Efficient quadtree based block-shift filtering for deblocking and deringing

Abstract

The existing implementations of block-shift based filtering algorithms for deblocking are hard to achieve good smoothing performance and low computation complexity simultaneously due to their fixed block size and small shifting range. In this paper, we propose to integrate quadtree (QT) decomposition with the block-shift filtering for deblocking. By incorporating the QT decomposition, we can easily find the locations of uniform regions and determine the corresponding suitable block sizes. The variable block sizes generated by the QT decomposition facilitate the later block-shift filtering with low computational cost. In addition, large block based shift filtering can provide better deblocking results because the smoothing range of large blocks spans over the conventional 8 × 8 block size. Furthermore, we extend the proposed QT based block-shifting algorithm for deringing JPEG2000 coded images. Experimental results show the superior performance of our proposed algorithms.

Introduction

The most popular still image coding standards, JPEG and JPEG2000, are based on block discrete cosine transform (DCT) and wavelet transform (WT), respectively. At low bit rates, due to the independent transformation and quantization of image blocks, JPEG coded images often exhibit grid-pattern false edges aligned with the block boundaries, which is known as blocking artifact. On the other hand, although the new WT based JPEG2000 provides higher coding efficiency and better visual quality at high compression ratios, it often causes spurious oscillations around strong edges at low bit rates, which is known as ringing phenomenon.

For blocking artifacts, in the last two decades, numerous deblocking algorithms have been proposed. According to their underlying mathematical models, the existing algorithms can be generally divided into two categories: image enhancement based algorithms and image restoration based algorithms [1]. The image enhancement based algorithms typically utilize the domain knowledge that the blockiness only occurs around block boundaries, and apply various filtering approaches to smooth the boundaries including fixed [2] and adaptive [3] approaches, and spatial domain [3] and transform domains (DCT, WT, etc.) [4], [5], [6], [7], [8] approaches. In addition to smooth the boundaries, Al-Fohoum and Reza [9] pointed out that the quality degradation in the interior part of an image block due to heavy quantization of DCT coefficients, mostly observable around image edges, should also be alleviated.

In the second category, the restoration based algorithms typically make assumptions on image or noise models and use certain prior knowledge of the coding process. The classical image restoration methods adopted for deblocking include projection onto convex sets (POCS) [10], [11] and maximum a posterior estimation (MAP) [12], [13]. In general, these restoration based deblocking algorithms can suppress both blockiness and other edge-related artifacts and yield superior processing results. However, the major problem is that they rely on iterative processes to obtain the optimal solution, which has high computation complexity.

Among various deblocking algorithms, block-shift filtering [4], [5] is a fast adaptive algorithm for reducing image artifacts. In our previous work, we have proposed an efficient spatial-domain non-iterative block-shift filtering based deblocking algorithm [14]. Although the proposed algorithm can effectively suppress both blockiness and other edge-related artifacts, there still exists much room for further improvement in terms of computational efficiency and deblocking performance.

In particular, in our previous work and all the other existing block-shift filtering based deblocking algorithms, the block size is always fixed to 8 × 8. However, it is observed that in low bit rate JPEG images there often exist a large number of uniform regions, whose sizes are much larger than 8 × 8. If we could use appropriate large blocks to process these uniform regions, the computation cost can be greatly reduced.

In addition, in all the former implementations of block-shift based deblocking, in order to lower computational complexity, the shifting range is typically fixed to half of the image block size for each individual block, i.e. [−4, 4] for 8 × 8 blocks. Such a small shifting range limits the smoothing performance of block-shift based deblocking.

Therefore, in this paper, we propose to integrate quadtree (QT) decomposition with the block-shift filtering for deblocking. QT decomposition [15], [16], [17], [18] is a powerful yet low-complexity multi-resolution image segmentation technique, which can effectively partition an image into many homogenous regions subjected to some predefined rules. By incorporating the QT decomposition, we can easily find the locations of uniform regions and determine the corresponding suitable block sizes. The variable block sizes generated by the QT decomposition facilitate the later block-shift filtering with low computational cost. In addition, large block based shift filtering can provide better deblocking results because the smoothing range of large blocks spans over the conventional 8 × 8 block size. Furthermore, we extend the proposed QT based block-shifting algorithm for deringing JPEG2000 coded images. Experimental results show the superior performance of our proposed algorithms.

The rest of the paper is organized as follows. Section 2 introduces the proposed algorithm for deblocking JPEG images. Section 3 describes how to apply the same framework of the integration of QT and block-shift filtering for deringing JPEG2000 images. Experimental results and comparative studies are provided in Section 4. Finally, Section 5 concludes the paper.