A Predictive Coding Method for Lossless Compression of Images

Compression is a process, in which the given size of data is compressed to a smaller size. Storing and sending images to its original form can present a problem in terms of storage space and transmission speed. Compression is efficient for storing and transmission purpose. Based on the reconstruction quality, compression can be lossy or lossless. Images used for biomedical research or astrophysical categorization must retain all the quality of the original image because machine analysis can be performed on these images to find very specific details that the human eye cannot detect. If these images are to be transmitted or stored, lossless image compression is needed. In the first part of this thesis, we propose a comparative study, based on literature survey, between the prediction based methods and the transform based methods. From information theoretic point of view, prediction can be explained as an estimation of some unknown quantity from known observation. In prediction, the value of a current pixel is being predicted from some of the previous pixels. These pixels are already stored in the memory when the image is scanned. Subtracting the predicted pixel value from the current pixel value at the same spatial location, we get the error image. This error image is entropy coded. But, in the transform based method, spatial image pixel values used to convert into transform coefficient values. It produces as many coefficients as there are pixels in the image. These coefficients can then be compressed more easily because the information is statistically concentrated in just a few coefficients. These coefficients are quantized and the quantized values are entropy coded. But, in literature survey, we found that for lossless compression, prediction based methods are better than the transform based methods in terms of complexity. The second part of the thesis described a new lossless adaptive prediction based algorithm for continuous tone images. In continuous tone images, spatial redundancy exists. Our approach is to develop a new backward-adaptive prediction technique to reduce the spatial redundancy in a image. The new prediction technique known as Modified Gradient Adjusted Predictor (MGAP) is developed. MGAP is based on the prediction method used in Context based Adaptive Lossless Image Coding (CALIC). An adaptive selection method which selects the predictor in a slope bin in terms of minimum entropy improves the compression performance.