Abstract
Many scientific applications require that image data be stored in floating-point format due to the large dynamic range of the data. These applications pose a problem if the data needs to be compressed since modern image compression standards, such as JPEG2000, are only defined to operate on fixed-point or integer data. This paper proposes straightforward extensions to the JPEG2000 image compression standard which allow for the efficient coding of floating-point data. These extensions maintain desirable properties of JPEG2000, such as lossless and rate distortion optimal lossy decompression from the same coded bit stream, scalable embedded bit streams, error resilience, and implementation on low-memory hardware. Although the proposed methods can be used for both lossy and lossless compression, the discussion in this paper focuses on, and the test results are limited to, the lossless case. Test results on real image data show that the proposed lossless methods have raw compression performance that is competitive with, and sometime exceeds, current state-of-the-art methods.
Highlights
Floating-point image compression has not received a lot of attention by the compression community
Lossless results are given since the correctness of the compression algorithms can be unconditionally verified, it avoids the issue of doing post compression rate distortion optimization, it makes it easy to directly compare the results with other compression algorithms, and the lossless case is the case emphasized in the literature for floating-point compression
The results for the proposed method should be considered as slightly optimistic since they do not include all of the overhead needed for packet formation in JPEG2000
Summary
Floating-point image compression has not received a lot of attention by the compression community. One could attempt to mimic the lossy/lossless decompression of the proposed methods by applying the fixed-point JPEG2000 method directly to the 4-byte integer representation (for single precision) floating-point data. The lossy partial decompression of this data would not be rate distortion optimal (and would be really bad for data with widely varying exponents) since the JPEG2000 method assumes a linear representation for its rate distortion computations, and the mapping from floating-point value to 4-byte integer is nonlinear. The coding passes, context computations, bit scanning, lifting for wavelet transform, and arithmetic encoder are all identical to those used in JPEG2000 While this fact may not be important to end users, it is very beneficial to persons or companies implementing the proposed algorithms. This is the case most emphasized in the literature on floating-point compression
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
