Murilo B. de Carvalho

Multidimensional signal compression using multiscale recurrent patterns

In this paper we propose a new multidimensional signal lossy compression method based on multiscale recurrent patterns, referred to as multidimensional multiscale parser (MMP). In it, a multidimensional signal is recursively segmented into variable-length vectors, and each segment is encoded using expansions and contractions of vectors in a dictionary. The dictionary is updated while the data is being encoded, using concatenations of expanded and contracted versions of previously encoded vectors. The only data encoded are the segmentation tree and the indexes of the vectors in the dictionary, and therefore no side information is necessary for the dictionary updating. The signal segmentation is carried out through a rate-distortion optimization procedure. A two-dimensional version of the MMP algorithm was implemented and tested with several kinds of image data. We have observed that the proposed dictionary updating procedure is effective in adapting the algorithm to a large variety of image content, lending to it a universal flavor. For text and graphics images, it outperforms the state-of-the-art SPIHT algorithm by more that 3 dB at 0.5 opp, while for mixed document images, containing text, graphics and gray-scale images, by more than 1.5 dB at the same rate. Due to the way the images are segmented, they appear slightly blocky at low rates. We have alleviated this problem by proposing an effective way of reducing the blockiness in the reconstructed image, with no penalty in signal-to-noise ratio performance in most cases. We conclude the paper with a theoretical analysis of the approximate matching of Gaussian vectors using scales, which gives a justification of why approximate multiscale matching is a good option, specially at low rates.

Scanned Compound Document Encoding Using Multiscale Recurrent Patterns

In this paper, we propose a new encoder for scanned compound documents, based upon a recently introduced coding paradigm called multidimensional multiscale parser (MMP). MMP uses approximate pattern matching, with adaptive multiscale dictionaries that contain concatenations of scaled versions of previously encoded image blocks. These features give MMP the ability to adjust to the input images characteristics, resulting in high coding efficiencies for a wide range of image types. This versatility makes MMP a good candidate for compound digital document encoding. The proposed algorithm first classifies the image blocks as smooth (texture) and nonsmooth (text and graphics). Smooth and nonsmooth blocks are then compressed using different MMP-based encoders, adapted for encoding either type of blocks. The adaptive use of these two types of encoders resulted in performance gains over the original MMP algorithm, further increasing the performance advantage over the current state-of-the-art image encoders for scanned compound images, without compromising the performance for other image types.

Improving multiscale recurrent pattern image coding with least-squares prediction mode

The Multidimensional Multiscale Parser-based (MMP) image coding algorithm, when combined with flexible partitioning and predictive coding techniques (MMP-FP), provides state-of-the-art performance. In this paper we investigate the use of adaptive least-squares prediction in MMP. The linear prediction coefficients implicitly embed the local texture characteristics, and are computed based on a blocks causal neighborhood (composed of already reconstructed data). Thus, the intra prediction mode is adaptively adjusted according to the local context and no extra overhead is needed for signaling the coefficients. We add this new context-adaptive linear prediction mode to the other MMP prediction modes, that are based on the ones used in H.264/AVC; the best mode is chosen through rate-distortion optimization. Simulation results show that least-squares prediction is able to significantly increase MMP-FPs rate-distortion performance for smooth images, leading to better results than the ones of state-of-theart, transform-based methods. Yet with the addition of least-squares prediction MMP-FP presents no performance loss when used for encoding non-smooth images, such as text and graphics.

Intra-prediction for color image coding using YUV correlation

In this paper we present a new algorithm for chroma prediction in YUV images, based on inter component correlation. Despite the YUV color space transformation for inter component decorrelation, some dependency still exists between the Y, U and V chroma components. This dependency has been previously used to predict the chrominance data from the reconstructed luminance. In this paper we show that a chrominance component can be more efficiently predicted by using the reconstructed data from both the luminance and the remaining chrominance signal. The proposed chroma prediction is implemented and tested using the Multidimensional Multiscale Parser (MMP) image encoding algorithm. It is shown that the new color prediction mode outperforms the originally proposed prediction methods. Furthermore, by using the new color prediction scheme, MMP is consistently better than the state-of-the-art H.264/AVC for coding both for the luminance and the chrominance image components.

Improving H.264/AVC Inter Compression with Multiscale Recurrent Patterns

In this paper we describe the ongoing work on a new paradigm for compressing the motion predicted error in a video coder, referred to as MMP-Video. This new coding algorithm uses the multidimensional multiscale parser image coding algorithm to encode the residue error, in a H.264/AVC based video coder. MMP has shown to perform very well as a universal still image coding method, particularly when it is combined with intra prediction schemes. In addition, previously published preliminary results have also presented MMP as a promising video coding method. In this paper, we propose new dictionary updating techniques for MMP-video. Along with other functional optimizations, these techniques allow for a significant improvement in the encoder performance. Thus, we were able to achieve considerable gains over H.264/AVC for B slices, specially for medium and high bit-rates, while maintaining equivalent performance for the P slices.

