| Summary: | 碩士 === 國立臺灣大學 === 資訊工程學研究所 === 103 === Method of integrating a high-resolution (HR) image from multiple observed
low-resolution (LR) images is called multi-frame super-resolution (SR).
There are basically two stages of reconstruction-based SR: registration of LR
images and reconstruction of HR image.
In this thesis, we based on the assumption of intensity consistency, and
tried several optical flow methods as registration method. Also, based on another
assumption: ”forward-backward flow consistency”, we calculated the
confidence of a flow, then brought confidence into HR image reconstruction
to reduce the error caused by mis-registration. But in the test sets like ”resolution
chart”, there will be some errors caused by some patterns with frequencies
that is too high. The errors violates the assumption of intensity consistency,
which will cause fail registration of optical flow method. Thus, we proposed
to applying blur before calculating the flow. The method can prevent optical
flow from failing.
Also, due to the resolution of images nowadays becomes higher and higher,
which will make the reconstruction of HR image need enormous amount of
memory usage and time. The thesis proposed to divide the reconstruction
to multiple parallelable data blocks to reduce memory and time usage, and
proposed multi-thread and GPU speed-ups. As for algorithm speed-up, we
proposed combining nearest neighbors (NN) reconstruction and linear reconstruction to achieve acceleration.
With the method proposed by this thesis, we can make using optical flow
in multi-frame reconstruction-based SR more robust, and reduce the reconstruction
time and peak memory usage.
|
|---|
