Codebook improvements for a CMOS imager with focal-plane vector quantization

In previous works a 32 × 32 pixel CMOS image sensor featuring focal-plane image compression based on differential pulse-code modulation and vector quantization was presented. Images compressed with this imager were reconstructed at peak signal-to-noise ratios around 18 dB, for bit rates below 0.94 bit per pixel. In this work, we focus on the improvement of the vector quantization stage of the previously proposed imager. We use experimentally captured photographs to update the code-book that is used at the decoder side. With the new codebooks, we are able to reconstruct texture feature vectors with a mean squared error (MSE) that is consistently smaller than the MSE obtained with the originally designed codebook.