AbstractThis paper deals with clock truncation coding for gray‐scale images. An image region is segmented into unequal‐sized square blocks according to local luminance variations and luminance values in the block are approximated by triangular plane patches constructed from the luminance values at the four vertices of the block. Computational complexities of the unequal‐sized block segmentations are discussed. Segmentations are classified into three types: top‐down splitting; bottom‐up merging; and generalized split‐and‐merge method. It is derived from the theoretical model that initial block size in the split‐and‐merge method for a given image of size 256 × 256 is 17 × 17. Segmentation by accumulated square error is preferred to human visual sensitivity. Experimental results using a standard image have clarified that the segmentation by the accumulated square error improves the quality of approximated image to 1 to 2 dB compared with segmentation by the mean square error. Moreover, it is confirmed that the approximation by triangular plane patches is superior to JPEG for the compression of the images that h