• Multi-resolution approach
Other Motion
Estimation • Translational alignment
approaches • Block matching approach
• Image Registration
Multi-resolution approach or Hierarchical
approach or Coarse-to-Fine motion estimation
Coarse-level means low resolution image
Fine-level means high resolution image
Idea is to reduce displacement between the frames by half in each
level till we get a shift or a displacement of around 1 pixel between
the video frames. Here we can apply LK method for motion
estimation.
Procedure:
Step 1: Construct a Guassian pyramid as shown on the right.
High resolution image is subsampled to get a lowest level image
with small motion between successive frames (1pixel)
Step 2: Motion estimation at Coarse-level using LK method.
Step 3: Using OF to generate finer level image 2 from image 1 by
warping and upsampling. Warped & upsampled image will be closer
to original image 2 at that resolution.
Step 4: Step 1 to Step 3 repeated till we get highest resolution
images
[Link] [Link]
Procedure for down-sampling & up-sampling
Downsampling (Finer to Coarser level)
• A smoothing filter (to avoid aliasing or
overlapping of pixels) is first applied on the
original image. This smoothed image is then
subsampled, most commonly by a factor of
two in both the horizontal and vertical
directions.
Upsampling (Coarser to Finer level)
• First upsample followed by interpolation
filter (preserves image features). Motion between the frames is more
than 1 pixel (LK not possible)
Watch youtube video from Columbia university for more information on Multi-resolution approach
[Link]
Translational Alignment
• Process of adjusting the position of
one image relative to another by only
considering translations, which
include movements along the x and y
axes.
• Difference in radiographic
measurements in the first and second
EOS [Link] to the image on the
right
• [Link]
4418/11/11/2141
If the motion vector
is fractional, then
interpolation
functions used to
align the images
more accurately.
[Link]
Block matching approaches
Block Matching Algorithm is a way of locating matching macroblocks in a sequence of digital video frames for the
purposes of motion estimation.
Number of block-matching algorithms do exist.
Block matching algorithms used for Video compression.
This is local motion estimation method as motion vectors are estimated for each block rather than entire image.
Matching criteria is used to find the best match for each block in the frames at t & t-1.
Basic underlying assumptions:
• No change in the ambient lighting.
• Objects are rigid.
• Objects are translated on a 3D scene parallel to the image plane.
U=[d1,d2] displacement vector
Block size is (m1,m2)
Block appear at (n1,n2) in image xt
• Block matching is a standard technique of encoding motion in video compression
applications.
• There exist two forms of block matching. The first is Fixed Size Block Matching
Algorithm (FSBMA) and the other is Variable Size Block Matching Algorithm
(VSBMA).
• In fixed size block matching, an image is subdivided into blocks of uniform size,
whereas in variable size block matching small block cover areas of complex
motion, while regions of uniform motion are spanned by large blocks.
H.261, H.263 video compression standards or codecs are based on Block-matching
[Link]
Image Registration
It is a process that overlays two or more images from various imaging
equipment or sensors taken at different times and angles, or from the
same scene to geometrically align the images for analysis
• [Link]
/10.1142/S0219519409003085
Proposed Methodology
Figure 1 shows the flowchart of the proposed algorithm. There are
three main modules: preprocessing, image registration, and image
subtraction. The two images used for matching are referred as
reference and floating image. Reference image is an image captured
before the treatment (pre-operative) and floating image is captured
after the treatment (post-operative). The aim of this algorithm is to
match reference image with registered floating image and to detect
the change between them. Intra oral periapical X-ray images are only
used by this algorithm. Importance is given to root apex region to
identify bone formation.
Spectral subtraction
The main goal of registration algorithm is to detect the change between the refer- ence and registered images. Image
subtraction is used to get the difference image between the two images. Various types of image subtraction algorithms exist
in the literature. Spectral subtraction is tried in this algorithm. In fast fourier transforms (FFT) both images are computed. FFT
of reference image with a subtraction factor of 0.9 is subtracted from FFT of registered image. Subtraction parameter of 0.9
gives a better result. Spectral subtraction gives a better result than spatial subtraction. Original size of the reference and
registered images are used for image subtraction.
Answer following questions:
1. Matching criteria used for Image Registration?
2. How spectral subtraction is performed?
3. Why is it called multi-resolution approach?
4. What is the application of this method in dentistry?
5. How 2D digital images are obtained?