SPRUJ28E November 2021 – September 2024 AM68 , AM68A , TDA4AL-Q1 , TDA4VE-Q1 , TDA4VL-Q1
With combination of just level 1, levels 1+2, and levels 1+2+3, horizontally and vertically, the result feature detection filter kernels are shown in Figure 6-85. Note that all coefficients are powers of 2, and therefore do not require multiplication.
G1...9 are the counter-part reconstruction kernels. Each of G1...9 is a row-wise and column-wise mirror of the counterpart F function.
The kernel representation in Figure 6-86 illustrates a filter alignment.
Figure 6-86 is showing an academic 8x8 pixel image frame which has been border extended on all 4 edges. Superposed onto this border extended frame (shown in dotted lines) is the kernel alignment for all 3 different kernel sizes for each of the three levels:
For each of the 9 different kernels, a matrix of (width + 7) x (height +7) decomposition values result. All of these result matrixes are exactly aligned.
The kernel alignment is top-left aligned for the first pixel. Which translates in the recursive approach as a X-Y shift. In order to be mathematically aligned in the Recursive approach, the Level 1 output needs to be shifted 6 positions in both the X and Y direction relative to the Level 3 output. Additionally, the Level 2 output needs to be shifted 4 positions in both the X and Y direction relative to the Level 3 output. These shifts are critical such that Level 1, 2, and 3 values are cycle aligned when presented to the thresholding circuit.
One can reach the same conclusion by looking at the figure and thinking about which line delay or which horizontal pipelined pixel position is used in order for the kernels to output their results in the same pipeline cycle.