Adaptive Control Mechanism



1, f or G (i, j) ≥ threshold 0, otherwise

(4-10)

Finally, the contend-based subsample mask (CSM) is generated by merging the edge mask and the subsample mask, as shown in (4-11). According to the cal-culation of CSM, the subsample rate in CSA (CSR), denoted as R_s, is N² -to-csm cnt, where -to-csm cnt is the number of 1’s in CSM and N²is the macro-block size. Figure 4-6 demonstrates an example of CSM in which the subsample rate is 64-to-27.

CSM (i, j) = SM_8:m(i, j) ∨ EdgeMask (i, j) , 0 ≤ i, j ≤ N − 1.

(4-11)

4.4.3 Adaptive Control Mechanism

In the previous subsection, the content-based subsample mask is generated by merging the regular generic subsample mask with edge mask of high frequency pixels. Since the edge pixels is determined by the edge threshold parameter, the lower the edge threshold parameter, the more the edge pixels will be took. In the prior work, we used constant edge threshold parameter to determine the edge mask [39]. However, the same edge threshold parameter turns out diverse number of edge pixels for various video clips. In Table 4.I., we analyzed the effect of edge threshold parameter on the subsample pixels by setting the edge threshold parameter from 0.05, 0.1, 0.2, 0.3, 0.4, 0.5 to 0.6. As the simulation results shown in the table, standard deviation is up to 13.846 between these 21 test clips for the same edge threshold parameter. Since the subsample rate (N²-to-subsample

Merge

Cotent-Based Subsample Mask

Edge Pixels

Figure 4-6: The components of a content-based subsample mask (CSM)

pixels) is related to how many corresponding processing elements will be disabled for saving power consumption, the diverse property of edge threshold parameter causes obstacle to implement accurate power-aware architecture.

In order to conquer the obstacle, the content-based subsample algorithm pro-poses an adaptive mechanism to self-optimize the edge threshold parameter to converge the subsample pixels. Figure 4-7 shows the block diagram of the adap-tive control mechanism. As described in the previous section, the host processor receives the status of battery pack and decides the best target subsample pixels (trg cnt) according to the battery profile. By the target subsample pixels, the adaptive mechanism recursively updates the edge threshold parameter m^t+1₁ (x, y) according to the difference between target subsample pixels and current

subsam-Table 4.I.: Analyze the effect of edge threshold parameter m1 on subsample pix-els.

m₁ 0.05 0.1 0.2 0.3 0.4 0.5 0.6

akiyo 205.414 176.522 136.769 110.035 92.805 82.260 74.973 children 178.668 149.623 119.201 101.613 89.612 81.177 74.757 coastguard 226.269 200.438 157.329 125.502 103.123 87.978 77.955 container 211.320 183.773 145.082 118.056 99.330 86.446 77.758 dancer 213.506 191.270 153.491 123.626 102.685 88.545 77.913 destruct 204.833 170.554 128.369 103.568 88.558 79.085 72.849 flower 204.992 177.966 140.167 113.887 95.937 84.277 75.844 foreman 190.506 153.741 116.262 96.951 85.588 78.253 73.124 hall monitor 201.773 172.598 136.236 112.599 96.687 85.678 77.743 mobile 184.738 158.141 124.977 104.341 90.300 80.638 74.081 mother daughter 208.210 178.797 139.426 113.509 96.484 85.003 76.194 news 197.857 168.008 130.289 107.245 92.678 82.839 75.606 paris 187.107 157.579 124.165 104.327 91.431 82.612 76.093 sean 203.642 173.333 135.699 110.464 94.222 83.595 75.872 silent 203.741 173.749 132.628 106.703 90.764 80.832 74.483 singer 207.794 179.213 141.512 116.214 98.662 86.565 77.972 stefan 203.708 177.468 141.179 115.379 97.696 85.556 77.092 table tennis 215.165 184.407 138.681 109.122 91.020 80.171 73.374 tempete 197.329 165.714 128.063 104.663 89.659 79.982 73.699 waterfall 221.736 191.497 144.527 113.015 93.324 81.379 74.054 weather 169.356 149.140 125.092 108.309 97.317 87.226 80.078 Average 201.794 173.025 135.197 110.435 94.189 83.338 75.786 σ 13.749 13.846 10.522 7.052 4.646 3.059 1.967

Edge

Figure 4-7: The block diagram of the edge-determination unit with adaptive con-trol mechanism.

ple pixels (csm cnt) frame by frame, as shown in (4-12).

m^t+1₁ (x, y) = m^t₁(x, y) + K_p· (csm cnt − trg cnt) ; if (m^t+1₁ (x, y) < 0) {m^t+1₁ (x, y) = 0};

if (m^t+1₁ (x, y) > 1) {m^t+1₁ (x, y) = 1};

(4-12)

where m^t+1₁ (x, y) is the threshold parameter of macro-block (x, y) in the (t + 1)-th frame and K_pis the control parameter. Figure 4-7 illustrates the block diagram of the the edge-determination unit with the proposed adaptive control mechanism.

