THE PROBLEM OF CONTOURS IN TRANSFORM IMAGE CODING
Parrilli, Sara (2008) THE PROBLEM OF CONTOURS IN TRANSFORM IMAGE CODING. [Tesi di dottorato] (Inedito)
Full text disponibile come:
Object contours contribute to a large extent to the perceived quality of an image but are typically quite hard to compress. As a matter of facts, many coding algorithms fail to describe efficiently this information. The problem of contours in image compression is the leading thread of this thesis work, as well as wavelet inefficiency in describing this piece of information. This very same subject is treated in this PhD thesis work under three different scenarios, that is, three different attempts to overcome wavelet limits on images contours: object-based coding; new directional transforms; and adaptive lifting scheme. As regards the object-based image coding paradigm, we analyze costs and advantages of an object-based scheme based on Li and Li’s wavelet shape-adaptive (SA-WT) and shape-adaptive SPIHT. Our aim is to assess the rate-distortion performance of such an object-based coder by means of numerical experiments in typical situations of interest, and single out, to the extent possible, the individual phenomena that contribute to the overall losses and gains. We also extend the object-based paradigm to the class of multispectral images. In this context the object-based scheme can be declined in two cases: class-based and region-based paradigms. The analysis of the rate-distortion performance for both schemes, with reference to remote-sensing images, prove the potential of object-based paradigms for multispectral images. The second scenario refers to the new directional transforms. In this context, we present a new compression technique based on the contourlet transform. Preliminary results on NLA quality lead us to use actually a hybrid wavelet-contourlet decomposition. Then, the SPIHT coder is adapted to the new transform, with the main design problem being the definition of suitable significance trees that took into account the correlation of coefficients across scales, space and directions. Even if the transform is slightly redundant the rate-distortion performance is good, especially for highly textured images, and the visual quality of directional details is better than that of the conventional wavelet/SPIHT coder. The last solution analyzed is the adaptive lifting scheme. We show how to estimate the coding distortion in the transform domain for two interesting classes of adaptive lifting schemes. The basic idea is that the nonlinearity of these schemes can be seen as a time-variable behavior. In this way, we can compute the weights allowing us to estimate the distortion in the transform domain via a weighted average of subband distortions. Experimental results show that by using these weights the distortion assessment becomes very reliable. As a consequence, coding techniques based on distortion minimization benefit from a better distortion estimation and provide better performance.
Solo per gli Amministratori dell'archivio: edita il record