Visualization and Graphics ➤ Interaction ➤ Dept ICS ➤ Faculty of Science ➤ UU

//webspace.science.uu.nl/~telea001/Site/TopBar

Spline-Based Dense Medial Descriptors for Lossy Image Compression

You have a grayscale or color image and you want to encode (compress) it efficiently. How do you do that? Many so-called lossy compression methods exist, such as JPEG. Are such methods the best we can reach?

This is not the case. We propose a method, called Spline-Based Dense Medial Descriptors (SDMD) that represents such images by means of dense skeletons. We efficiently encode approximations of these skeletons using splines. Our end-to-end method allows us to represent, manipulate, simplify, and compress images at levels similar or superior to comparable state-of-the-art methods.

Pipeline

The figure below shows the pipeline of our method with the following steps:

  • compute threshold-sets, or layers, from the input image
  • select a small subset of these layers
  • compute the medial descriptors of these layers
  • fit splines to these medial descriptors; these are the SDMD representation
  • reconstruct the image from the SDMD representation

Our SDMD method inherits from the SMAT method which covers steps 1,2,3, and 5 above.

Results

We have extensively compared SDMD with JPEG, JPEG 2000, and BPG. Below we show some results:

In the above figure, (a) are original images, from the medical domain. (b) show JPEG's result. (c) show SDMD. We see that SDMD reaches very similar, and often higher, quality than JPEG.

A second figure below compares SDMD with JPEG 2000 and BPG. We see that SDMD leads to better image quality.

Super-resolution

SDMD represents images by a vector model - splines encoding the medial descriptors of all selected layers in the image. As such, it allows us to easily generate super-resolution images, i.e., increasing the resolution of the encoded pictures with very limited artifacts.

The image below shows this by comparing SDMD with gradient meshes. We see how SDMD generates smoother super-resolution images, up to the detail level.

Implementation

SDMD is implemented in C++ using CUDA to efficiently do all image processing. The full source code and examples are available here.

References

J. Wang, J. Kosinka, A. Telea (2021) Spline-Based Dense Medial Descriptors for Lossy Image Compression. J Imaging 7(8), 153, MDPI