A pixel-wise directional intra prediction method

Abstract

In a decade where the prominence of video applications has become increasingly prevalent, the optimization of video coding processes remains as important as ever. The directional intra prediction has been proved as an effective tool for intra picture coding. However, in a directional intra prediction based coding structure, there is a considerable gap between the transform efficiency of DCT and the optimal transform KLT. In this paper, we analyze these problems and propose a pixel-wise directional intra prediction (PDIP) method to solve this problem. This method exploits the reconstruction value of the adjacent pixels to predict the current pixel, and keeps the block-based lossy coding structure of H.264/AVC. The proposed method can significantly improve the coding efficiency with little decoding complexity increment.

Introduction

The digital video compression techniques can be traced to the principle that some accuracy of pixels composing a static image can be sacrificed without notable deterioration in the image’s viewing quality to the human eyes. Video compression takes advantage of the reasonable level of error allowed for the pixels of an image in order to analytically predict a majority of the image’s pixels based on knowledge of only a few reference pixels [1].

Techniques for video compression prediction can be classified into inter prediction coding and intra prediction coding. Inter prediction coding predicts pixels using the relationship between temporal adjacent frames for temporal estimation. Intra prediction coding removes the spatial redundancy in each separate frame to compress videos. Since the intra coding is applied within one picture, it is essentially the same as the static image compression. There are increasing requirements for intra only video compression, such as film-making, digital cameras and video editing applications. Efficient intra coding is also important for random access applications. In this paper, we will study the intra coding of current standards, and improve the intra coding method to introduce significant coding gain, with the tradeoff being only a relatively small increment in complexity.

The state of the art video coding standard H.264/AVC [2], [3] developed by VCEG (ITU-T SG16 Q.6) and MPEG (ISO/IEC JTC 1/SC 29/WG 11) has been well-known for achieving a much higher coding efficiency than the last generation video coding standards [4]. For example, it saves more than 50% bit-rate compared to MPEG-2 [5] while preserving almost the same visual quality [3].

The most highlighted intra picture coding technique of H.264/AVC is the directional spatial prediction [2]. In this coding structure, the reconstructed pixels from the adjacent blocks are used to predict the current coding block and generate the residue block (The detailed procedure will be illustrated in Section 2.1). The residue block is then transformed, and the coefficients are then scanned, quantized and entropy coded sequentially, as illustrated in Fig. 1.

For a classic intra coding structure like MPEG-2, DCT has been proved to be able to archive very close coding efficiency to KLT [6]. However, [7] proved that the transform efficiency of DCT is far less than KLT while directional intra prediction is employed, and the coding efficiency can be further improved by more than 4% while a separable mode dependent transform derived from KLT is used for each directional intra prediction mode [8]. The defect of the mode dependent KLT is that we should design one transform matrix for each mode. Employing several transformations will significantly improve the implementation complexity of both the encoder and decoder, especially for hardware implementation. Moreover, a KLT is normally not mathematically separable, also does not have fast algorithms. There were some other researches focused on further simplifying the KLT for directional intra prediction, and achieved considerable coding gain [9].

This paper proposes a much simpler method, which improves the prediction process and keeps the transform process unchanged. Test results show a performance improvement of about 2.5% average bit-rate savings or equivalently 0.2 dB PSNR gain for intra coding.

The rest of this paper is organized as follows. The directional prediction in H.264/AVC is briefly introduced in Section 2; the problem brought by directional prediction is also discussed. The proposed prediction method which solves the problem is introduced in Section 3. Section 4 gives the experimental results, and the conclusion is drawn in Section 5.