Analog parallel processor solutions for video encoding

Abstract

This thesis deals with Cellular Nonlinear Network (CNN) analog parallel processor networks and their implementations in current video coding standards. The target applications are low-power video encoders within 3rd generation mobile terminals. The video codecs of such mobile terminals are defined by either the MPEG-4/H.263 or H.264 video standard. All of these standards are based on the block-based hybrid approach. As block-based motion estimation (ME) is responsible for most of the power consumption of such hybrid video encoders, this thesis deals mostly with low-power ME implementations.

Low-power solutions are introduced at both the algorithmic and hardware levels. On the algorithmic level, the introduced implementations are derived from a segmentation algorithm, which has previously been partly realized. The first introduced algorithm reduces the computational complexity of ME within an object-based MPEG-4 encoder. The use of this algorithm enables a 60% drop in the power consumption of Full Search ME. The second algorithm calculates a near-optimal block-size partition for H.264 motion estimation. With this algorithm, the use of computationally complex Lagrange optimization in H.264 ME is not required. The third algorithm reduces the shape bit-rate of an object-based MPEG-4 encoder.

On the hardware level a CNN-type ME architecture is introduced. The architecture includes connections and circuitry to fully realize block-based ME. The analog ME implemented with this architecture is capable of lower power than comparable digital realizations. A 9×9 test chip has also been realized. Additionally implemented is a digital predictive ME realization that takes advantage of the introduced partition algorithm. Although the IC layout of the ME algorithm was drawn, the design was verified as an FPGA.