<title>Improved compression performance using singular value decomposition (SVD)-based filters for still images</title>

Abstract

It is well known that random noise on images significantly affects the efficiency of compression algorithms. Traditional spectral filtering techniques are effective in many cases but may require some prior knowledge of the noise and image characteristics. Furthermore, the processing requirements of spectral filters strongly depend on their noise rejection properties. In this paper we present a block-based, non-linear, filtering technique based on the Singular Value Decomposition (SVD). Traditional applications of SVD to image processing rely on heuristics to estimate the noise power and are usually applied to the entire image. The proposed scheme employs a complexity-theoretical criterion for noise estimation which exploits the well known property that random noise is hard to compare. By combining SVD with a lossless compression algorithm, in our case lossless JPEG, we can estimate the noise power and derive accurate SVD thresholds for noise removal. Simulation results on grayscale images contaminated by additive noise show that the technique can effectively filter noisy images and improve compression performance with no prior knowledge of either the image or the noise characteristics. Furthermore, the technique does not cause any blurring, unlike linear filtering techniques or median filtering.