A hybrid between block-based and segmentation-based image coding is proposed. The coder divides each edge block into two fragments and transform encodes each fragment separately. The encoder is controlled by an optimality principle, to minimise the code length for any prescribed level of distortion.