Directionally classified subspace image vector quantization

This paper describes a new image coding scheme called directionally classified subspace vector quantization which is based on the dimensionality reduced subspace distortion measurement technique [1-4] and the classified vector quantization technique [5] for reducing the computational complexity and memory requirement of the image vector quantizer. In the new coding scheme, the 4×4 image block is classified into one of nine classes according to the directional content of the image block which is vector quantized using appropriate Hadamard transform subspace distortion measure. The classification is based on the horizontal and vertical gradients of the image block. The two gradient parameters form a 2-dimensional space which can be partitioned into 9 regions and each region correspond to a class of vectors. As the subspace vector quantization is applied on the restricted class of vector, extremely low dimensionality subspace distortion measures can be used. Thus, the computational complexity and memory requirement of the coder are both significantly reduced, while the reconstructed image quality is preserved for edges.