DCT Based SPIHT Architecture for Hyperspectral Image Data Compression

of the Thesis DCT Based SPIHT Architecture for Hyperspectral Image Data Compression by Jieming Xu Master of Science in Electrical and Computer Engineering Northeastern University, August 2016 Prof. Vinay K. Ingle, Adviser The wavelet transformation is leading and widely used technology in transform coding. Many coding algorithms are designed for this transformation based on its unique structure, such as EZW, SPIHT and SPECK. The correlation between each subbands naturally generates a special tree structure in the whole image. With bitplane and entropy coding technique, the compression ratio can be achieved at a very high value. In this thesis, we focus on the traditional discrete cosine transform (DCT) to design our compression system. After analyzing the performance of the SPIHT algorithm, we found that, the coefficients’ arrangement in DCT still has the features similar to these in wavelet transform and these features are vital to maximize the performance of SPIHT algorithm. For realistic implementation, a large hyperspectral data cube must be tiled into small size of code cubes for compression to achieve a fast compression speed. In JPEG standard, two applicable block sizes are give, which are 8×8 and 16×16 blocks. In our system, we extend these blocks into three-dimensional cube for hyperspectral image as 8×8×8 and 16×16×16 cubes for testing. To enhance the compression performance, PCRD algorithm is also applied in our system. Because the values in spectrum direction share very similar trajectory for each pixel, the power of several continuous bands is predictable. In this way, we optimized the PCRD algorithm for our system, and the truncation points can be chosen without calculation which saves time. Three AVIRIS hyperspectral data sets are tested. For DCT based compression, each image cube was tiled into sizes of 8×8×8 and 16×16×16 cubes for transformation and compression independently. For DWT based compression in small code cube setting, each image cube was tiled into sizes of 32×32×32 and 16×16×16 with a DWT decomposition level of five and four. For DWT compression in large code cube setting, the transformation (five levels of DWT) and compression algorithm was performed on the whole image cube in a size of 448× 448× 224. Results showed viii that, the DCT based compression with 16 × 16 × 16 code cube size has the best performance for lossy hyperspectral image compression and the bitplane arrangement is more effective for SPIHT algorithm.