Graphics Reference
In-Depth Information
13.3.2 Motion Compensation
The signal-to-noise (SNR) improvements and motion blur reduction comes directly
from the alignment of LDR images to produce an HDR image. Without motion
compensation, each pixel would integrate the captured light much like a camera with
long exposure. There are manymethods to enable image alignment. For example, one
can use a gyro and accelerometer to detect camera motion and align image frames.
However, this approach requires good initial calibration, precise timing, and drift
adjustment to work accurately. We advocate further sub-pixel accuracy in alignment
using image-based analysis.
In image-based alignment, feature based registration methods [
32
] work well to
achieve this if each LDR image has sufficient SNR. However, global methods are
preferred in SNR-limited low light conditions. Zhang et al. [
33
] have shown that a
hierarchical motion estimation algorithm using Laplacian pyramid methods exhibits
robust behavior for extremely low SNR (
1.0) imagery. For our approach, this robust
motion estimation is applied to each LDR image. We also reduced processing latency
by having the first captured LDR image from a subgroup (typically in a group of
four images) used as the reference. The estimated true image is thus the weighted
average of these four aligned LDR images.
∼
13.4 Contrast Enhancements
To improve the feature feasibility on a display or downstream video analytics, we
apply global and local contrast enhancement to the output HDR image. This would
also ensure better contrast and full range display on monitors or displays which
typically have low color range. For example, when we fuse LDR images (8-10 bits
per pixel, we can generate a HDR scene with higher dynamic range (10-14 bits per
pixel) and yet standard displays provide only 8 bit color depth. Using our contrast
enhancement step, we can automatically scale different regions of the image to fit
the highest dynamic range within the low color depth, while preserving the contrast.
Our global contrast enhancement is based on a logarithmic and exponential map-
ping function [
34
]. The processing would map the HDR image to the globally com-
pressed image based on the average of both the logarithmic and exponential mapping
functions. Our local contrast enhancement technique boosts the weak high spatial
frequency contrast features of an image while reducing the gain of its low spatial
frequency features in local regions. The source image is first decomposed into a
Laplacian pyramid [
35
]. Then the local contrast and a contrast map pyramid are
generated. At each pyramid level, this map is multiplied with the source pyramid to
create a normalized pyramid. Finally, the Contrast Normalized image is obtained by
the inverse transform of the pyramid. To ensure the best display, the output image is
typically gamma corrected, either by the algorithm or the monitor itself.
