Hybrid spatial-frequency compensation image coding technique

Abstract

Frequency Domain Coding has been widely and successfully applied in image coding. For example, the Discrete Cosine Transform (DCT) has been applied in the JPEG image coding standard. However, practical implementations of DCT-based image coding systems have been quite limited due to the nature of image signals. Some aperiodic signals, such as the unit step function and impulsive function, that appear in the image signal content make compression performance degraded by using Frequency Domain Coding method. A significant amount of information appearing in aperiodic signals is discarded so as to enhance the coding efficiency of Frequency Domain Coding. Image quality can degrade greatly, especially when many sharp edges are present in the image signal. In this thesis, Hybrid Spatial-Frequency Compensation (HSFC) image coding techniques are proposed. The HSFC image coding techniques decompose an image into a combination of spatial and frequency components which are then compressed 4 and coded separately. The concept, theory and implementation of HSFC image coding are discussed. Both one dimensional (1-D) and two dimensional (2-D) implementations of HSFC image coding are proposed. In the 1 -D implementation, aperiodic sharp edges are extracted in either the horizontal or the vertical direction and then coded in the spatial domain. In the 2-D implementation, however, the more complex, 2-D aperiodic sharp edges are extracted and coded by a pattern-based codebook method. Experimental results show that the 2-D implementation of HSFC image coding can achieve a 1dB improvement, in terms of Peak Signal-to-Noise Ratio (PSNR), over Frequency Domain Coding - JPEG for some natural images which contain sharp edges. Moreover, 2-D HSFC can yield at least 1.5 times enhancement, in terms of Compression Ratio (CR), when it is comparing with a Spatial Domain Coding - Vector Quantization (VQ) for general natural images. The HSFC image coding technique is also applied to cartoon images for which more than 10 dB improvement in PSNR is achieved.