Lossless region of interest coding

Abstract

This paper describes a method that uses the well known S+P and TT transforms to encode a specially important part of an image, a region of interest (ROI) in a lossless mode. Other parts of the image (the background) are given lower quality levels allowing higher compression. The ROI coding is done in the framework of an embedded wavelet-based image compression algorithm. Entire coding, regional and full image, is done in a naturally progressive manner all the way up to lossless.

Zusammenfassung

Dieser Artikel beschreibt eine Methode, die die wohlbekannten S+P und TT Transformationen benutzt, um einen besonders wichtigen Teil eines Bildes, die Region of Interest (ROI) ohne Informationsverlust zu kodieren. Andere Teile des Bildes (der Hintergrund) erhalten geringere Qualitätsstufen, die eine höhere Kompression erlauben. Die ROI-Kodierung wird im Rahmen eines eingebetteten und auf Wavelets basierenden Bildkompressionsalgorithmus durchgeführt. Die gesamte Kodierung (der ROI und des Gesamtbildes) wird auf eine natürliche, bis hin zu einer informationsverlustfreien, fortschreitenden Art und Weise durchgeführt.

Résumé

Cet article décrit une méthode utilisant les transformations S+P et TT bien connues pour encoder une partie spécialement importante d'une image, une région d'intérêt (ROI), en mode sans perte. D'autres parties de l'image (l'arrière-plan) se voient donner des niveaux de qualité moins élevés ce qui permet une compression accrue. Le codage de la ROI est opéré dans le cadre d'un algorithme de compression d'images basé sur les ondelettes. Tout le codage, aussi bien local que sur l'image complète, est réalisé de façon naturellement progressive jusqu'au codage sans perte.

Introduction

When handling digital images the main concern should of course always be the quality of the image delivered to the user. However, quality should be achieved with as compact representation of the image as possible. The research efforts in the field of digital image compression have been immense for many years. One possible classification of the techniques is into two categories, lossy and lossless algorithms.

The more common lossy algorithms take advantage of the fact that in most cases small losses can be tolerated. This gives more freedom on the choice of the algorithm. Higher compression efficiency can also be achieved if one is willing to sacrifice quality.

Lossless compression is used for high demanding applications where no loss of quality can be tolerated. However, this requirement will make the compression performance fairly modest.

Recently lossy and lossless algorithms have partially merged. This is due to new schemes that combine the two without any substantial loss of performance for either mode [11], [14]. This is done inside the framework of the recently very popular schemes which produce an embedded bitstream [8], [9], [11], [13], [15]. An embedded bitstream has the property that the information is encoded in the order of its importance. This means that with one single bitstream the encoder performs its best independently of where the bitstream is truncated. It is therefore possible to hit a specific bitrate or to easily provide different users with different quality images depending on their demands and access link. The embedded bitstream will also make the transmission progressive so that the receiver can use the transmitted information to display an image of gradually improving quality while receiving. It should be pointed out however that which information is the most important is not always an easy choice and sometimes a highly subjective one. For example in some applications or for some users different regions of an image might have different importance. Therefore it might be desirable to have a representation which is not only compact but also selective [16].

This paper presents an image compression algorithm that achieves progressive encoding and also achieves different quality levels for different parts of the image. Specifically an arbitrary region of interest (ROI) in the image can be encoded progressively up to lossless.

The different quality levels represent a rate distortion trade-off. For example, lossless quality might be required for a small part of the image where the user's real interest is located. The rest of the image, which is only important in a contextual sense, is then assigned a lower quality. This can maintain a fairly high compression while meeting the lossless requirement.

The algorithm is based on the completely reversible integer wavelet transforms S [4], S+P [12] and Two Ten (TT) [11]. The transformation coefficients are then encoded in an embedded fashion. The idea of embedded coding originated with wavelet transform and zero-tree coding [15]. Embedded coding can also be done in a DCT framework [9]. However, when lossless encoding is desired, reversible wavelet transforms is the method of choice, at least if the encoding should be naturally progressive all the way.

Another option is to use residual coding [16], where a lossy algorithm is first used to generate an approximation of the original image and then a simple or more elaborate lossless algorithm is used to encode the remaining differences. This puts restrictions also on the lossy part of the scheme since the lossy reconstruction has to be the same on all platforms. However, with careful implementation this is possible with any lossy scheme. For a transform coding scheme it means that the transform should be implemented to yield the same result on all platforms. Other lossy schemes, like vector quantization for example, naturally give the same results.

It should be noted that although it should be desirable in the same range of applications, there is a technical difference between achieving an ROI with better quality than the background and achieving a lossless ROI. In the former case the coefficients affecting the ROI the most can be improved upon to give it a better quality but when lossless encoding is the goal, all coefficients with a possibility of affecting the ROI should be reversibly encoded and should preferably be as few as possible.

The proposed scheme solves the ROI problem in such a way that:

• The coding is naturally progressive and without switching algorithm (no residual coding as in [10]).

• The ROI or ROIs can have arbitrary shape.

• The quality of the ROI is guaranteed and then the quality of the surroundings is degrading gracefully so that there are no visually annoying edges around the ROI.

• The functionality of the solution is not dependent on the entropy coding scheme. It is also independent of the method used for coding the shape of the ROIs.

The paper is organized as follows. Section 2 describes the method as it was implemented. In Section 2.1, an algorithm overview is given. Section 2.2 is a review of the S, S+P and TT transforms. Section 2.3 describes how to derive a mask so that these transforms can be used for progressive coding of images with perfect reconstruction of selected ROIs. Section 2.4 describes the normalization of the transform. 2.5 Successive refinement, 2.6 Coefficient ordering and coding give information on the progressive coding of the coefficients with modifications of the method described in [13]. Section 3 gives results and comparisons. Section 4 is a short discussion. Conclusions are drawn in Section 5.