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 neural networks. Kohonen self-organizing maps (SOMs) have been previously proposed for the implementation of full-search vector quantizers, and multilayer perceptrons (MLPs) have been proposed for the implementation of linear-transform-based vector quantizers. In this article, we introduce a complexity function that estimates how many CMOS transistors the proposed solutions will require, if they are implemented with analog hardware that is typical of focal-plane solutions for modern digital cameras. We also propose the use of non-linear MLPs with Gaussian, hyperbolic tangent, or polynomial Kernel functions, to implement new complexity-constrained vector quantizers. The complexity function is applied to several SOMs and MLPs (linear and non-linear). Numerical simulation results show that MLPs achieve complexity reduction factors around 15 with respect to the SOMs, while not losing more than 1.5dB in reconstruction quality or 0.04bpp in the minimal bit rate.