Image converted with DxO

Image converted with UFRaw

The image was Shot on a Nikon D40 at ISO 1600 so that there would be a reasonable amount of noise in the image to see how well DxO handles noise in the conversion.

I used bilinear interpolation in UFRaw so that I could see if DxO did anything with the pixels that came from the camera as well as filling in the gaps between pixels of the same colour.

Regions of interest in the image

The first region is an out of focus light grey area. Firstly the DxO Conversion, then the UFRaw conversion

Region 1 DxO

Region 1 UFRaw

Now just the luma for each conversion:

Region 1 DxO Luma

Region 1 UFRaw Luma

And the difference between the two luma conversions:

Region 1 Luma difference

Here I have enhanced the contrast and then split the image into different scales of detail. You can see there is some difference at scales larger than 1 or 2 pixels.

Region 1 Luma difference, contrast enchaced, at differnt detail scales.

And the chroma:

Region 1 DxO Chroma

Region 1 UFraw Chroma

The second area; A pile of thumb tacks.

Region 2 DxO

Region 2 UFRaw

The Luma of region 2:

Region 2 DxO Luma

Region 2 UFRaw Luma

Region 2 Chroma:

Region 2 DxO Chroma

Region 2 UFRaw Chroma

Region 3, pencils.

Region 3 DxO

Region 3 UFRaw

Region 3 Luma:

Region 3 DxO Luma

Region 3 UFRaw Luma

Region 3 Chroma:

Region 3 DxO chroma

Region 3 UFRaw chroma

Region 4, Out of focus black and white photo:

Region 4 DxO

Region 4 UFRaw

Region 4 Luma:

Region 4 DxO Luma

Region 4 UFRaw Luma

Region 4 Luma difference

Region 4 Chroma:

Region 4 DxO Chroma

Region 4 UFRaw Chroma

In this article I am going to discuss the affect that different design elements have on PNG compression. I will look at some good and bad things to do to get a small file size when creating or editing your image. I will not look into software that optimises the compression once the image is complete.
For example: Why is this image only 2816 bytes…

And this one 4924 bytes

How PNG works… for dummies

PNG looks for patterns running horizontally in an image. Patterns can be easily compressed, so the more patterns you have the more compression you’ll get. The patterns that the PNG algorithm finds are in the binary representation of the image. Sometimes it is not always obvious that there will be a pattern in the image from looking at it.

The examples are meant to illustrate principles. Your mileage may vary.

Example 1

A square that fades from black to white is on a white background. If the fade runs perfectly horizontally or vertically the file size is small. If the box with the fade is rotated, the file size becomes much larger. In the last example the fade is on an angle but the box is not.

474 bytes

427 bytes

1908 bytes

2903 bytes

Example 2

Different types of dither will have an impact on the file size.
Floyd-Steinburg 3202 bytes
Pattern 469 bytes

To be continued…

I downloaded one of the free sample pictures from
www.3D.sk: source of human photo references for 3D modeling and texturing and applied my healing technique to it. Once I had set up the layers (which took a bit of time to ‘despeckle’ on my old computer), it took less than ten minutes to get the results you can see in the picture above.

It could be a bit better but I wanted to see how quickly I could get some fairly decent results. It’s pretty quick, even if I say so myself!

This is an update to the old method I showed you.

This method is faster, less complicated, and is better at dealing with edges. You still get good results, and you can see what is happening in real time.

10 July 07 – Edited step number 3 so that the whole thing works now!

The setup

I used this image in my example (about 182kb): creation_of_adam.png

1. Duplicate the layer. Call this (top) layer ‘Shape’. Select this layer.

2. Open the Despeckle Dialog ( Filters -> Enhance -> Despeckle). Uncheck the ‘Adaptive’ and ‘Recursive’ check boxes. Set the Black Point to -1 and the White point to 256. Adjust the radius until the blemish that you want to heal disappears. In Gimp 2.2 make sure that radius is odd (there is a bug in 2.2 that is fixed in 2.3)

3. Change the layer mode to ‘Grain Extract’. Select ‘Copy Visible’ from the Edit menu. Paste this and convert it into a new layer. Name this layer ‘Texture’. Change the Mode to ‘Grain merge’. Change the ‘Shape’ Layer’s mode to ‘Normal’.

4. Show only the Texture layer. Select a hard brush and clone over the blemish. Make both the ‘Texture’ and ‘Shape’ layers visible to see the result. You can clone the texture layer while both layers are visible to see the results as you do it if you like.

