A generalization of quad-trees applied to image coding

Although quad-trees are not the most successful strategy in image coding, some generalized subdivision schemes have been proposed recently. This work exploits a moderate generalization of quad-trees where tiles are not restricted to be split in both dimensions, which leads to a previously developed graph of anisotropic tiles called “bush”. An algorithm is developed to find the minimal number of tiles to represent shapes, which is used to build a codec for bi-level and indexed color images. Also, a lossy codec based on tile-wise rate-distortion optimized quantization of low-frequency DCT coefficients is developed. The aim of this work is to investigate whether anisotropic tiles have an advantage over square tiles. The results indeed show significant improvements. The lossless algorithm is suitable for images with large continuous regions and high color payload, such as geographical maps. The lossy codec is able to compete with JPEG2000, especially for artificial images.