Sensitivity analysis of low-complexity vector quantizers for focal-plane image compression

Most high-performance block-coding systems for image compression, such as JPEG, have been designed for software or dedicated digital hardware implementations where the data are already assumed to be available in digital format. In modern CMOS photosensors, smart-pixel technologies have allowed the realization of basic signal processing tasks at the pixel level, in analog format before analog-to-digital (A/D) conversion. The elimination of A/D converters and implementation of block-coding directly over analog blocks of pixels in such sensors can be attractive both in terms of area savings and power consumption. The design of block encoders, under the strong hardware constraints that derive from the A/D converter removal, has been investigated in this paper. We present a comparison of three systems in terms of rate, distortion and complexity, and present a numerical simulation analysis of their sensitivity to implementation errors. The conclusion of the analysis is that linear-transform coding vector quantizers outperform full-search vector quantizers and warping hyperbolic-tangent neural networks, in terms of performance, complexity and robustness, for a CMOS imaging sensor implementation.