CODING ARTIFACT REDUCTION USING ADAPTIVE POST-TREATMENT M. Luong, A. Beghdadi, W. Souidene, A. Le Négrate * L2TI, Université Paris 13 * LEEC, Université Paris 13 99, avenue J. B. Clément 93430 Villetaneuse, France beghdadi, luong,souidene@galilee.univ-paris13.fr, aln@leec.univ-paris13.fr ABSTRACT This paper presents a post-processing method for reducing blocking and ringing artifacts resulting from lossy coding techniques based on the Discrete Cosine Transform (DCT) such as JPEG, H.261, H.263 or MPEG. These coding artifacts appear at low and medium bit rate and strongly affect the image quality. The proposed method is based on a constrained least-squares method using adaptive smoothing post-processing. The objective is to reduce blocking artifact and edge noise while preserving image details. To evaluate the visual image quality, a new blockiness visibility measure is proposed. The method has been tested on some degraded images and video sequences. The obtained results are satisfying and clearly demonstrate the efficiency of the proposed method. 1 INTRODUCTION Lossy image compression codecs suffer from various coding artifacts at low bit rate. One has to find a compromise between the bit rate and the image quality level. For some structured degradations, at our knowledge, there is no satisfying model which helps in finding the best solution for removing or reducing coding artifacts. Image and video compression standards using the Discrete Cosine Transform (DCT) such as JPEG, H.263 or MPEG, are prone to some particular distortions, such as blocking and ringing effects. These artifacts strongly affect the subjective image quality. The blocking effect, visible along block boundaries, is due to the independent and coarse quantization of individual blocks of fixed size (8 x 8 pixels for instance). At higher video bit rate, ringing artifacts which appear as a noise in the vicinity of object contours is clearly visible near contrasted edges. This noise is sometimes referred to as “mosquito” artifact in image sequence. The aim of our contribution is to propose a post-processing method to efficiently reduce these coding artifacts. Existing works on post-processing methods could be classified into image enhancement and image recovery. For image enhancement-based post- processing methods, distortions of decoded image are first identified, and then an operator like filtering is designed to reduce these distortions. Various kinds of filters, in particular, low-pass filters are proposed to attenuate blocking artifact [1]-[3]. Another approach which exploits some characteristics of the Human Visual System for quantifying the local distortions, in order to design an adaptive filter, has been proposed in [4]. This method consists in smoothing the block boundary until the visibility distortion becomes lower than a predefined visibility threshold. Other enhancement methods are based on examination of the structure of the DCT coefficients. The artifacts are then identified as “blocking” or “ringing” effects [5]. If smoothing can remove coding artifacts, it also tends to affect image details. To remedy this problem, the coding artifact problem can be formulated as an image recovery problem using a prior knowledge on the smoothness of the original image. Hence, the enhanced image must remain faithful to the decoded image while satisfying some constraints on the smoothness. Some classical methods consist in using either deterministic or stochastic models. Concerning the deterministic model methods, a Lagrangian technique [6] is proposed to find an optimum solution using a smoothness constraint at the block boundaries. In [7], non-linear model is used and a method based on projection onto convex sets (POCS) is proposed. In fact, this method is implemented through iterative algorithm. Since the reduction of coding artifacts is often performed using a smoothness constraint on block boundary pixels or adjacent blocks, one has to define a homogeneity measure. This homogeneity measure could then be used as a trade- off between block smoothing and edge preservation. Indeed, edges should not be affected by the smoothing process. For exemple, in [8] a region homogeneity measure is used to constrain the POCS algorithm. Other methods based on the same idea use a line process to model region and block borders [9]. In fact, a line process was introduced in [10] to model the presence or absence of discontinuity information. Here, we propose some improvements of the method developed by Kaup which we found very appealing. We propose a different adaptive smoothing scheme. It consists in adapting the smoothing parameters with respect to the local information. To assess the image enhancement an objective quality measure should be used. Many ad hoc image quality metrics have been proposed for such coding artifacts. In particular, we can refer to [11] and [12] for methods based on 1D-FFT and HVS. Here, to evaluate the visual image quality, we propose a new no reference image quality measure, the blockiness visibility measure. This paper is organized into five sections. In the following 0-7803-9243-4/05/$20.00 ©2005 IEEE 347