Download files
Abstract
In this work we investigate adaptive techniques which can be used to improve the performance of highly scalable video compression schemes under resolution scaling. We propose novel content adaptive methods for motion compensated 3D discrete wavelet transformation (MC 3D-DWT) of video. The proposed methods overcome problems of ghosting and non-aligned aliasing artifacts, which can arise in regions of motion model failure, when the video is reconstructed at reduced temporal or spatial resolutions. We also study schemes which facilitate simultaneous scaling of compressed video bitstreams based on both constant bit-rate and constant distortion criteria, using simple and generic scaling operations.
In regions where the motion model fails, the motion compensated temporal discrete wavelet transform (MC TDWT) causes ghosting artifacts under frame-rate scaling, due to temporal lowpass filtering along invalid motion trajectories. To avoid ghosting artifacts, we adaptively select between different lowpass filters, based on a local estimate of the motion modelling accuracy. Experimental results indicate that the proposed adaptive transform substantially removes ghosting artifacts while also preserving the high compression efficiency of the original MC TDWT.
We also study the impact of various MC 3D-DWT structures on spatial scalability. Investigating the interaction between spatial aliasing, scalability and energy compaction shows that the t+2D structure essentially has higher compression efficiency. However, where the motion model fails, structures of this form cause non-aligned aliasing artifacts under spatial scaling. We propose novel adaptive schemes to continuously adapt the structure of MC 3D-DWT based on information available within the compressed bitstream. Experimental results indicate that the proposed adaptive structure preserves the high compression efficiency of the t+2D structure while also avoiding the appearance of non-aligned aliasing artifacts under spatial scaling.
To provide simultaneous rate and distortion scaling, we study layered substream structure. Scaling based on distortion generates variable bit-rate traffic which satisfies the desired average bit-rate and is consistent with the requirements of leaky-bucket traffic models. We propose a novel method which also satisfies constraints on instantaneous bit-rate. This method overcomes the weakness of previous methods with small leaky-bucket buffer sizes. Simulation results indicate promising performance with both MC 3D-DWT interframe and JPEG2000 intraframe compression.