Lab 3 - Image compression with k-means segmentation October 28, 2020

August Hertz Bugge & Kris Back Kruse & sarphiv (redacted)

For the full rapport read report.pdf

The following is a teaser from the report.

Storing images is necessary for many AI tasks, however, data storage at large scale is costly. A widely used technique is to compress images using lossy algorithms like JPEG, but even a small improvement can save money. We compare k-means clustering as a lossy compression algorithm vs. JPEG on 1890 images using qualitatively chosen parameters for our k-means clustering. Our algorithm has a compression ratio of 35% better than JPEG. This indicates high compression ratios can be obtained if parameters are optimized for a given use case.

Storing large amounts of images requires a lot of storage space which is costly. Many algorithms exist, and is used in different situations, to compressed images. We want to compare the compression of an image, using k-means clustering and the widely used "JPEG" algorithm, in terms of space gained. Our hypothesis is that, using the k-means clustering algorithm, we will get within 20% points of the compression ratio of the "JPEG" algorithm, when compared to the equivalent bitmap file.

• Image data set: http://images.cocodataset.org/zips/val2017.zip

• Source code: https://github.com/sarphiv/dtu-intro-ai-lab-03

Processing many images requires a lot of computational power. In the future, more time should be spent error-checking, optimizing, and preparing as this could ironically lead to a loss of time.