5 Image visualization and enhancement Flashcards
Perception of color
•Color perception takes place in the human eye and the associated part of the brain
•Tristimulus mode
–The retinas in human eyes have cones and rods.
•The knowledge of the Tristimulus phenomenon led to the development of color television screens and monitors
–Color screens are composed of a large number of small dots arranged in a regular pattern of groups of three (Red/Green/Blue)
–Color is created by mixing different amounts of Red/Green/Blue
–The mixing takes place in our brain.
–We get the same impression when we see yellow light as when we see a mixture of red and green.
Perception of color - Tristimulus mode
Cones
- Cones are light-sensitive receptors that send signals to the brain when hit by photons in different parts of EM spectrum
- There are three different types of Cones corresponding to red/green/blue lights.
- The signal received by human brain gives a sensation of color.
Perception of color - Tristimulus mode
Rods
- Rods sense brightness.
- They can sense low-intensity light
- Do not contribute to color vision
Color spaces - RGB (Red/Green/Blue)
–Three colors are added: additive color scheme
–Light beams that are superimposed mix their colors additively
–Used in color television screens and monitors
–White: all three colors mixed equally
–Magenta: red and blue
–Yellow: red and green
–Cyan: green and blue
Color spaces - YMC (Yellow/Magenta/Cyan)
–Three colors are subtracted: subtractive color scheme
–Each component acts like a filter that subtracts a primary color from whitelight
•Cyan = white - red
•Magenta = white - green
•Yellow = white - blue
–YMC color scheme is used in printers
Color spaces - IHS (Intensity/Hue/Saturation)
–A more natural reflection of our sensation of colors
–A color is described by three values
•Hue: name of color e.g. red, green, purple (degrees 0 360)
•Intensity : a color is light or dark (0 1)
•Saturation : description in terms of purity (percentage 0 100)
Image display thingy (1/2)…
- A computer screen usually have a certain pixel depth (e.g. 8 bits)
- When displaying an 8-bit gray scale image on a monitor with 8-bit pixel depth, a pixel having the value zero is displayed as black and a pixel having the value 255 will be shown as white
Image display thingy (1/2)…
Single-band data are shown as
–Grey scale image e.g. Osh region, Kyrgyzstan, August 2016, SPOT-6
–Pseudo color image e.g. land surface temperature; land use & land cover
Image display thingy (2/2)…
- Computer screens usually have three channels, each fed by grey scale image
- When we have multi-band imagery, any combination of bands can be used as input to the RGB channels of the monitor
- The choice should be made based on the intended use of the RGB image
•Color composites
–True color composite
–False color composite
–Pseudo natural color composite
Image display thingy (2/2)…
Color composites: True (natural)
–True color composite corresponds to how we see the world –Vegetation appears green –Water appears blue to black –Snow and clouds appear white –Bare soil appears light gray and brown
Image display thingy (2/2)…
Color composites: False
–NIR –> R
–Red –> G
–Green –> B
–Vegetation appears in a red purple color
Image display thingy (2/2)…
Color composites: Pseudo-natural
–Also called false-color composite
–Different pseudo-natural color composites can be formed for different applications
Sentinel-2 band combinations
What do you know about Sentinel-2?
