35
Figure 4.2 : (a) White-light image of object mask shaped as the Chinese character
’’light”. Terahertz images reconstructed via compressed sensing using (b) 300 and
(c) 600 magnitude measurements, which are respectively about 30% and 60% of the
total number of image pixels. Both figures display a 32 × 32 image and the pixel size
is Imm x lmm.
4.3 Imaging results: Amplitude-only images
The amplitude-only imaging experiment measures the magnitude of the detected THz
waveform at a frequency of 100 GHz (λ = 3 mm), because near this frequency the
incident THz beam is large enough to illuminate the entire object and has a high
spectral amplitude. Figure 4.2(b) and (c) show the CS reconstruction results, 32 ×
32 images of the object in Figure 4.2(a) with Imm × Imm pixel size, using 300 and
600 measurements respectively. The reconstruction algorithm uses minimization of
the total variation (min-TV) [9]. Appendix A describes the details of a fast algorithm
for min-TV, called the Fast Total Variation deconvolution (FTVd), which uses less
than 10 seconds to compute the reconstruction results in Figure 4.2 in Matlab on
a standard PC [58]. Our system accurately reconstructs both the size of the object
and its millimeter-scale features. Reconstruction using more measurements yields
a sharper image but also adds some artifacts. Sources of noise include laser power
fluctuation and alignment errors between patterns.
Figure 4.3 plots the normalized mean-squared error (MSE) of the reconstructed
images of the object in Figure 4.2 using different number of masks (CS measurements).