Application of the Princen-Bradley filter bank to speech and image compression

The authors provide an efficient fast Fourier transform implementation, a set of optimal windows, and a new proof of the perfect reconstruction property for a critically sampled filter bank introduced by Princen and Bradley. Based on the new proof, necessary conditions for perfect reconstruction are also derived for the oversampled case. This filter bank is well suited to compression of speech and images due to its good frequency resolution, overlapped output, computational efficiency, and low delay. The authors evaluated its performance in a 32 band, 16-kb/s adaptive speech compression system, and found both its objective and subjective performance to be comparable to a more complex quadrature mirror filter bank. In image coding experiments, the Princen-Bradley filter bank demonstrated quality similar to DCT-based systems at high rates and reduced blocking artifacts at rates of 0.25-0.5 bpp. A psychophysically based bit allocation algorithm that provides a significant perceptual improvement over a related algorithm that seeks to minimize the mean-square error is introduced.<
>