Block adaptive linear predictive coding with adaptive gain and bias

DETAILED DESCRIPTION According to one aspect of the present invention, the object is achieved by taking into account the block-to-block variations of the centroid (preferrably mean) of the error signal from zero as well as the block-to-block variations of the range (preferrably sigma value) of the error signal, and adapting to these variations by selecting from a set of two-variable quantizers optimally designed to minimize the square difference between the input error signal and the coded error signal.
Excluding the first pixel in each image line and the first line of pixels (which are directly quantized) the error signal is formed for a given pixel by making a prediction of the pixel value from previously coded error signals that are in close proximity to the pixel value in two dimensions, and subtracting the prediction from the pixel value.
The error signals in a linear block are then quantized by selecting from a set of two-variable quantizers designed for the class of images to which the current image belongs, by performing the following steps: (1) forming independent linear blocks of error signals; (2) forming, maximum likelihood (ML) estimates for a function of the mean and sigma value of the error signals in a block; (3) from the minimum square error decision regions of the mean and sigma values for the class of images, determining which two-variable decision region contains the ML estimates; (4) using the minimum square error reconstruction values for sigma (gain) and mean (bias) from the selected region to construct the minimum square error quantizer for the error signal in the block; (5) quantizing the error signal in the block employing the constructed quantizer; (6) transmitting or storing the quantized (coded) error signal and a code word for the selected quantizer; and (7) proceeding to the next block and repeating the steps (1)-(6) until the line of image signals is processed.
According to another aspect of the present invention, the object is achieved by identifying runs of blocks having low contrast, and encoding these blocks with a flag indicating a run, and a value indicating the number of blocks in the run