High ratio wavelet video compression through real-time rate-distortion estimation.

Abstract

The success of the wavelet transform in the compression of still images has prompted an expanding effort to exercise this transform in the compression of video. Most existing video compression methods incorporate techniques from still image compression, such techniques being abundant, well defined and successful. This dissertation commences with a thorough review and comparison of wavelet still image compression techniques. Thereafter an examination of wavelet video compression techniques is presented. Currently, the most effective video compression system is the DCT based framework, thus a comparison between these and the wavelet techniques is also given. Based on this review, this dissertation then presents a new, low-complexity, wavelet video compression scheme. Noting from a complexity study that the generation of temporally decorrelated, residual frames represents a significant computational burden, this scheme uses the simplest such technique; difference frames. In the case of local motion, these difference frames exhibit strong spatial clustering of significant coefficients. A simple spatial syntax is created by splitting the difference frame into tiles. Advantage of the spatial clustering may then be taken by adaptive bit allocation between the tiles. This is the central idea of the method. In order to minimize the total distortion of the frame, the scheme uses the new p-domain rate-distortion estimation scheme with global numerical optimization to predict the optimal distribution of bits between tiles. Thereafter each tile is independently wavelet transformed and compressed using the SPIHT technique. Throughout the design process computational efficiency was the design imperative, thus leading to a real-time, software only, video compression scheme. The scheme is finally compared to both the current video compression standards and the leading wavelet schemes from the literature in terms of computational complexity visual quality. It is found that for local motion scenes the proposed algorithm executes approximately an order of magnitude faster than these methods, and presents output of similar quality. This algorithm is found to be suitable for implementation in mobile and embedded devices due to its moderate memory and computational requirements.