Line-based Image Coding

With the development of digital images in multimedia, the increase of image resolutions and the emergence of new image modalities, image coding is confronted with new challenges in either coding efficiency or computational and memory complexities. Nowadays mainstream image and intra frame coding adopt a block-based coding structure, among which the state-of-the-art coding standard H.264/AVC employs a block-based prediction and a2D transform for intra frame coding. However, restricted by the block structure, the prediction is not accurate enough; thus harms the coding efficiency. On the other hand, in order to exploit inter-block correlations, block-based coding introduces strong dependence among blocks, making it not quite feasible for highly parallel coding to speed up the encoding, while it is also not a good choice in terms of memory complexity. Therefore, this paper proposes the line-based coding to accommodate these circumstances.We first analyze the theoretical performances of three different coding structures: block, line and pixel, when coding a stationary Gaussian source. It turns out that at high bitrates, line-based and pixel-based coding can have the same tight upper bound of the coding gain as that of block-based coding. However, in practical coding, the local statistics cannot be estimated accurately in block-based coding, while with a smaller coding unit, line-based and pixel-based coding own good coding flexibility and local adaptivity. One disadvantage of pixel-based coding is that it is sensitive to quantization noises. These analyses reveal that at least at high bitrates the block-based structure is not always the best while the line-based coding is a good choice either from its good adaptivity or from its low sensitivity to quantization noises. Based on the analyses, we design a hybrid block and line coding scheme for image coding. In particular, three adaptive prediction methods are proposed in line-based coding to adapt to different image characteristics. Experimental results show that line-based and pixel-based coding can achieve comparable or even better performances than the state-of-the-art block-based coding. When the three structures are combined into one scheme to adapt to different image characteristics and bit rates, a superior coding performance is achieved. Up to2dB and1.5dB gains are observed on natural images compared with H.264/AVC and KTA, respectively and an up to8dB gain on compound images.Considering the high coding efficiency of line-based coding, this paper further proposes a line-by-line coding (LBLC) for parallel coding on many cores. In LBLC, an image is coded line by line, while each line is divided into equal-length line units. The coding of line units in the same line are completely independent from each other; therefore they can be assigned to different cores and coded concurrently. In this way, it can achieve a constant and high parallelism. Simulations on a multi-core platform show that LBLC can achieve a near-linear parallelism and a13.9times speed-up is achieved with15cores at the encoder. Ideally, such near-linear speeding relation with the number of cores can be kept for more than100cores. Besides high parallelism, LBLC can achieve comparable or even better performances than H.264above middle bitrates. At near lossless and lossless coding, it achieves a more than lOdB gain and14%bits saving compared with H.264lossy and lossless coding, respectively.Due to the low delay and low memory requirement of line-based coding, this paper further proposes the Line-Cast scheme for continuous satellite image acquisition and transmission to the ground with limited bandwidth, onboard memory and computational complexities. The line-by-line coding and transmission endows it with low delay and low memory, while the distributed source coding approach not only exploits the anisotropic correlations among pixels but also endows it with low encoder complexity. Due to the semi-analog transmission, Line-Cast can achieve a graceful degradation. Theoretical analysis shows that Line-Cast can achieve Shannon’s optimum attainable performance when a high-dimensional lattice is used for bandwidth-matched case. Experimental results on satellite images show that Line-Cast can achieve comparable or even better performances than the state-of-the-art analog scheme, Softcast, and a more than5dB gain over JPEG2000.