Adaptive colour decorrelation for predictive image codecs François Pasteau, Clément Strauss, Marie Babel, Olivier Déforges, Laurent Bédat To cite this version: François Pasteau, Clément Strauss, Marie Babel, Olivier Déforges, Laurent Bédat. Adaptive colour decorrelation for predictive image codecs. European Signal Processing Conference, EUSIPCO, Aug 2011, Barcelona, Spain. pp.1-5. hal-00600354 HAL Id: hal-00600354 https://hal.archives-ouvertes.fr/hal-00600354 Submitted on 14 Jun 2011 HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non, lished or not. The documents may come from émanant des établissements d’enseignement et de teaching and research institutions in France or recherche français ou étrangers, des laboratoires abroad, or from public or private research centers. publics ou privés. ADAPTIVE COLOR DECORRELATION FOR PREDICTIVE IMAGE CODECS Franc¸ois Pasteau, Clement´ Strauss, Marie Babel, Olivier Deforges,´ Laurent Bedat´ IETR/Image group Lab CNRS UMR 6164/INSA Rennes 20, avenue des Buttes de Coesmes¨ 35043 RENNES Cedex, France ffpasteau, cstrauss, mbabel, odeforge, [email protected] ABSTRACT codecs such as LAR, JPEGLS or H264 to improve both When considering color images and more generally multi decorrelation and prediction processes of color images. This component images, state of the art image codecs usually new method being completely reversible can be applied to achieve component decorrelation through static color trans- both lossless and lossy coding. forms such as YUV or YCoCg. This approach leads to This new method can be decomposed in two processes. suboptimal results as statistics of the image are not taken First, adaptive decorrelation performed during the prediction into account. The new approach proposed here offers to re- process aimed at removing the residual correlation. A pre- move the correlation of one component according to another vious study [1] has shown good improvement over compres- adaptively during the prediction process of an image codec. sion rate when applying such decorrelation after the predic- Through two jointly used processes, one aiming at choos- tion process. However using adaptive decorrelation during ing the best predictor of a component and another aiming at the prediction process improve the effectiveness of the pre- improving the predictor’s effectiveness, this new approach dictor and therefore leads to both better compression rate and improves both image quality and compression ratio. This better quality. new technique has been applied to the LAR codec and shows In addition, inter component prediction aims at choosing an improvement over previous studies up to 20% in rate and the best fitted predictor for a component using information 0.5db in PSNR at low bit rates. from another previously encoded component of the image. The paper is organised as follows. The following section 1. INTRODUCTION introduces the concept of inter-component prediction. In sec- tion 3 Adaptive decorrelation perfomed during prediction is Despite the general use of color images, developement of im- explained. In section 4 both inter component prediction and age codecs such as JPEG, JPEG2K or the newly standardized adaptive decorrelation are jointly used. Finally an applica- JPEGXR has been primarily focused on giving the best per- tion of these two processes on the LAR codec is shown. formance on one component images. To handle color im- ages, state of the art codecs usually rely on color transforms such as YUV, YCgCo and/or subsampling to achieve both 2. INTER COMPONENT PREDICTION good compression ratio and good visual quality. However In this section we present an approach aiming at improving after performing color transforms and/or subsampling each the prediction of a component using the best predictor of an- component is encoded independently without using the resid- other previously encoded component. In order to propose ual correlation still existing between components. In a pre- algorithms and figures, notations described below need to be vious study [1], we showed that using the RGB color space defined. with an adaptive decorrelation and classification leads to bet- let Y be the first component of the image ter results than using static color transform such as YUV, let C be one of the other components of the image YCgCo. th let Yi be the i value of the Y component of the image When considering lossless coding, coding techniques j rely on statistical analyzes of the image to perform compres- let Ybi be the prediction of Yi through predictor j j sion. As subsampling would cause losses, only reversible let Yei be the reconstruction of Yi through predictor j