5. You’re done.

Theres plenty of information around about different methods of Bayer Interpolation. Most of those examples are for images that have very little noise. I wanted to see how different methods would compare for a very noisy image. For example, one that is very underexposed, or shot at a very high ISO setting.

For this experiment I found an image I took at night where the flash on my Fuji S5600 did not fire. I had a 1/10^th second exposure so I did get something to work with. I used UFRaw 0.12, converting the image into 8-bit TIFF. I then cropped items of interest and converted to PNG. No noise-reduction was used.

I used the follow methods: Variable Number of Gradients (VNG), Patterned Pixel Grouping (PPG), Adaptive Homogeneity-Directed (AHD) and bilinear.

A Car

Click on the thumbnail for actual size (678kb)

A Face

Click on the thumbnail for actual size (447kb)

What I think

This is a quick and dirty experiment but I think VNG gives the images with the least noise.

I have revised this technique. See the latest version here.

Here is an alternative to using the healing brush in GIMP 2.3. I think it its good because:

• It achieves good results
• It will work in GIMP 2.2
• Once you have set up the layers it is very quick to do
• My old computer, a 1GHz Athlon, has no problems with showing the brush strokes in real time.
• You have a lot more control

Read on to learn all about it.

For this demo I used a photo I found on flickr:

http://flickr.com/photos/gusdiaz/29802203/

It is licensed under Creative Commons Attribution-NonCommercial-ShareAlike 2.0

Part 1 – The set up

If you can write scripts for the GIMP you can automate this process. I can’t write scripts, so if you do write one I’d love to get a copy.

1. Duplicate the layer twice

2. Name the top layer ‘16th size’

3. Name the middle layer ‘quarter size’

4. Resize ‘16th size’ layer to 6% original size with linear scaling

5. Resize ‘Quarter size’ layer to 25% original size with linear scaling

6. Rescale both ‘16th size’ and ‘quarter size’ layers back to the original image pixel dimensions using ‘Sinc’ interpolation (or Cubic if you are not using Gimp 2.2). Do not use percentages, use pixel values. Check that the both width and height values match your image size.

-At this stage you should have 3 layers of equal size, the original image, a somewhat blurry layer (called quarter size) and a very blurry layer (16th size)

7. Hide the ‘16th size’ and set the ‘quarter size’ layer mode to Grain extract. This will give you a predominantly grey image with just the fine details or texture of the original.

Click to enlarge

8. Select ‘copy visible’ from the Edit menu. Paste this and make it a New Layer – Name this layer ‘Texture’

9. Make the ‘16th size’ and ‘quarter size’ layers visible. Hide the others. Set the ‘16th size’ layer mode to ‘Grain extract’

10. Select ‘copy visible’ from the Edit menu. Paste this and make it a New Layer – Name this layer ‘Features’

11. Delete the ‘Quarter size’ layer. Rename the layer ‘16th size’ layer as ‘Shape’ and set the mode to ‘normal’ . Arrange the layers so in this order, top to bottom. Texture, Features, Shape, Background.

12. Set both ‘texture’ and ‘features’ layers to ‘Grain merge’. This will make it look like you have your original image back. Try hide different layers to see what happens.

- What you have done is split your image into three layers based on the size of the objects in the image. The ‘texture’ layer has only small sized objects in it. The ‘Shape’ Layer has only large sized objects in it. The ‘features’ layer has medium sized objects in it. Now we can edit different sized objects largely independently of the others.

Part 2 – Let the healing begin

1. Zoom in to 100%. Find the part of the image you want to edit. Show only the ‘Texture’
layer. You may need to engage the ‘Display filters’ Contrast setting to get a better idea of what you are doing

2. Select a hard edged brush. Clone over the wrinkles near the eye.

-Make all layers visible. You will see that most of the wrinkles have gone, and the skin looks realistic. The large dark wrinkle below the eye still exists, but we will remove that.

3. With all the layers visible, add a Mask to the ‘Features’ layer. Using a soft edged brush paint black onto the Mask in the position of the dark wrinkle below the eye. It should disappear. If you are keen, you can paint black onto the mask where there are any small blemishes.

Click to enlarge

If you decide that the fine texture is to great (e.g. skin pores too obvious), add a mask to the texture layer use it to vary the opacity of the layer in selective spots.

The final result (109kb):
Click to enlarge

Notes

• When you are resizing your different layers, keep in mind the size of the ‘texture’ in the picture. For large images you may need to resize the layers by 1/8^th and 1/64^th. Some trial and error may be required.
• When you are using the mask on the features layer you will find that things may start to look a bit strange near edges. Instead of using the mask to make the layer transparent, you may find you get a better results by blurring, or cloning the features layer.