| In this dissertation the two most significant problems associated with Generalized Lapped Orthogonal Transformation (GenLOT) based image/video compression are addressed and computationally simple yet effective solutions are developed.; Specifically, first ringing artifacts as a compression induced distortion is addressed. In this direction, a mathematical morphology based post-filtering algorithm is developed which successfully eliminates ringing artifacts with exceptional edge and fine detail protection performance. Through the introduction of a morphologically generated “edge mask” edge and fine details of the input image are protected against undesired filtering. A second morphologically generated mask namely the “filtering mask”, is designed to mark the image regions to be filtered against ringing. Further within the support of the filtering mask, human visual system (HVS) masking properties are exploited to avoid undesired filtering and blurring of genuine image detail locally masking the ringing artifacts. For the desired smoothing within the filtering mask, gray-level morphological filters are employed. In spite of being a post-processing scheme, the proposed algorithm introduces only a negligible post-processing delay. The proposed algorithm also has very low computational complexity and is very well suited for hardware based real-time implementation.; Also, a new objective measure namely the “visible ringing measure” (VRM), is introduced to quantify the ringing distortion in a manner consistent with HVS. VRM avoids the need for the artifact-free original image and achieves exceptional compatibility with human perception.; As the second problem, lost transform domain data concealment within the context of GenLOT encoders is considered. In this direction, two algorithms one in the spatial domain and the other in the frequency domain, are developed for the restoration of the impaired decoding region. The spatial domain approach exploits a vector-space constraint owing to the special overlapping structure of the GenLOT basis functions, to reduce the error power of an optimal interpolation based on a simple autoregressive signal model. The frequency domain based approach, makes use of the enhanced interblock correlations of GenLOTs within a linear optimal estimation framework. Both algorithms are computationally simple and successful in their performance. |
