Near-lossless data compression
MetadataShow full item record
Near-lossless data compression algorithm and its applications are investigated in this thesis. Near-lossless data compression is an alternative to the lossless compression scheme. The near-lossless compression techniques studied give quantitative measure about the type and the amount of distortion introduced. The near-lossless compression techniques can yield a much higher data compression ratios than the lossless compression techniques. The algorithm proposed in this thesis is a predictive, context-base, graph searched, entropy-coded DPCM (Differential Pulse Code Modulation) technique with a window-based error criterion. The near-lossless data compression algorithm is developed in the context of image compression. An image is compressed in such a way th a t the total error between the original and the coded images over a W x W window around each pixel does not exceed e > 0 in magnitude. This near-lossless error criterion is developed to preserve the image brightness and color. Issues related to practical implementations are discussed in the thesis to solve the zero-state problem. Methods implemented to solve the zero-state problem are: violate the per-pixel error criterion, use lossless coding when zero-state occurs, set the maximum allowed per-pixel error dynamically, limit the number of states at each stage and the ML-algorithm. The near-lossless data compression are applied in two areas: the edge preserving image compression and near-lossless EEG compression. Images are compressed using the near-lossless data compression scheme to preserve the edge information as defined by a Laplacian operator. The EEG signals are compressed based on the predictive models and the sample-to-sample reconstruction errors are limited to a specified range. Compression results obtained for both applications suggest th a t the proposed near-lossless data compression methods are useful in achieving a high compression ratio while preserving the specific information in the original data.