Removal of tile artifacts using projection onto scaling functions for JPEG2000

images divided into independent tiles in the pixel domain is important. However, tiling introduces artifacts when transform coefficients are quantized. This paper presents a " detiling " solution applicable to wavelet transform coefficients. The solution, called Projection Onto Scaling Functions, eliminates the blocking artifacts by computing approximations that match the smoothness of the compressed image inside of a tile. This solution is applicable to any wavelet filter – extending previous work that applied only to a limited class of wavelet filters – and specifically targets use in decoding of JPEG2000 compressed images. 1 INTRODUCTION Image compression systems used in real world applications have to incorporate memory and bandwidth constraints into their architecture. A typical approach to reduce memory cost and/or to allow parallel processing is to divide images into tiles and compress each tile independently. Such a tiling is standard in the DCT-based JPEG compression system where tiles consists of blocks of 8x8 pixels. Naive decoding for independent tiles can create visible tile boundary artifacts. For DCT-based compression they are commonly called block-artifacts. Significant work has been done in block-artifact removal in DCT-compressed images. Approaches include simply smoothing the decompressed image at the boundaries with a lowpass filter, computing a polynomial fit through the given coefficients (Section K.8 of [1]), Projection Onto Convex Sets [2] and denoising via wavelet shrinkage [3]. Those methods rely on assuming an image model for describing smoothness across block boundaries. Compression artifacts in wavelet-based compressed images have very different characteristics than DCT artifacts. Without any tiling, in so-called full-frame compression, a DCT system would produce severe Gibbs artifacts over the entire image. Wavelet-based systems would produce some ringing around edges and some " wavelet noise " in smooth regions by reconstruction of isolated wavelet basis vectors. Both artifacts are determined by the smoothness of the wavelet system. The Gibbs artifacts in DCT compression are reduced by choosing very small tiles. As a consequence, tile artifacts overcome the Gibbs artifacts. Wavelet-based systems allow for much larger tiles, and both types of artifacts, full-frame and tile boundary artifacts are clearly visible. In this paper a detiling solution for wavelet-based compressed images is proposed with the goal to modify an image at the tile boundaries such that the image quality near those boundaries