CV Cheat Sheet

Images Division Scaling

• color depth = bit depth = bits per pixel = I(x, y) = I1 (x, y) ÷ I2 (x, y) • inverse scaling prefered (from dest. to source)
bbp OR to avoid gaps.
• dithering: putting 2 colors close to give I(x, y) = I1 (x, y) ÷ C
illusion of a 3rd color. inhomogeneous
• grayscale img: 8 bpp, black = 0 & white = Blending     
x2 s 0 x1
255. I(x, y) = I1 (x, y) ∗ C + I2 (x, y) ∗ (1 − C) = x
y2 0 sy y1
• color: 24 bpp, 8 bits for each channel of
RGB, & possible alpha channel. α = 0 =⇒
Logical Operations
homogeneous
fully transp., α = 1 =⇒ fully opaque. NOT, AND/NAND, OR/NOR, XOR/XNOR
  
sx 0 0 x1
 
Conversion x2
BMP Format 1. Bitwise: convert values to binary base and 0
 y2  =  sy 0   y1 

file header apply operators bitwise. 1 0 0 1 1
(size, offset, . . . ) 2. Thresholding:
info header (DIB) ( convert each pixel value to 1
0, I > 127,
(width, height, . . . ) bit: Inew = Subsampling (shrink)
optional color palette 1, I ≤ 127.
1. Replacement: replace a group of pixel by a
image data Applications chosen one (upper left).
File Header • AND/NAND: intersection bet. 2 images. 2. Interpolation: mean value of group.
14 bytes • OR/NOR: union bet. 2 images. Upsampling (magnify)
• NOT: negative of input image.
• magic identifier: 2 bytes 1. Replication: fill group of pixels with same
• file size : 4 bytes Logical NOT value of original pixel.
• 2 reserved places: 2 bytes each Ways to apply NOT: 2. Interpolation: get missing values at
• offset to image data: 4 bytes boundaries, then interpolate by distance.
• normal boolean NOT.
Info Header • grayscale: I 0 (x, y) = 255 − I(x, y).
• float pixel format: I 0 (x, y) = −I(x, y),
Reflection
40 bytes
followed by normalization. About x − axis:
• header size in bytes: 4 bytes    
Bitshift Operations 1 0 0 x1

• width and height : 4 bytes each x2
• 0
 y2  =  −1 0   y1 

number of color planes: 2 bytes usage: fast multiplication and division
• number of bits per pixel: 2 bytes 1 0 0 1 1
• Shift left i bits = ∗2i (multiplication)
• compression (0 to 3): 4 bytes, 0 = none • Shift right i bits = ÷2i (division)
• image size in bytes: 4 bytes About y − axis:
Empty Places
Note that the order is B → G → R, & all bytes    
−1 0 0 x1

1. fill with zeros. x2
are written in reverse order.
 0
 y2  =  1 0  y1 

2. fill with ones.
Color Palette 3. fill with bits from other side (rotate). 1 0 0 1 1
• If present, then a pixel is stored in ≤ 1 bytes.
Geometric Operations About a general axis:
• Each color entry is in RGBA format with 4
bytes. Translation 1. Translate to origin.
• If not present, of f set = 14 + 40 = 54, else 2. Rotate abt origin to fit on x- or y-axis.
inhomogeneous
of f set = 54 + 4 ∗ nColors. 3. Reflect about this axis.
4. Rotate back.
     
x2 x t
= 1 + x
Arithmetic Operations y2 y1 ty 5. Translate back.

Addition homogeneous Affine Transformation

I(x, y) = I1 (x, y) + I2 (x, y)     • 6 degrees of freedom.
1 0 t x x1

OR x2 • needs 3 pairs of points to estimate.
0
 y2  =  1 t y   y1 

I(x, y) = I1 (x, y) + C
1 0 0 1 1 inhomogeneous
Overflow       
x2 a a12 x1 t
I(x, y) > max (255)? = 11 + x
Rotation y2 a21 a22 y1 ty
1. Wrapping: I 0 (x, y) = I(x, y) − (max + 1)
• inverse rotation prefered (from dest. to
2. Saturation: I 0 (x, y) = max homogeneous
source) to avoid gaps.
• y upwards =⇒ counterclockwise is +ve,  0   
Subtraction x a11 a12 a13 x
y downwards =⇒ clockwise is +ve. y 0  = 
a21 a22 a23  y 

usage: detect changes between 2 images.
I(x, y) = I1 (x, y) − I2 (x, y) Rotate about origin: 1 a31 a32 a33 1
OR inhomogeneous
I(x, y) = I1 (x, y) − C      How to get matrix A from 3 pairs of points?
x2 cos θ − sin θ x1
Underflow =    −1  0 
y2 sin θ cos θ y1 a11 x1 y1 1 x1
I(x, y) < 0? a12  = x2 y2 1 x0 
2
homogeneous a13 x3 y3 1 x03
1. Wrapping: I 0 (x, y) = I(x, y) + (max + 1)    
cos θ − sin θ 0 x1

