irst, we used a default value
of λ = 0.0 to generate one hundred halftone QR codes and run the
manual scan procedure as in experiment I. We found that 91% of the
results can be successfully read by all barcode readers, meaning that
our representation model could handle large type of images with no
difficulty. For the rest of 9% results which failed in one or more
barcode readers, we regenerated results with λ = 0.3 and λ = 0.6,
and repeated the experiment. We hypothesize that the readability
of these results could be effectively improved as we increase the λ
(see Figure 8). Figure 7(b) shows the average success rates against