an image 1/25 the final filesize (assuming an average 5 times enlargement) can save a lot of processing and wait time. So software re-sizing, especially upsizing, may just turn the graphics world upsize down.

Jacques Surveyer is a consultant and photographer, he can be reached at jbsurv@thePhotoFinishes.com


Figure 1 - Top Image Upsizing Methods

 




References:

Compression Algorithms, A.Hassan, Addison Wesley
Adobe Photoshop 7 - $900, www.adobe.com
Celartem VFZoom 2.2 - $120 Adobe plugin, $2980 for 50-user Server, www.celartem.com
Corel PhotoPaint 10 - $700, www.corel.com
Jasc PaintShop Pro 7 - $100, www.jasc.com
LizardTech Genuine Fractals 2.5 - $200, www.lizardtech.com