2. Saturation: I 0 (x, y) = 0 x2    −1  0 
3. Absolute: I 0 (x, y) = |I(x, y)| a21 x1 y1 1 y1
 sin θ
 y2  =  cos θ 0   y1 

1 1
a22  = x2 y2 1 y20 
0 0 1 a23 x3 y3 1 y30
Multiplication
usage: enhance contrast Rotate about arbitrary pt:    −1  
a31 x1 y1 1 1
I(x, y) = I1 (x, y) ∗ I2 (x, y) 1. Translate to origin. a32  = x2 y2 1 1
OR 2. Rotate about origin. a33 x3 y3 1 1
I(x, y) = I1 (x, y) ∗ C 3. Translate back.
Homography Transformation Triangular Filter
• 8 degrees of freedom. • similar to mean filter, but weights are diff.
• needs 4 pairs of points to estimate. • filter is always normalized (divide by sum of
 0    weights).
x h11 h12 h13 x
y 0  = h21 h22 h23 

 y  1 2 3 2 1
1 h31 h32 h33 1 2 4 6 4 2
1
3 6 9 6 3 ∗ 5 × 5 pyramid. kernel
h11 x + h12 y + h13
0
81
x = 2 4 6 4 2
h31 x + h32 y + h33
1 2 3 2 1
h21 x + h22 y + h23
y0 = 0 0 1 0 0
h31 x + h32 y + h33 0 2 2 2 0
How to get matrix H from 4 pairs of points? 1 2 5 2 1 ∗ 1
5 × 5 cone kernel (R = 2.5)
25
Let h33 = 1, M = 0 2 2 2 0
x1 y1 1 0 0 0 −x01 x1 −x01 y1
 
0 0 1 0 0
0 0 0 x1 y1 1 −y10 x1 −y10 y1 
x2 y2 1 0 0 0 −x02 x2 −x02 y1  Gaussian Filter (Blur)
 
0 0 0 x2 y2 1 −y20 x2 −y20 y2 
 
Gaussian in 1D:
x3 y3 1 0 0 0 −x03 x3 −x03 y3 
 
2
0 x3 y3 1 −y30 x3 0
−y3 y3  1 − x2
0 0
 
G(x) = √ e 2σ
x
4 y4 1 0 0 0 −x04 x4 −x04 y4  2πσ
0 0 0 x4 y4 1 −y40 x4 −y40 y4 Gaussian in 2D:
  0
h11 x1
 2 2
1 − x +y
h12   y10  G(x, y) = √ e 2σ2

h13  x02 
    2πσ
h21   y20 
   
M =  1 4 7 4 1
h22  x03  4 16 26 16 4
h23   y30 
    1
h  x0  7 26 40 26 7 ∗ 272
5 × 5 Gauss. (σ = 1)
31 4 4 16 26 16 4
h32 y40 1 4 7 4 1
Digital Filters Median Filter
Convolution • reduces noise, but preserves details.
• replace each pixel with the median of its
For an n × m kernel K:
m n
neighbourhood.
2
X 2
X • sort values (keep duplicates) and pick the
I(x, y) = I1 (x + k, y + l)K(k, l) middle one (or average of two middles if
k=− m
2
l=− n
2 even).
• median(I1 + I2 ) 6= median(I1 ) + median(I2 ).
Noise
Kuwahara Filter
1. Salt & pepper noise: the color of a noisy
pixel has no relation to surrounding pixels. • edge-preserving filter, doesn’t disturb
2. Gaussian noise: each pixel is changed from sharpness and position of edges.
PN
its original value by a small amount. 2
i=1 (I(xi ) − mean)
Variance: σ 2 =
Smoothing Filters N
Linear Filters
1. Calculate mean and variance of each 3 × 3
LF ilter(I1 + I2 ) = LF ilter(I1 ) + LF ilter(I2 ) region (upper left, upper right, lower left, &
1. Uniform (mean) filter lower right).
2. Triangular filter 2. Output value of center pixel = mean value of
3. Gaussian filter region of smallest variance.
Non-linear Filters
1. Median filter
2. Kuwahara filter
Uniform (Mean) Filter
• replace each pixel with the mean of its
neighbourhood.
• all coeffs. in kernel have same weights.
• smoothing effect increases with kernel size.
• filter is always normalized (divide by sum of
weights).
1 1 1
1
1 1 1 ∗ 9
3 × 3 rectang. OR circ. kernel (R = 1.5)
1 1 1
1 1 1 1 1
1 1 1 1 1
1
1 1 1 1 1 ∗ 25
5 × 5 rectang. kernel
1 1 1 1 1
1 1 1 1 1
0 1 1 1 0
1 1 1 1 1
1
1 1 1 1 1 ∗ 21
5 × 5 circ. kernel (R = 2.5)
1 1 1 1 1
0 1 1 1 0