Fast Scaling Technology In Video Transcoding

Fast image scaling with high fidelity plays a key role in video processing and codec by providing various resolution, reducing bitrates to facilitate video transmission and achieving the best visual quality at the receiver. In fact, scaling technology is an essential component in many image/video processing such as to magnify/minify still image, to achieve special effect like picture-in-picture, to compress or transform digital video with spatial scalability etc. In this thesis we mainly employ this technology in video transcoding, which is a powerful but computationally intensive tools to convert one kind of video format into another, meeting different demand such as lower bitrates or various display terminals. In sum, there are two kinds of scaling schemes: one is in spatial domain and the other in transform domain. The scheme in transform domain can provide very good performance, that is, efficient computation and desirable image quality, whereas it has some disadvantage which is impossible or hardly to overcome, especially some scenarios require arbitrary ratio zoom in horizontal and vertical, and excellent visual effect. However, the other scheme in spatial domain can easily tackle these problems via FIR filter, providing more flexibility and better quality. Note that we achieve this goal at expense of numerous computation and frequent memory I/O. Actually, it is not feasible to put so much workload on scaling which is only a subsection of whole transcoder. Through polyphase decomposition of FIR filter, we can avoid a mass of redundant computation and decrease memory I/O. We can further speed up the filter process by taking advantage of MMX instructions that operate on a set of data in parallel. After simplification of FIR filter structure and optimization of implementation process, we not only meet the real-time demand but also get perfect image quality.