Lifting-based Wavelet Transform with Directionally Spatial Prediction

This paper incorporates the directionally spatial prediction (DSP) into the conventional lifting-based wavelet transforms and proposes a novel, efficient and flexible lifting structure (referred as DSP-lifting hereafter). Each lifting stage can be performed at the direction, where pixels have a strong correlation, rather than always at the horizontal or vertical direction. The proposed DSP enables the predicting and updating process at the fractional pixel without any constraint on the interpolation, which fully exploit the spatial correlation among pixels. Furthermore in the 2D transform, the direction of the first 1D transform is no longer requested to be vertical to that of the second so that the second 1D transform can utilize the spatial correlation better. The proposed DSP-lifting guarantees the perfect reconstruction. We also investigate the techniques to efficiently estimate and code the directional data, thus increasing the precision of spatial prediction and reducing the overhead bits. Experimental results have shown that the proposed DSP-lifting scheme can significantly outperform JPEG 2000 in terms of PSNR (up to 2dB) and visual quality.