engleski

Adaptation methods of 2-D non-separable wavelet filter bank

We present several adaptation methods of 2-D nonseparable wavelet filter bank. Proposed filter bank is obtained using lifting scheme in which nonseparable quincunx decimation and true 2-D filters have been applied. Each filter bank is split into a fixed and a variable part. Due to the fixed part, basic convergence and regularity properties of the limit wavelet functions and scales are preserved. Parameters of the variable part of the filter bank are changed at each pixel in order to adapt to the analyzed image. Adaptation criterion is calculated from the wavelet coefficients. We applied one-dimensional and true two-dimensional adaptation algorithms. 1-D recursive least squares (RLS) and windowed least squares (LSW) are compared to 2-D windowed least squares and to 2-D best linear unbiased estimate (BLUE) algorithm. Adaptation methods were tested on synthetic and real-world images. It was shown that true 2-D algorithms outperform 1-D algorithms and adaptive filter banks outperform fixed (nonadaptive) wavelet filter banks.

2D wavelet filter bank; non-separable wavelets; adaptive zero moments

nije evidentirano