UNIVERSAL IMAGE CODING USING MULTISCALE RECURRENT PATTERNS AND PREDICTION Nuno M. M. Rodrigues 1,2 , Eduardo A. B. da Silva 3 , Murilo B. de Carvalho 4 , S´ ergio M. M. de Faria 1,2 , V´ ıtor M. M. da Silva 1,5 1 Instituto de Telecomunicac ‚ oes, Portugal; 2 ESTG, Instituto Polit· ecnico Leiria, Portugal; 3 PEE/COPPE/DEL/Poli, Univ. Fed. Rio de Janeiro, Brazil; 4 TET/CTC, Univ. Fed. Fluminense, Brazil; 5 DEEC, Universidade de Coimbra, Portugal. e-mails: nuno.rodrigues@co.it.pt, eduardo@lps.ufrj.br, murilo@telecom.uff.br, sergio.faria@co.it.pt, vitor.silva@co.it.pt ABSTRACT In this paper we present a new method for image coding that is able to achieve good results over a wide range of image types. This work is based on the Multidimensional Multiscale Parser (MMP) algorithm [1], allied with an in- tra frame image predictive coding scheme. MMP has been shown to have, for a large class of image data, including texts, graphics, mixed images and textures, a compression efficiency comparable (and, in several cases, well above) to the one of state-of-the-art encoders. However, for smooth grayscale images, its performance lags behind the one of wavelet-based encoders, as JPEG2000. In this paper we propose a novel encoder using MMP with intra predictive coding, similar to the one used in the H.264/AVC video coding standard. Experimental results show that this method closes the performance gap to JPEG- 2000 for smooth images, with PSNR gains of up to 1.5dB. Yet, it maintains the excellent performance level of the MMP for other types of image data, as text, graphics and com- pound images, lending it a useful universal character. 1. INTRODUCTION Despite the well known intrinsic limitations of the trans- formed-based image encoding methods, this class of algo- rithms is generally considered as state-of-the-art, both in image and video compression. These methods assume that an image has a low-pass nature and expect most of the trans- form coefficients for the higher frequencies to be negligi- ble or of little importance. This is then exploited by using coarse quantization or by simply ignoring these frequency components. When used to encode images with frequency distributions other then low-pass, like text and graphics, the efficiency of these methods deteriorates noticeably. The method presented in this paper is built upon an algo- rithm that is not based on the transformation-quantization- encoding paradigm. It is referred to as Multidimensional Multiscale Parser (MMP) [1], because it uses an adaptive dictionary of vectors to approximate variable-length input vectors. These vectors result from recursively parsing an original input block of the image. Scaling transformations are used to resize each dictionary element to the dimension of the block segment that is being considered. In this work we introduce a new development of MMP, called MMP-Intra, that combines MMP with predictive co- ding techniques, like those used in H.264/AVC Intra coding prediction. Experimental results show that this new method is able to achieve relevant coding gains over the original MMP, making it competitive with state-of-the-art encoders, like JPEG2000 [2] and H.264/AVC [3]. When text, graphics and texture images are considered, then MMP-Intra achieves significant gains over the other tested encoders, which indi- cates that it has a universal character. Unlike previous extensions of the MMP [4], MMP-Intra does not make any assumptions about the nature of the in- put image, achieving good coding results both for smooth grayscale images as well as for text, graphics and combined text and grayscale images. In the next section the MMP algorithm for image coding is briefly described. Section 3 presents the MMP-Intra me- thod, discussing the joint use of Intra-like prediction sche- mes and MMP. In section 4 some experimental results are shown and section 5 presents the conclusions. 2. THE MMP ALGORITHM Although the MMP algorithm was initially proposed as a generic lossy data compression method, it is easily extenda- ble to work with n-dimensional data, and has been success- fully applied to image coding. In this section we describe the most important aspects of the MMP algorithm applied to image coding. An exhaustive description of the method can be found in [1]. MMP is based on approximations of data segments (in 0-7803-9134-9/05/$20.00 ©2005 IEEE II-245