Master Thesis of Nico Schertler
Improving JPEG Compression with Regression Tree Fields
Chair of Computer Graphics and Visualization
Student: Nico Schertler
Supervisor: Prof. Dr. Stefan Gumhold
Abstract
Storing digital images is a common task in a variety of domains, ranging from family photo albums to medical imaging. Where many images need to be saved, efficient coding is essential. JPEG, as a well-established codec for still-image compression, is used in a variety of use cases. This thesis analyzes improvements on compression performance by introducing a prediction step through regression tree fields. RTFs are used to predict coefficients of JPEG's discrete cosine transform, which allows to encode only differences of small magnitude instead of the original coefficients. The adapted compression scheme's degrees of freedom (e.g. prediction order, RTF loss function, and encoding procedure) are analyzed with respect to their influence on compression performance.
Results
This thesis' results show that the general idea of predicting coefficient images can improve the compression performance significantly. Especially loss-specific optimization poses a great gain, which induced the development of an entropy-based loss function. Furthermore, the results are evident that RTF models are not suited for predictions in the frequency domain with respect to compression, which suggests further work with different models and image representations based on the developed scheme.
DCT coefficient images
Idea of entropy loss