Abstract: For reducing the data rate of digitized images, image data are first transformed block by block, subsequently coded and finally temporarily stored in a buffer memory before they are output. Each data block is assigned on the basis of its own image statistics a classification code or factor from a collection of predetermined classes of different high-local resolution and the image data blocks are differently coded according to the class assigned to them. In order to prevent the buffer memory overflowing, a control variable is derived from the extent of the temporarily stored data, with which control variable the data extent of image data blocks is changed.To prevent disturbing worsening of the local resolution of single or several image data blocks, the assignment of the classes to the image data blocks is changed depending on the control variable. In this manner, although fewer structures are resolved, these can be reconstructed with less error, which leads to a better subjective impression.