Hybrid transform, spatial decorrelation and unified coding system for image and video compression

Abstract

In this thesis, a novel compression model for still images and video was developed. Instead of coding image indiscriminately as a whole, as in conventional compression scheme such as JPEG and JPEG2000, the new algorithm analyzes image features, classifies into different types and processes each portion accordingly. When there is spatial coherence, spatial decorrelation is performed prior to applying transform with either Discrete Cosine Transform (DCT), Discrete Wavelet Transform (DWT), or their combination (hybrid transform). When spatial coherence does not exist, either DCT, DWT or Hybrid Transform is chosen as the transform technique for best available compression efficiency. In short, the model combines available spatial domain coherence, DCT and short-tap-length DWT in a non-trivial manner to achieve high compression efficiency. Operationally, the model exploits the fact that typical images are always consisting of various image features. Specifically, there are edge-regions, smooth-regions, patterns and irregular regions, which behave rather differently in terms of bit-rate and Signal-to-Noise-Ratio (SNR) when compressed. The model selectively applies pure DCT, pure DWT or the hybrid transform to various regions with respect to their corresponding image-feature characteristics. As a result, the psycho-visual perception of the decompressed image is significantly better than pure DCT schemes but yet the computation overhead for inverse transform is less than typical 9-7 tap wavelet and comparable to classical DCT schemes. The model takes advantage of spatial domain correlation and could manage to further reduce bit-rate by feeding back image feature information and processed image data to the decode loop. In conjunction with the reduced bit-rate, the visual quality is also better. Owing to the hybrid transform scheme and correlation-reduction using processed image data and image-feature information, the transformed coefficients now become much smaller in magnitude and better defined. A new entropy coding scheme for transformed coefficients was developed to take advantage of the structure of coefficients. The scheme fits equally well to Motion-Estimated frames, which is also characterized by the small magnitude coefficients. When comparing with the start-of-the-art JPEG2000 implementations, the compression scheme developed in this thesis is competitive in terms of bit-rate / SNR for “still” and “video” frames. In most cases, the proposed scheme out-performs JPEG2000 for various still images especially the well-known “Barbara” image and “Bike” image, to name a few. For video frames, the scheme always out-performs JPEG2000, especially for fast motion video sequence. The coding scheme, as it can be applied to both “still images” and “video frames”, is referred to as an unified coding scheme.

Date of Award	4 Oct 2004
Original language	English
Awarding Institution	City University of Hong Kong
Supervisor	Yiu Keung CHAN (Supervisor)