A Comparative Study of the Complexities of Neural Network Based Focal-Plane Image Compression Schemes

摘要

Block coding of images can be implemented efficiently by neuralnetworks. Kohonen self-organizing maps (SOMs) have been previouslyproposed for the implementation of full-search vector quantizers,and multilayer perceptrons (MLPs) have been proposed for theimplementation of linear-transform-based vector quantizers. Inthis article, we introduce a complexity function that estimateshow many CMOS transistors the proposed solutions will require, ifthey are implemented with analog hardware that is typical offocal-plane solutions for modern digital cameras. We also proposethe use of non-linear MLPs with Gaussian, hyperbolic tangent, orpolynomial Kernel functions, to implement newcomplexity-constrained vector quantizers. The complexity functionis applied to several SOMs and MLPs (linear and non-linear). Numerical simulation results show that MLPs achieve complexityreduction factors around 15 with respect to the SOMs, while notlosing more than 1.5 dB in reconstruction quality or 0.04 bpp inthe minimal bit rate.