Analog multilayer perceptron implementation of low complexity VQ for image compression

This work proposes a multilayer perception (MLP) method to perform direct mapping from an analog image block into a vector index, i.e. implement vector quantization (VQ) without the need for carrying out first an analog-to-digital (A/D) conversion of each analog sample. The design focuses on an analog hardware implementation and is based on observing that the use of a careful binary index assignment causes hyperplanes associated with the formation of the same bit at the MLP output to have similar spatial orientation. Experimental results show that, for the same rate-distortion performance, MLPs can be implemented with less than half the computational complexity of the Kohonen self-organizing map (KSOM) approaches advanced earlier.