Simultaneous Compression and Denoising of Imagery Using Non-linear Interpolative Vector Quantization

Quantization of a noisy source is a classic problem. Although, optimal solutions have been shown to exist under certain assumptions, they are usually either extremely difficult to implement or computationally very intensive. The vector quantization (VQ) based strategy we present in this paper is simple, efficient, and capable of joint compression and denoising of images corrupted by additive white Gaussian noise of fixed variance. The VQ is trained on pairs of clean original images and their respective noisy versions. This VQ is then capable of taking a noisy image as input to its encoder, compressing it, and then producing a less noisy image at the output of its decoder. Experiments performed on test images (inside and outside the training set) produced significant denoising at various bit rates. Results also indicate that our system is capable of handling a wide range of noise variances, while designed for a particular noise variance.