color transforms can be used. However, even after apply- let aC be the decorrelation factor for the C component ing static color transforms, residual correlation still exists let x0 be x with adaptive decorrelation applied between components [1]. This underlaying correlation re- j let modeprediction(Y ; j) be the process returning Y sults in an suboptimal compression rate as decorrelation has i bi let entropycoder(x) be the entropy coding of x been done statically without consideration of the image it- let Q(x) return the value of x after quantization self. A critical application of lossless coding of colour im- −1 ages concerns cultural digital libraries [2]. Museums actu- let Q (x) return the value of x after inverse quantization −1 ally try to safely digitalize their belongings and thus produce let [x]Q be Q (Q(x)) large quantities of lossless colour pictures. Current digital let sign(x) = -1 if x < 0, 1 otherwise cameras are wide spread and generate high resolution colour let jxj be the absolute value of x images. Professional photographers tend to prefer lossless By applying different predictors from a predictor set on Y, compression of their pictures to avoid artifacts due to image selection of the best fitted to Y can be achieved. Therefore compression. this predictor can be used on component C to obtain a better In this paper we present a new method for predictive prediction. The Algorithm 1 explains the search for the best predictor. The process corresponds to a minimization of the Algorithm 2 Adaptive Decorrelation Algorithm j 0 distance between Ybi −Yei according to the norm 1. 1: fInitializationg 2: cpt = 0 Procedure 1 BestPrediction 3: NC = 0 0 4: DC = 0 Input: Y ;Y i ei 5: aC = 0 Output: bp 6: for all i do 1: best = +inf 0 7: Ybi = modeprediction(Yi;0) 2: bp = 0 0 3: for all j do 8: Cbi = modeprediction(Ci;0) j 0 0 0 4: Ybi = modeprediction(Yi; j) 9: Yei = Ybi + [Ybi −Yi]Q j 0 10: 5: if jYbi −Yei j < best then j 0 11: fAdaptive Decorrelation on predictorg 6: best = jYbi −Yei j 00 0 0 0 12: C = C + a × [Y −Y ] 7: bp = j bi bi C bi ei Q 00 00 00 8: end if 13: Cei = Cbi + [Cbi −Ci]Q 9: end for 14: 10: freturn the best predictor according to Yg 15: faC updateg 11: return bp 0 0 16: if DC > 0 or [Ybi −Yei ]Q 6= 0 then 00 0 0 0 17: NC = NC + (Cei −Cbi ) × sign([Ybi −Yei ]Q) Predictor 0 corresponds to the predictor used on com- 0 0 ponent Y. The effectiveness of such a method is highly de- 18: DC = DC + j[Ybi −Yei ]Qj pendent on the number of predictors and their effectiveness. 19: aC = NC=DC To keep a low computational cost, the number of predictors 20: cpt = cpt + 1 should be kept low. Such process already exists in literature, 21: for example in JPEGLS [3][4]. However the main advantage 22: fEnsure local adaptationg here is that the selection of the best predictor does not need to 23: if cpt > 1000 then be transmitted to the decoder as the decoder can execute the 24: NC = NC=4 same process. Therefore inter component prediction is cost 25: DC = DC=4 free. As it involves modification of predictors, inter compo- 26: cpt = 0 nent prediction is also lossless. 27: end if 28: end if As inter component prediction aims at choosing the right 0 0 predictor for a component, an other approach would be to 29: entropycoder(Q(Ybi −Yei )) improve the effectiveness of the predictor by removing the 00 00 30: entropycoder(Q(C −C )) residual correlation between Y and C. This approach is de- bi ei 31: end for scribed in the next section. 3. ADAPTIVE DECORRELATION DURING 00 PREDICTION PROCESS dictions after decorrelation Cbi to ensure an acurate update of aC. To avoid DC = 0 we perform a point reflection on the In a previous study devoted to inter component decorrelation points of coordinate f(x;y)jy < 0g around the origin (0,0). [1], we perform decorrelation as an independent process of 0 the prediction scheme. By doing so, only the compression Therefore only the absolute value of [Ybi −Yi]Q is taken into ratio is improved, as extra information given by the decor- consideration in DC. To ensure a locally adaptive decorre- relation process is not used to improve the prediction itself. lation, we divide the numerator and denominator of aC after Therefore, by performing adaptive decorrelation during the cpt iterations. The number of iterations and the divider of the prediction process both compression ratio and quality can be numerator and denominator have been empirically evaluated. improved in a single pass. Figure 1 represents the functional As adaptive decorrelation can be realised at both encoder implementation of such a technique. The algorithm used to and decoder side, the process itself is costfree.