As shown in (4-12), the control parameter K_p that will affect the settling time and stationary state error of subsample rate. If the control parameter is not well-selected, the settling time would be too long to have real-time switching and the CSR error would be so large to make the setting of power consumption mode in-accurate and the power-awareness worse. The control parameter K_p in Fig. 4-7 is the major factor to affect the settling time and the CSR error. In order to analyzing the effect of Kpupon the subsample rate and stationary error, we simulated 21 test

Table 4.II.: Average stationary error for 21 video clips with K_p = 0.2.

Target 96.000 128.000 160.000 192.000 224.000 HPF 95.913 127.827 159.771 191.359 221.731 Error 0.091% 0.135% 0.143% 0.334% 1.013%

SBL 96.026 127.818 159.712 191.261 221.659 Error 0.027% 0.142% 0.180% 0.385% 1.045%

MPH 96.553 128.015 159.734 191.370 222.179 Error 0.576% 0.012% 0.166% 0.328% 0.813%

clips for 30 frames with 1 : 1 of the initial subsample rate and 8 : 5 of the target subsample rate. Figure 4-8 shows the effect of the K_p selections of four clips as illustrations.

Obviously, the higher the value of K_p, the shorter the settling time and the worse the stability of the CSR are. If the K_pis too large, the real subsample pixels will be overshoot and oscillatory around the target subsample pixels. Although the settling time is under three frames, the overshoot is up to 32 pixels worse in the weather clip and hard to converge. On the other hand, the settling time will be longer than twenty frames while the Kp is set too tiny. As shown in the response plot of various K_p, the suitable range of K_p is from 0.1 to 0.3. In this range, all the test clips can be converge to the target subsample pixels under ten frames and less overshooting. That is, for a real-live video clip with 30 frames each second, the system can be operated on the target power mode under 0.33 second.

Table 4.II. presents the average subsample pixels and station error of three filters with the proposed adaptive mechanism. The simulation results clearly show that the stationary error can be kept less than 1.045%.

Finally, Table 4.III. shows the simulation results by the proposed adaptive con-trol mechanism. In order to comparing with the fixed edge threshold parameter shown in Table 4.I., we set the target subsample pixels as the average

subsam-0 5 10 15 20 25 30

Response time of Dancer clip

Frame number

Response time of News clip

Frame number

Response time of Paris clip

Frame number

Response time of Weather clip

Frame number

Figure 4-8: Response time of four clips. (a) The Dancer Clip. (b) The News Clip.

(c) The Paris Clip. (d) The Weather Clip.

Table 4.III.: Analyze the effect of controlled edge threshold parameter on subsam-ple pixels.

Target 201.794 173.025 135.197 110.435 94.189 83.338 75.786 akiyo 201.430 172.929 135.121 110.373 94.155 83.813 76.256 children2 200.458 172.210 135.142 111.296 95.576 85.236 78.055 coastguard 201.198 172.537 134.989 110.477 94.563 84.081 76.705 container 200.960 172.690 135.206 110.601 94.411 84.011 76.518 dancer 200.848 172.613 135.572 111.167 95.083 84.667 77.370 destruct 200.837 172.315 134.867 110.587 94.899 84.504 77.361 flower 201.033 173.026 135.515 110.816 94.898 84.843 78.578 foreman 201.463 172.843 134.975 110.415 94.773 84.639 77.552 hall monitor 201.043 172.787 135.279 110.733 94.602 84.321 76.948 mobile 200.948 172.907 135.129 110.859 95.157 84.401 76.925 mother daughter 200.324 172.884 135.071 110.562 94.665 84.139 76.720 news 201.223 172.951 135.313 110.496 94.291 83.890 76.603 paris 200.924 172.855 135.240 110.786 94.741 84.260 76.926 sean 200.409 171.941 134.242 109.320 93.614 83.268 76.264 silent 199.972 172.756 135.039 110.336 94.025 83.898 76.366 singer 201.256 172.798 135.075 110.273 94.159 83.605 76.281 stefan 201.532 173.145 135.469 110.790 94.985 84.371 77.060 table tennis 200.460 171.941 134.546 110.115 94.315 83.886 76.715 tempete 201.527 173.155 135.454 110.875 94.894 84.144 76.853 waterfall 201.794 173.080 135.254 110.526 94.368 83.790 76.459 weather 198.632 172.282 136.317 113.543 98.618 88.744 81.883 Average 200.870 172.697 135.182 110.712 94.800 84.405 77.162

σ 0.690 0.365 0.404 0.763 0.981 1.089 1.234

ple pixels of these 21 test clips with the fixed edge threshold parameter. As the results showed, the standard deviation can be reduced to 0.365 and the worse is only 1.234, which is much better then the result 13.846 with uncontrolled edge threshold parameter.