04.10.

2017 Image Kernels expla ned v sually

Back

Image Kernels
Expla ned V sually
Tweet Beğen 762 Paylaş

By Victor Powell

An image kernel is a small matrix used to apply effects like the ones you might nd in Photoshop or Gimp, such as blurring,
sharpening, outlining or embossing. They're also used in machine learning for 'feature extraction', a technique for determining the
most important portions of an image. In this context the process is referred to more generally as "convolution" (see: convolutional
neural networks.)

To see how they work, let's start by inspecting a black and white image. The matrix on the left contains numbers, between 0 and 255,
which each correspond to the brightness of one pixel in a picture of a face. The large, granulated picture has been blown up to make
it easier to see; the last image is the "real" size.

Let's walk through applying the following 3x3 sharpen kernel to the image of a face from above.

sharpen

0 -1 0

( -1

0
5

-1
-1

0
)
Below, for each 3x3 block of pixels in the image on the left, we multiply each pixel by the corresponding entry of the kernel and then
take the sum. That sum becomes a new pixel in the image on the right. Hover over a pixel on either image to see how its value is
computed.

( 139 + 192 + 190

http://setosa. o/ev/ mage-kernels/ 1/5

04.10.2017
( Image Kernels expla ned v sually
× 0 × -1 × 0

+ 139 + 191 + 197

× -1 × 5 × -1

+ 149

× 0
+ 191

× -1
+ 190

× 0
)
= 236

kernel:
sharpen

input image output image

One subtlety of this process is what to do along the edges of the image. For example, the top left corner of the input image only has
three neighbors. One way to x this is to extend the edge values out by one in the original image while keeping our new image the
same size. In this demo, we've instead ignored those values by making them black.

Here's a playground were you can select different kernel matrices and see how they effect the original image or build your own
kernel. You can also upload your own image or use live video if your browser supports it.

Dosya Seç Dosya seç lmed L ve v deo

0 -1 0

-1 5 -1

0 -1 0

sharpen

The sharpen kernel emphasizes differences in adjacent pixel values. This makes the image look more vivid.

For more, have a look at Gimp's excellent documentation on using Image kernel's. You can also apply your own custom lters in
Photoshop by going to Filter -> Other -> Custom...

For more explanations, visit the Explained Visually project homepage.

Or subscribe to our mailing list.

Ema l address Subscribe

28 Comments Explained Visually 

1 Login

Sort by Best
 Recommend 29 ⤤ Share

http://setosa. o/ev/ mage-kernels/ 2/5

04.10.2017 Image Kernels expla ned v sually

Join the discussion…

LOG IN WITH
OR SIGN UP WITH DISQUS ?

Name

Keeguon • 3 years ago

Just as a side note, though image kernels are a good way to produce a blurring effect on pictures it's not really the most efficient way to do it
on large kernels. If you want to achieve more complex blurs efficiently it's more efficient to use a Fourier transform.
4△ ▽ • Reply • Share ›

Zz Tux > Keeguon • a year ago

I prefer using firwin. Using one or more FIR filter works better than ifft.
After all a kernel is just a filter :D.
26 △ ▽ • Reply • Share ›

tar van krieken > Keeguon • 2 years ago

well, simply running the same kernel over the output picture multiple times gives the desired effect aswell and is slightly more efficient
than large kernels :)
△ ▽ • Reply • Share ›

renan jegouzo > tar van krieken • a year ago

for the blur it's faster to do it with 2 passes, one horizontally, one vertically
△ ▽ • Reply • Share ›

Adam Hancock > renan jegouzo • 6 months ago

And it's faster still to use FFT to perform the convolution of the blur kernel in the frequency domain
△ ▽ • Reply • Share ›

Christopher Silvia • 3 years ago

So how do Convolutional Neural Nets use these Kernels to detect features?
3△ ▽ • Reply • Share ›

pwais > Christopher Silvia • 3 years ago

CNNs will learn a weighted combination of these kernels in order to match parts of objects (e.g. part of the wheel on a car). This is a
nice visualization! Other helpful resources:
* See DeepViz: https://github.com/bruckner...
* See Layers >=2 plotted here: http://arxiv.org/pdf/1311.2...
* Another tool one can use to play with convolutional kernels is ShaderToy ( https://www.shadertoy.com/ )
5△ ▽ • Reply • Share ›

