A multi-dimensional measure for image quality

Summary form only given. It is necessary to develop a quality measure that is capable of determining (1) the amount of degradation, (2) the type of degradation, and (3) the impact of compression on different frequency ranges, in a reconstructed image. We discuss the development of a new graphical measure based on three criteria. To be able to make a local error analysis, we first divide a given image (the original or a degraded) into areas with certain activity levels using, as in the case of Hosaka plots, a quadtree decomposition. The largest and smallest block sizes in our decomposition scheme are 16 and 2, respectively. This gives us 4 classes of blocks having the same size. Class i represents the collection of i×i blocks; a higher value of i denotes a lower frequency area of the image. After obtaining the quadtree decomposition for a specified value of the variance threshold, we compute three values for each class i (i=2,4,8,16), and normalize them according to: (1) the number of pixels/the number of pixels in the entire image; (2) the number of distinct pixel values/the number of possible pixel values; and (3) the average of the standard deviations in the blocks/a preset maximum standard deviation. The essential characteristics of the image are then displayed in a normalized bar chart. This lays the foundations for designing optimized image coders