838 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 12,NO. 7, JULY 2003 Removing the Blocking Artifacts of Block-Based DCT Compressed Images Ying Luo and Rabab K. Ward, Fellow, IEEE Abstract—One of the major drawbacks of the block-based DCT compression methods is that they may result in visible artifacts at block boundaries due to coarse quantization of the coefficients. In this paper, we propose an adaptive approach which performs blockiness reduction in both the DCT and spatial domains to re- duce the block-to-block discontinuities. For smooth regions, our method takes advantage of the fact that the original pixel levels in the same block provide continuity and we use this property and the correlation between the neighboring blocks to reduce the discon- tinuity of the pixels across the boundaries. For texture and edge regions we apply an edge-preserving smoothing filter. Simulation results show that the proposed algorithm significantly reduces the blocking artifacts of still and video images as judged by both ob- jective and subjective measures. Index Terms—Blocking artifacts, JPEG, low bit rate coding, MPEG, post-processing. I. INTRODUCTION B LOCK-BASED DCT coding has been successfully used in image and video compression applications due to its energy compacting property and relative ease of implementa- tion. After segmenting an image into blocks of size N N, the blocks are independently DCT transformed, quantized, coded, and transmitted. One of the most noticeable degradation of the block transform coding is the “blocking artifact.” These arti- facts appear as a regular pattern of visible block boundaries. This degradation is a direct result of the coarse quantization of the coefficients and of the independent processing of the blocks which does not take into account the existing correlations among adjacent blocks pixels. To cope with the blockiness problem, different types of post-filtering techniques are developed to reduce the high frequency components near the block boundaries. These methods can be classified as spatial filtering methods [1]–[4], DCT-based filtering methods [5]–[7] or hybrid filtering methods [8]–[13]. Earlier on, Reeve and Lim [1] proposed a symmetric, two-dimensional Gaussian spatial filtering method. However, due to its low-pass nature, this spatial filtering approach has the drawback of visible smoothing of the image. Manuscript received September 24, 1999; revised February 4, 2003. This work was supported by the Natural Sciences and Engineering Research Council of Canada. The associate editor coordinating the review of this manuscript and approving it for publication was Dr. Touradj Ebrahimi. Y. Luo is with Gennum Corporation, Burlington, ON, Canada L7R 3Y3 (e-mail: yluo@gennum.com). R. K. Ward is with the Electrical and Computer Engineering Department, University of British of Columbia, Vancouver, BC, Canada V6T 1Z4 (e-mail: rababw@ece.ubc.ca). Digital Object Identifier 10.1109/TIP.2003.814252 To try to minimize this blurring, various other schemes have been suggested. In [3], [4], and [8]–[11], an edge classification procedure is used to distinguish between monotone, textured and strong edge regions of the image. The spatial filtering scheme is driven by the output of these edge classifiers. Edge and texture regions are usually not filtered since humans are more sensitive to low frequency errors than to high frequency ones. The edge classi- fication is done in the spatial or in the DCT domain. In [3] and [4] spatial methods such as the gradient/threshold and the his- togram methods are used. In [8]–[11], the block classification is performed by examining the DCT coefficients. Projection onto convex sets (POCS) based recovery algorithm [12] is a powerful DCT domain filtering approach. The basic idea is to optimize the value of the quantized transform coeffi- cients, subject to some smoothness and quantization constraints. The major drawback of this approach is its high computational complexity. It takes at least 3 5 iterations for this algorithm to converge, where both forward and inverse DCT are required in each iteration, in addition to other required computational over- head. A filtering strategy, which relies on the estimation of the quantization error for each spatial block is proposed in [13]. is determined by the quantization step and the proba- bility distribution of the original unquantized DCT coefficients. The latter is estimated from the moments of the quantized coefficients. To alleviate the discontinuity offset problem caused by smoothing the boundary pixels, the smoothing process is repeated until the center of the block is reached. Because of quantization, it is difficult to get a good estimate of the probability distribution of the DCT coefficient at each frequency. If we cannot get an accurate estimate of for each block, we may risk over-smoothing or under-smoothing in the filtering process. While the above mentioned methods carry their computa- tions in the spatial or both spatial and DCT domains, many DCT-based methods have also been proposed. In [5], the blocking effects are reduced in the DCT domain by suppressing the block-DCT coefficients. The suppression parameters are determined from the distribution of the DCT coefficients in the coding process. In [6], the transform coefficients are estimated using the local statistics of the quantized coefficients. The reconstructed trans- form coefficients are confined to their original quantization in- tervals. The above two methods require no iterations. In this paper, we propose a new DCT based method for re- ducing the blocking effects in smooth regions of the image. Our 1057-7149/03$17.00 © 2003 IEEE