Rafael Espericueta > Christopher Silvia • a year ago

Filters can be designed to find edges, which are places in an image with maximum information, as an edge implies a spatial change.
Areas of an image with the same color contain very little information. So filters are used in image processing (even by our own
retinas!) to extract information from images.
△ ▽ • Reply • Share ›

Lee Meng • 3 days ago

Wonderful explanation!
△ ▽ • Reply • Share ›

Amadeo • 5 months ago

Deep learning aside, how do you derive kernels like blur, sobel, etc?
△ ▽ • Reply • Share ›

grubberr • a year ago

wow how simple it can be. Thanks a lot for explanation !!!
△ ▽ • Reply • Share ›

jagadkanihal • a year ago

Awesome post, very intuative
http://setosa. o/ev/ mage-kernels/ 3/5
04.10.2017 Image Kernels expla ned v sually
△ ▽ • Reply • Share ›

Orion Osborn • a year ago

How are the negative numbers translated to the greyscale?
△ ▽ • Reply • Share ›

Adam Hancock > Orion Osborn • 6 months ago

by scaling the result or anything less than 0 is 0 and anything greater than 255 is 255
△ ▽ • Reply • Share ›

Luis Javier Merino > Adam Hancock • 17 days ago

That is normally known as clamping, or saturating arithmethic.
△ ▽ • Reply • Share ›

Edgar Cheung • a year ago

I am just learning about image kernel recently, and I would like to ask whether the kernels like the sobel kernels has been flipped before
convolution? Or the kernels shown above have already been flipped?
△ ▽ • Reply • Share ›

Pedro Tytgat • a year ago

Very nice page. Thank you!
△ ▽ • Reply • Share ›

Prajwel P.J • a year ago

Beautiful explanation!
△ ▽ • Reply • Share ›

TechnoWings Project • a year ago

Thank you for a wonderful article! You have taken lot of effort to explain convolution. Thank you so much!!!
△ ▽ • Reply • Share ›

Harry Santoso • a year ago

so if we want sharpening image with convolution we must make it into grayscale right?
△ ▽ • Reply • Share ›

Juraj > Harry Santoso • 10 months ago

No, the effect is best visible in grayscale. You can sharpen color images with this method as well..
△ ▽ • Reply • Share ›

CommanderWaffles • a year ago

Excellent demo :)
△ ▽ • Reply • Share ›

Let Me • 2 years ago

Good explanation of kernel effect. Could not thank you enough. I was really confuse as to how in Opengl is doing this by visualizing.
△ ▽ • Reply • Share ›

Abhijit • 3 years ago

nice work man !!!
△ ▽ • Reply • Share ›

Muggin • a year ago

Hey! First of all, thanks for the great post! I would just like suggest one adjustment. I believe this part "In this context the process is referred to
more generally as "convolution"" is not entirely correct. In ConvNets a kernel is still called a kernel, or a filter, and convolution is the process
of applying the kernel to the processed image.
△ ▽ • Reply • Share ›

Anton Bacaj > Muggin • a year ago

The author said that the process is referred to as "convolution", I'm pretty sure you misunderstood what the author wrote. He did not
say that kernels are called convolutions, but the act of applying the kernel in a CNN is referred to as the "convolution" part of the "C" in
"CNN". Your suggestion needs to be re-evaluated.
△ ▽ • Reply • Share ›
http://setosa. o/ev/ mage-kernels/ 4/5
04.10.2017 Image Kernels expla ned v sually

Archi • 2 years ago

Is Kernel matrix decided?? Or we can Choose manually kernel matrix?
△ ▽ • Reply • Share ›

Juraj > Archi • 10 months ago

Both yes and no. Depends what you want to do. With sharpening, sum in matrix should be 1. So in the example above you have (4 x
-1) + 5 = 1. If you put -2s there, you have to change 5 for 9. Effect is much stronger now, image is overshrpened. Change again, and
make negative numers -0.3, so central one should be 2.2 now. Sharpen is much weaker. You can change intensity, but some rules
must be kept.
1△ ▽ • Reply • Share ›

✉ Subscribe d Add Disqus to your siteAdd DisqusAdd 🔒 Privacy

http://setosa. o/ev/ mage-kernels/ 5/5