Codebook Calibration Method for Vector Quantizers Implemented at the Focal Plane of CMOS Imagers

Abstract

We have recently designed and tested a CMOS imager fabricated in a 0.35- $\mu \text{m}$ CMOS technology, with focal-plane image compression using vector quantization (VQ) for texture representation. In Lloyd-Max VQ design theory, the input space partition is matched to the codebook and vice versa, by means of well-known partition and centroid conditions. Although both the partition and codebook were designed with vectors extracted from the same digital image base, only the partition is nonideally implemented, owing to mismatches and errors produced by the fabrication process. Whereas this paper does not have a focus on hardware or system architecture, some hardware details are described, to illustrate the most important error sources. In this paper, we introduce a method for updating the codebook as binary indices are experimentally generated by the camera, so that texture reconstruction with smaller mean squared error is consistently obtained. We illustrate the method by means of experiments carried out with captured and improved photographs.