Lanczos3 algorithm as a way to produce better image downscaling

Image scaling options

At IDR Solutions recently, I have been looking for ways to provide higher quality image quality when down-sampling images in Java. I stumbled across Lanczos3 and this gives really good results. So the purpose of this article is to introduce the algorithm and show the better results you can obtain.

By default Java offers a choice of algorithms to use when scaling down an image. The default ImageIO library in Java uses the image scaling algorithms nearest neighbour, bilinear and bicubic to scale images (with a trade off between speed and quality).

Of these bicubic scaling (which takes into account other rows and colums) is best (and slowest!) option.  But over 200% the generated results are not good enough to maintain image clarity and essential details start to disappear. So I looked further afield and decided to use the Lanczos3 filtering algorithm. This is not available in ImageIO so I have had to implement it in our own JDeli image library to show you the results.

So let us look at some examples of all these scaling algorithms so you can see why I am such a fan of Lanczos3.

Image scaling examples

puzzle

[click on the image above to view the original image (2048×2048) ]

Let us consider the image above which is created in simple paint application on windows and we resized the image into 128×128 dimension with several algorithms

  1. java Nearest Neighbour (ImageIO)
  2. java bilinear (ImageIO)
  3. java bicubic (ImageIO)
  4. Lanczos3 (QualityThumbnail.java – examples produced using JDeli)

 

nearest-neighbour     bilinear    bicubic    lanczos

[click on any images above to view the thumbnail version (128×128) ]

With all of these algorithms, there is a trade off between time and quality. The best results (BiCubic and Lanczos3) require the slightly more processing time than bicubic but produces the best results. Only Lanczos3 retains the key details of the image as we make is smaller.

The only downside with Lanczos3 is that it is not implemented in Java as a default option. You will either need to implement it yourself (details on wikipedia) or use a third party library such as our commercial JDeli Image Library.

In JDeli it is a single line code to get downscaled thumbnail version

BufferedImage thumbnail = QualityScaler.getScaledImage(sourceImage,128,128);

What does Quality Scaler do in JDeli?

Generates high quality downsampled thumbnails and scaled images from Bufferedimage fastest convolution scaling in java.

Do you think the better results are worth the extra time?

If you’re a first-time reader, or simply want to be notified when we post new articles and updates, you can keep up to date by social media (TwitterFacebook and Google+) or the  Blog RSS.

Related Posts:

The following two tabs change content below.

suda

Java EE developer at IDRSolutions
Suda is the Senior Java EE Developer at IDR Solutions, and specialises in XFA, Fonts, True-Type Fonts, application servers and conversions. He is a keen science-fiction fan in his spare time.
suda

About suda

Suda is the Senior Java EE Developer at IDR Solutions, and specialises in XFA, Fonts, True-Type Fonts, application servers and conversions.
He is a keen science-fiction fan in his spare time.

2 thoughts on “Lanczos3 algorithm as a way to produce better image downscaling

  1. Ggg

    It is named lanczos algorithm, Please use proper spelling if you want to be a good pro selling a commercial product…
    https://en.m.wikipedia.org/wiki/Lanczos_algorithm

    • Hi Gianone
      thanks for pointing the error we corrected it

Leave a Reply

Your email address will not be published. Required fields are marked *

You may use these HTML tags and attributes:

<a href="" title=""> <abbr title=""> <acronym title=""> <b> <blockquote cite=""> <cite> <code> <del datetime=""> <em> <i> <q cite=""> <s> <strike> <strong>