Multiscale recurrent pattern matching approach for depth map coding

In this article we propose to compress depth maps using a coding scheme based on multiscale recurrent pattern matching and evaluate its impact on depth image based rendering (DIBR). Depth maps are usually converted into gray scale images and compressed like a conventional luminance signal. However, using traditional transform-based encoders to compress depth maps may result in undesired artifacts at sharp edges due to the quantization of high frequency coefficients. The Multidimensional Multiscale Parser (MMP) is a pattern matching-based encoder, that is able to preserve and efficiently encode high frequency patterns, such as edge information. This ability is critical for encoding depth map images. Experimental results for encoding depth maps show that MMP is much more efficient in a rate-distortion sense than standard image compression techniques such as JPEG2000 or H.264/AVC. In addition, the depth maps compressed with MMP generate reconstructed views with a higher quality than all other tested compression algorithms.

H.264/AVC based video coding using multiscale recurrent patterns: first results

The Multidimensional Multiscale Parser (MMP) algorithm has been proposed recently as a universal data coding method. MMP has proved to be a very powerful coding method for images, as for other types of signals. Experimental tests showed that MMP is able to achieve better results than the traditional transform-based image coding methods, particularly for images that do not have a low-pass nature. These promising results motivated the use of MMP for residual error encoding in hybrid video coding algorithms. This paper presents the first results of these experiments, performed using a H.264/AVC based video encoder, but using MMP to encode the motion compensated residual data, for the P and B slices. Experimental results show that, even in this not fully optimised version, this method is able to achieve an approximately equivalent performance to the H.264/AVC. This demonstrates that MMP is an alternative to the transform-quantisation paradigm for hybrid video coding that is worth investigating.

Improving Multiscale Recurrent Pattern Image Coding with Deblocking Filtering

The Multidimensional Multiscale Parser (MMP) algorithm is an image encoder that approximates the image blocks by using recurrent patterns, from an adaptive dictionary, at different scales. This encoder performs well for a large range of image data. However, images encoded with MMP suffer from blocking artifacts. This paper presents the design of a deblocking filter that improves the performance the MMP. We present the results of our research, that aims to increase the performance of MMP, particularly for smooth images, without causing quality losses for other image types, where its performance is already up to 5 dB better than that of top transform based encoders. For smooth images, the proposed filter introduces relevant perceptual quality gains by efficiently eliminating the blocking effects, without introducing the usual blurring artifacts. Besides this, we show that, unlike traditional deblocking algorithms, the proposed method also improves the objective quality of the decoded image, achieving PSNR gains of up to about 0.3 dB. With such gains, MMP reaches an almost equivalent performance to that of the state-of-the-art image encoders (equal to that of JPEG2000 for higher compression ratios), for smooth images, while maintaining its gains for non-smooth images. In fact, for all image types, the proposed method provides significant perceptual improvements, without sacrificing the PSNR performance.

Lossy and lossless image encoding using multi-scale recurrent pattern matching

In this study, the authors investigate the use of multi-scale recurrent pattern matching paradigm for lossless image compression. The multi-scale multidimensional parser (MMP) algorithm is a successful implementation of this paradigm for lossy image compression, and can naturally perform lossless compression since it was first derived from a Lempel–Ziv lossless scheme. However, neither its recently adopted coding tools had been adapted for lossless coding nor a thorough analysis of its performance had been carried out. In this work, the authors evaluate MMPs lossless compression capability, proposing modifications for some of its predictions modes, as well as the inclusion of an adaptive prediction mode based on least squares. The residual information is also coded with well-known techniques used in lossless compression. Experimental results for MMP show that the algorithm achieves a good performance for images such as computed generated graphics and scanned documents, whereas keeping a competitive performance for natural images. Since the algorithms structure is exactly the same for lossless and lossy compression, the obtained results suggest that MMP is able to achieve a high compression performance for a wide range of images and rates, from lossy to lossless, without any prior analysis of the image to be coded.

Optimizing Memory Usage and Accesses on CUDA-Based Recurrent Pattern Matching Image Compression

This paper reports the adaptation of the Multidimensional Multiscale Parser (MMP) algorithm to CUDA. Specifically, we focus on memory optimization issues, such as the layout of data structures in memory, the type of GPU memory – shared, constant and global – and on achieving coalesced accesses. MMP is a demanding lossy compression algorithm for images. For example, MMP requires nearly 9000 seconds to encode the 512 ×512 Lenna image on a 2013’s Intel Xeon. One of the main challenges to adapt MMP to manycore is related to the dependency over a pattern codebook which is built during the execution. This forces the input image to be processed sequentially. Nonetheless, CUDA-MMP achieves a 12× speedup over the sequential version when ran on an NVIDIA GTX 680. By further optimizing memory operations, the speedup is pushed to 17.1×.