Study On High-ratio Remotely Sensed Image Compression Using Structural Characteristics Of Wavelet Coefficients

EZW (Embedded Zerotree Wavelets) coding algorithm has not only high compression rate but also some new features such as progressive coding/decoding, low computational complexity, etc. In recent years, many improved algorithms have been developed. SPIHT (Set Partitioning in Hierarchical Trees) is an outstanding one among them. However, remotely sensed image compression has its own characteristics, such as weak data correlation, complex texture, big data volume and high real-time request. Direct usage of standard SPIHT cannot provide satisfactory reconstructed image quality or executing performance in this type of applications. Although the zerotree description of wavelet coefficients proposed by J. M. Shapiro can remarkably removes the statistical redundancy of insignificant coefficients, statistical experiments show that some directional and positional dependencies still exist between significant coefficients in each subband. The new points presented in this paper are that based on this structural characteristic, the RLC is for the first time incorporated into SPIHT and a new 2×2 block-based "multiresolution"scanning scheme is also proposed to replace the widely used row-by-row scan path, which clusters significant coefficients'1-D distribution and , in turn, reduces coefficients'such structural redundancy to some extent. The compression efficiency is further increased in applications of remotely sensed image coding at high ratios. As to the performance assessment for our proposed improvements, we employed both the traditional PSNR metric for image fidelity and a new perceptual image quality measurement based on images'structural similarity. Theoretic analysis and experiments indicate: (1)Based on significant wavelet coefficients'structural characteristics, after introducing RLC into SPIHT, not only is the low computational complexity preserved, but also increases the PSNR up to 1 dB at very high compression ratios, and the average improvement is about 0.3dB at 0.1-0.5 bpp for standard test images used. The visual quality of reconstructed images is also significantly improved. (2)Based on the arithmetic coding algorithm, we redesigned codec's configuration according to the general correlation of LIS and LIP nodes'significance in SPIHT, which was to further exploit the nonlinear dependencies of image's wavelet coefficients. However, this modification is not efficient. The general performance is only equivalent to the original one.