Smart Lossy Compression of Images Based on Distortion Prediction

Abstract

Images of different origin are used nowadays in numerous applications spreading the tendency of world digitalization. Despite increase of memory of computers and other electronic carriers of information, amount of memory needed for saving and managing digital data (images and video in the first order) increases faster making crucial the task of their efficient compression. Efficiency means not only appropriate compression ratio but also appropriate speed of compression and quality of compressed images. In this paper, we analyze how this can be reached for coders based on discrete cosine transform (DCT). The novelty of our approach consists in fast and simple analysis of DCT coefficient statistics in a limited number of 8x8 pixel blocks with further rather accurate prediction of mean square error (MSE) of introduced distortions for a given quantization step. Then, a proper quantization step can be set with ensuring the condition that MSE of introduced errors is not greater than a preset value to provide a desired quality. In this way, multiple compressions/decompressions are avoided and the desired quality is provided quickly and with appropriate accuracy. We present examples of applying the proposed approach.