Whilst cairo is known as a VectorGraphics-Library - it supports fast Blending-Operators
on its Pixel-Surfaces (aka ImageSurfaces) - and most wrappers around this library will
offer (in addition to cairos blending-ops) a few additional Pixel-Massaging-Methods.
The vbRichClient has 6 of that category already implemented:
- cCairo.PreMultiplyAlpha
- cCairo.DeMultiplyAlpha
- cCairoSurface.FastBlur
- cCairoSurface.GaussianBlur
- cCairoSurface.Sharpen
- cCairoSurface.AdjustColors (Brightness, Contrast, Gamma - all in one routine)
Now, there's certainly much more pixel-algorithms out there in the wild -
but the above few are what I restricted myself to, since they are often needed.
To accomodate those who want to perform their own PixelOps on those Surfaces, I've choosen
to make pixel-access as easy as possible - and so there exist two Array-Binding-Methods:
- cCairoSurface.BindArray(ByteArray) ... and its companion .ReleaseArray ByteArray
- cCairoSurface.BindArrayLong(LongArray) ... and its companion .ReleaseArrayLong LongArray
This is done without copying any Data - I guess the underlying SafeArray-technique is quite
wellknown in the meantime - and the Arrays in question need to be dynamic and uninitialized ones.
Well, that said - let's look at what the Demo has to offer:
Edge-Detection is a wide field - and I decided not to implement "the whole bunch of algos for comparison",
instead I concentrated on a very simple one (the "Roberts-Cross") - with focus on performance.
Since the algorithm is so simple (only using a small 2x2 Kernel), it was quite easy to avoid an
"inner Kernel-Loop" - and since the Formulas were not all that complex in this case too, coming
up with some Lookup-Tables for them was also relative easy to accomplish.
I've tuned it a bit though:
Instead of a SquareRoot on the "squared differences", I found that a logarithmic scaling produced
somewhat nicer results - and since the kernel is so small (compared to Sobel, Prewitt & Co.), the
algo captures fine details quite well, but tends to be "noisy" - so, a slight (and fast) PreBlurring was
added into the processing-queue, to remedy that.
For comparison (direct links instead of the forums img-tag, to get "the real image" -
and to easier compare - after loading them all - by just clicking between Browser-Tabs):
https://softwarebydefault.files.word..._grayscale.jpg
https://softwarebydefault.files.word..._grayscale.jpg
http://vbRichClient.com/Downloads/Lo...ithPreBlur.png
The last link of the above three, was generated with the Demo.
Performancewise, the Algo (including the steps: PreBlur, GreyScaling, LogarithmicRoberts, GammaCorrection, Sharpen)
takes in Sum only about 20msec total when native compiled (performing all steps on a 800x600 TrueColor-Image).
The Demo includes a few Demo-Images - here some ScreenShots:
Monarch-Original-Image:
![]()
Monarch-EdgeDetected and in ColorMode (algo performed on each color-channel separately).
The GreyScale-Result-Link I've posted already further above...
![]()
And another Result of one of the Demo-Images in Color-Mode:
![]()
Here's the Source-Zip:
EdgeDetection.zip
Have fun,
Olaf
on its Pixel-Surfaces (aka ImageSurfaces) - and most wrappers around this library will
offer (in addition to cairos blending-ops) a few additional Pixel-Massaging-Methods.
The vbRichClient has 6 of that category already implemented:
- cCairo.PreMultiplyAlpha
- cCairo.DeMultiplyAlpha
- cCairoSurface.FastBlur
- cCairoSurface.GaussianBlur
- cCairoSurface.Sharpen
- cCairoSurface.AdjustColors (Brightness, Contrast, Gamma - all in one routine)
Now, there's certainly much more pixel-algorithms out there in the wild -
but the above few are what I restricted myself to, since they are often needed.
To accomodate those who want to perform their own PixelOps on those Surfaces, I've choosen
to make pixel-access as easy as possible - and so there exist two Array-Binding-Methods:
- cCairoSurface.BindArray(ByteArray) ... and its companion .ReleaseArray ByteArray
- cCairoSurface.BindArrayLong(LongArray) ... and its companion .ReleaseArrayLong LongArray
This is done without copying any Data - I guess the underlying SafeArray-technique is quite
wellknown in the meantime - and the Arrays in question need to be dynamic and uninitialized ones.
Well, that said - let's look at what the Demo has to offer:
Edge-Detection is a wide field - and I decided not to implement "the whole bunch of algos for comparison",
instead I concentrated on a very simple one (the "Roberts-Cross") - with focus on performance.
Since the algorithm is so simple (only using a small 2x2 Kernel), it was quite easy to avoid an
"inner Kernel-Loop" - and since the Formulas were not all that complex in this case too, coming
up with some Lookup-Tables for them was also relative easy to accomplish.
I've tuned it a bit though:
Instead of a SquareRoot on the "squared differences", I found that a logarithmic scaling produced
somewhat nicer results - and since the kernel is so small (compared to Sobel, Prewitt & Co.), the
algo captures fine details quite well, but tends to be "noisy" - so, a slight (and fast) PreBlurring was
added into the processing-queue, to remedy that.
For comparison (direct links instead of the forums img-tag, to get "the real image" -
and to easier compare - after loading them all - by just clicking between Browser-Tabs):
https://softwarebydefault.files.word..._grayscale.jpg
https://softwarebydefault.files.word..._grayscale.jpg
http://vbRichClient.com/Downloads/Lo...ithPreBlur.png
The last link of the above three, was generated with the Demo.
Performancewise, the Algo (including the steps: PreBlur, GreyScaling, LogarithmicRoberts, GammaCorrection, Sharpen)
takes in Sum only about 20msec total when native compiled (performing all steps on a 800x600 TrueColor-Image).
The Demo includes a few Demo-Images - here some ScreenShots:
Monarch-Original-Image:
Monarch-EdgeDetected and in ColorMode (algo performed on each color-channel separately).
The GreyScale-Result-Link I've posted already further above...
And another Result of one of the Demo-Images in Color-Mode:
Here's the Source-Zip:
EdgeDetection.zip
Have fun,
Olaf