Biomedical Engineering Reference
In-Depth Information
Equalized
Output
Predicted
Hologram
Measured Hologram
(Bayer Pattern Image)
Equalization
of red &
green pixels
Estimation
of blue
pixels
Back
Propagation
Reconstructed
Lensfree Image
Recovered
Hologram
Iterative
refinement of
blue pixels +
Phase recovery
Mask
Holographic
Reconstruction
Object
Support
Fig. 4.7
Summarizes our de-Bayering algorithm that creates monochrome holographic images
from Bayer patterned raw outputs of the lensfree cell-phone microscope shown in Fig.
4.6
. Red and
green channels of the acquired raw holographic image are equalized using a background image that
was recorded with identical illumination conditions as the object. Blue pixels are then estimated
from their red and green neighbors using an edge-aware interpolation approach. The predicted
holograms are further refined (i.e., digitally corrected) through an iterative recovery process with
the help of an automatically generated object support constraint [
11
]. The recovered hologram is
digitally reconstructed using our custom-developed algorithms
calculate a grayscale de-Bayered holographic image with high signal-to-noise ratio
(SNR) using the uncompressed raw Bayer pattern image captured with the cell-
phone camera [
11
]. We specifically capture raw images as opposed to compressed
color images as the color images are demosaiced, which may result in loss of high-
frequency information in the holograms. In order to utilize most of the pixels of
the color sensor with high SNR, we employed an LED with a center wavelength
of
587 nm, where
both
the green and the red pixels of the sensor are highly
responsive. Since a Bayer pattern is composed of two green pixels, one blue and one
red pixel, only the blue pixels constituting 25 % of the total pixels are inefficiently
illuminated by the LED. As a result, the information in these pixels have very
low SNR, and our de-Bayering algorithm aims to recover the values of these blue
pixels with high SNR based on the information in the neighboring green and
red pixels, using an edge-aware interpolation algorithm [
11
]. In this computation
step, the red and green channels of a recorded raw holographic image are initially
equalized using a background image, serving as a calibration measurement that only
needs to be done once, which is recorded with identical illumination conditions in
the absence of objects. After channel equalization and the initial interpolation to
estimate the values of the missing blue pixels, we further refine the recovered pixel
values through an iterative recovery process that uses the object size as a priori
information, which is indeed digitally obtained with an automatically generated
object-support mask. In this iterative refinement process, holograms are propagated
back and forth between the sensor and the object planes using the object-support
constraint as described in Sect.
4.2
, and both the
amplitude
and
phase
values of the
recovered holograms are updated for the blue pixels, while only the
phase
is updated
