Image Registration: Preprocessing Operations
Arthur Goshtasby Wright State University Image Registration and Fusion Systems
Preprocessing operations
All operations performed on images that improve the registration performance. These include:
Noise filtering Deblurring Region extraction Edge detection Ed d t ti
A. Goshtasby
Noise smoothing
Given image f(x,y) and smoothing filter h(i,j), noise smoothing is defined by:
The intensity of pixel (x,y) in the output is obtained from a weighted sum of intensities of pixels at and around (x,y) in the input. h(i,j) is the weight of pixel (x+i,y+j) in the neighborhood of (x,y), and the sum of the weights over all i and j is1 is1.
A. Goshtasby 3
Mean filtering
When the weights defined by the filter are all the same, the operation is known as mean filtering filtering. Intensities of all pixels at and around a point in input have the same effect on the intensity at the same point in output. This operation is not rotationally invariant if the filter kernel is not circular.
Image containing g g Zero-mean noise
Filter-radius 2 pixels Filter-radius 4 pixels p p Computed using FFT algorithm
A. Goshtasby 4
Gaussian filtering
If the weights in filter kernel represent g p Gaussian coefficients, a Gaussian filter is obtained. Gaussian filtering is effective when image noise is zero-mean. Gaussian filtering is rotationally invariant. A 2-D Gaussian can be decomposed into 2 1-D Gaussians: G( ) = G( ) * G(y); th f G i G(x,y) G(x) G( ) therefore, filtering filt i can be carried in 1-D rather than in 2-D
A. Goshtasby 5
Computation of mean and Gaussian filtering G i fi i
Although filtering is a convolution operation and can be computed using the FFT algorithm, since FFT considers an image is a periodic signal, if left and right image borders, or top and bottom image borders are not the same, artifacts will appear near the image borders. To avoid this, carry out the computations directly. i di l
Image containing g g Zero-mean noise
Computed with FFT Computed directly p p y Gaussian filter of = 2 pixels
A. Goshtasby 6
Image segmentation
This is the process of partitioning an image into meaningful parts. There are two main approaches to image segmentation segmentation.
Methods that use information within regions Methods that use information on the boundary between regions
A. Goshtasby
Intensity thresholding
Assuming an image that contains objects with intensities that represent a G Gaussian di ib i and a background with i distribution d b k d ih intensities that represent a different Gaussian distribution, the objects can be separated from the background using the intensity at the valley between the two histogram modes. modes This method works well when an image contains homogeneous objects and a homogeneous background and the properties of the objects and the background are different. different
Threshold value
Original
Smoothed
Histogram
Thresholded image
8
A. Goshtasby
Intensity thresholding
Threshold value
A Landsat image
Histogram of the image
Thresholding at the valley between the first two peaks.
Thresholding at the average intensity of highest-gradient pixels.
A. Goshtasby 9
Threshold selection
The threshold value is computed by:
Finding the valley between the modes of the histogram of the image. image Finding the intensity that represents the average of intensities of high gradient pixels. high-gradient Finding the intensity at which a change in the intensity will minimally change the segmentation y y g g result.
A. Goshtasby
10
Interactive thresholding
Delineated tumor
Data courtesy of Kettering Medical Center Software courtesy of Marcel Jackowski, Yale University Jackowski
A. Goshtasby 11
Edge detection
Edge detection methods can be categorized into those g g that search for locally maximum image gradient magnitudes and those that search for zero-crossings of the Laplacian of an image. image Methods that search for gradient peaks do not pick false edges but the ones picked could be disconnected. Methods that search of the zero-crossings of the Laplacian image find closed boundaries, but parts of the boundaries could be false.
A. Goshtasby 12
LoG edge detector
Determination of the LoG of an image involves computation of:
or,
A. Goshtasby
13
LoG edge detection 2 D 2-D
An x-ray angiogram
LoG zero-crossing edges
A. Goshtasby
14
LoG edge detection: 3 D 3-D
A volumetric MR brain image LoG edges Data courtesy of Kettering Medical Center
A. Goshtasby 15
Edge detection by intensity ratio
Zero-crossing edges can be determined by convolving d i db l i the negative and positive parts of the LoG with an image separately and subtracting the convolved images and locating the zeroc oss gs. crossings. If instead of subtracting corresponding values in the convolved images, we divide g , the values and locate the one-crossings, intensity-ratio edges will be obtained.
A. Goshtasby
LoG operator in 2-D
Positive part Negative part
16
Intensity ratio edges
(a)
(b)
(c)
(a) Intensity ratio edges. (b) 30% highest-gradient ratio edges. (c) Intensity difference edges. (d) 30% highestgradient difference edges.
Intensity ratios detect more edges in dark areas while intensity differences detect more edges in bright areas. areas
A. Goshtasby
(d)
17
Canny edge detector
This method smoothes the image with a Gaussian to reduce g noise and then locates locally maximum gradient magnitudes in the gradient direction and uses them as the edges.
An x-ray angiogram x ray
Canny edges with = 2.5 pixels
A. Goshtasby 18
Edges in color images
Edge detection in a color image can be considered edge detection in a 2 D vector field 2-D field. If u(x,y) and v(x,y) represent color gradients in x and y directions, edges can be considered points where color gradients are locally maximum in the gradient direction. If R(x,y), G(x,y), and B(x,y) represent red, green, and blue color components at (x,y), respectively, color gradients are:
r, g, and b are unit vectors along red, green, and blue axes, g g g respectively, in the color space.
A. Goshtasby 19
Color edges
Gradient direction at (x y) is the direction (x,y) maximizing
and is obtained from
A. Goshtasby
20
Edge detection in l i Ed d t ti i color images
A color image
Edges of the color image
A. Goshtasby
21
Preprocessing references
1. 2. 3. 4. 5. 6. 7. 7 8. D. Marr and E. Hildreth, Theory of edge detection, Proc. R. Soc. Lond., 207:187217 (1980) 207:187 217 (1980). J. J. Clark, Authenticating edges produced by zero-crossing algorithms, IEEE Trans. Pattern Analysis and Machine Intelligence, 11(1):4357 (1989). J. Canny, A computational approach to edge detection, IEEE Trans. Pattern Analysis and Machine Intelligence 8:679 714 (1986) Intelligence, 8:679714 (1986). L. Ding and A. Goshtasby, On the Canny edge detector, Pattern Recognition, 34:721725 (2001). M. Bomans, K.-H. Hohne, U. Tiede, and M. Riemer, 3-D segmentation of MR images of the head for 3-D display IEEE Transactions on Medical Imaging, 3 D display, Imaging 9(2):177183 (1990). A. Goshtasby and H-L Shu, Edge detection by curve fitting, Image and Vision Computing, 13(3): 169177, 1995. A. Goshtasby, A Goshtasby On edge focusing, Image and Vision Computing, 12(4): 247 256 focusing Computing 247256, 1994. L. Zagorchev, A. Goshtasby, and M. Satter, R-snakes, Image and Vision Computing, 25: 945959, 2007.
A. Goshtasby
22