Abstract

In a digital image processing system, a CPU and a memory is provided to an image signal processing subsystem for computing the output vector of an inverse discrete cosine transform. The inverse discrete cosine transform is represented as a linear system and the output vector is computed using a forward mapping procedure where system matrix columns scaled by the non-zero quantized corresponding transform domain coefficient selected from the input vector are successively accumulated into the output vector. Dequantizations and scalings are performed as a combined single step by looking up the kernel values of the scaled reconstruction kernels from lookup tables corresponding to the selected transform domain coefficients' positions in the input vector. The lookup tables are highly optimized exploiting the symmetry characteristics of the reconstruction kernels, the inherent properties of quantization and the statistical attributes of the quantized transform domain coefficients. Furthermore, the calculations of the table values are also optimized. As a result, significant improvement in IDCT efficiency is achieved for the digital image processing system.