Comparative Study of Near-Lossless Compression by JPEG XR, JPEG 2000, and H.264 on 4K Video Sequences - Advances in Image and Graphics Technologies - page 31

Macroblock is a basic data unit in JPEG XR. The size of a macroblock is prede-
fined by the standard. One tile consists of NxM Macroblocks and one macroblock
consists of 4x4 blocks which consists of 4x4 pixels.
The data flow of JPEG XR encoding is shown in Fig. 2.
Fig. 2. JPEG XR Encoding Diagram
First of all, pre-processing is applied to the input image for better coding efficien-
cy. Then the color conversion module that converts the original image color to anoth-
er color space which is better suited for down-sampling is applied.
Then a transform called photo core transform (PCT) is applied to decompose an
image into frequency components. The PCT is applied to a rectangular area called a
macroblock. A transform called photo overlap transform (POT) is used with the PCT
to eliminate the artifacts in the boundary between blocks. The influence of overlap
filter on coding efficiency is studied in Section 2.2
Next, the transformed coefficients are quantized. Quantization parameters (QP)
give a great impact on the image quality. After the coefficient prediction, an adaptive
scanning is processed and coefficients are rearranged from two-dimensional form to
one dimensional form.
Finally, the scanned coefficients are entropy coded using adaptive Huffman tables.
Different from JPEG, JPEG XR encodes DC, LP, HP frequency bands separately. Not
all values are entropy encoded. A process called normalization is used to decide
which part of the bits should be kept as plain or be encoded. The plain bits in the DC
and LP bands are called refinement while in the HP band they are called flexbits as
shown in Fig. 3.
Flexbits is sent uncoded. For lossless 8 bit compression, Flexbits may account for
more than 50% of the total bits. Flexbits forms an enhancement layer which may be
omitted or truncated, where the parameter is called “TrimFlexBits”.