Abstract
We propose new image compression algorithms by predicting each block using an improved gradient estimation at the block boundary followed by applying an orthogonal transformation to the prediction error. Compared to the previously proposed polyharmonic local cosine transform where the DC components of adjacent blocks were used for the gradient estimation, our new methods directly use multiple combinations of the neighboring pixel values to estimate the gradients at block boundary more accurately. Hence, we can improve the prediction of each block using such gradient information. Consequently, we can improve image reconstruction quality and reduce the blocking artifact and aliasing noise.
