Task-specific metrics and optimized rate allocation applied to Part 2 of JPEG2000 and 3-D meteorological data

Olga M Kosheleva, University of Texas at El Paso


In this dissertation, we show how the state-of-the-art image compression techniques—JPEG2000—can be utilized for compressing multi-dimensional data; specifically, we look for compressions that minimize mean squared error (MSE) and maximum absolute error (MAE). We also show how to use JPEG2000 to guarantee the given value of the reconstruction error (i.e., how to solve the corresponding inverse optimization problem) by adjusting bit error rates. As a case study, we consider meteorological data produced by the Battlescale Forecast Model. ^ In the independent 2-D approach, the 3-D volume of data is considered as a set of layers, which are successively compressed (and stored or transmitted). Another alternative considered is the 3-D approach which consists of, first, preprocessing the data in the vertical direction by applying Karhunen-Loève Transform, and then compressing slice by slice the decorrelated data. The JPEG2000 Part 2 extension includes this 3-D approach, but does not provide a method for optimal bit rate allocation to the individual slices. One of the main contributions in this dissertation is the solution of this problem. Our first approach makes use of experimentally acquired rate-distortion data, thus providing optimal solution, but requiring a significant amount of computational effort. In the second, more computationally efficient approach, we use a mixed model approximation to the actual rate-distortion data. The model is a piece-wise combination of the traditional high-resolution model with a model that gives good approximation at low bit rates [MF98]. ^ The optimization problem is solved for both MSE and MAE. For MSE, we use an approach similar to the Post-Compression Rate-Distortion (PCRD) optimization approach used in JPEG2000 for selecting the optimal truncation points for code-blocks [TM02]. For MAE, for 3-D approach, we derive the upper bound on MAE and minimize this upper bound by using Lagrange multipliers; for 2-D approach, we provide an explicit solution to the MAE optimization problem. ^ Finally, in the Appendices, we demonstrate, on the example of mammography, FLIR, and SMD images, how various task-specific quality metrics can be used to evaluate lossy compression degradation. These metrics can be extended to 3-D meteorological data. ^

Subject Area

Engineering, Electronics and Electrical

Recommended Citation

Kosheleva, Olga M, "Task-specific metrics and optimized rate allocation applied to Part 2 of JPEG2000 and 3-D meteorological data" (2003). ETD Collection for University of Texas, El Paso. AAI3118163.