Sentinel-2, an optical remote sensing sensor, launched by the European space agency, is a constellation with twin satellites which carry sensors called multi-spectral instrument (MSI)
Sentinel-2 band combinations
True color composite (4,3,2)
fact 1/2
–Healthy vegetation: green
–Recently cleared vegetation: very light
–Unhealthy vegetation: brown and yellow
–Roads and urban features: white and gray
–Contrast can be low due to the scattering of blue light in the atmosphere
Sentinel-2 band combinations
True color composite (4,3,2)
fact 2/2
–Clouds and snow appear white and difficult to distinguish
–Shallow water is not distinguished from soil
–Difficult to distinguish vegetation types
–Difficult to distinguish healthy vegetation from stressed vegetation
Sentinel-2 band combinations
The disadvantage of true-color composite (4,3,2)
- high scattering in blue spectrum resulted in low contrast image
- Cant distinguish between cloud and snow
- It is difficult to distinguish healthy vegetation in true-color image
Sentinel-2 band combinations
False color composite (5,4,3)
fact
–Usually called Color Infrared
–Chlorophyll has a high reflection in NIR and low reflection in visible bands
- Vegetation: shades of red
- Soils: dark to light brown
- Urban areas: cyan/yellow/grey
- Water: dark blue
Sentinel-2 band combinations
The advantage of False color composite (5,4,3)
Esier to distinguish different types of vegetation than in true color composite
Sentinel-2 band combinations
Pseudo natural color composite (11,8,2)
fact
–Chlorophyll has a high reflection in NIR
–High reflectance in SWIR1 when water content is low.
–Vegetation: shades of green
*notes
Healthy vegetation = band blue and SWIR1 (red) small, NIR 1 (green) big. So healthy vegetation appears green
Sentinel-2 band combinations
The advantage of Pseudo natural color composite (11,8,2)
Easier to distinguish different types of vegetation than in true color composite
Image enhancement
•Histogram operations
–Global contrast enhancement
–Location of pixels are not important
•Filter operations
–Enhancing some aspects of the image and suppress other aspects
–Based on neighboring pixel values
Image histogram
- A histogram describes the distribution of pixel values in a band of an image
- without considering the location of pixels
A histogram can be summarized by descriptive statistics
–Minimum and maximum
–Mean
•Average of all pixel values
•Does not necessarily coincide with the most frequent value
–standard deviation (low std means histogram not spread all over the certain range of an image)
•Indicates the spread of the DNs around the mean
•Small standard deviation: narrow histogram, low contrast
–n% and 100 n% values
Image enhancement
Histogram operations
- Remote sensors are designed to detect a wide range so they are capable of measuring different illuminations/emission conditions
- The full range is not always given within one particular scene
•Gray scale transformation stretches the histogram over the entire range to achieve a better visualization
–Linear contrast stretch
–Histogram equalization
Image enhancement
Histogram operations: Linear contrast stretch
–Cut off values are selected •Values higher/lower than cut off values are considered as noise •Min/Max from the data •N% cut off •mean NxStd
–Values are linearly transferred in a way that selected cut off values become 0/255
–Straightforward method for contrast enhancement
–Works best when the input image has a narrow histogram, but close to a uniform distribution
Image enhancement
Histogram operations: Histogram equalization
–Nonlinearly transforms the values in a way that spreads out the most frequent intensity values
–Histogram equalization improves the global contrast of the image Histogram operations
Image enhancement
Filter operations - applications
–Enhance image e.g. reduce noise (low pass filters) or sharpen a blurred image (high pass filters)
–Extract features e.g. edges and lines
Image enhancement
Filter operations - image filters
*note
Filtering is usually carried out on a single band
–Defined through a kernel
–The kernel is moved through the image
–A new value is calculated for each pixel as a linear combination of itsneighboring pixels.
–The kernel defines the size of neighborhood to be considered
Image enhancement Filter operations
Moving average filter
–Calculates the local average of pixels within a kernel around each pixel
–Does not need to be square
–Is used to smooth the image
Image enhancement Filter operations
Gaussian filter
–Is used to smooth the image
–An approximation of Gaussian function
–Gives less weight to pixels further away from the center
–Is used to smooth the image
Sharpening
- Smooth filter the image
- Subtract the smoothed image from the original to highlight the details
- Add the details to the original image
original - smoothed = detail
original + detail = sharpened
Image enhancement Filter operations
Median filter
–The median of all values within a kernel is assigned to the pixel
–Useful in reducing random noise
Image enhancement Filter operations
Sobel edge detection
–Approximate derivatives in two directions # Hx Horizontal kernel -> if we have short change in horizontal direction, it can be detected by this Hx Horizontal Kernel # Hy Vertical kernel -> If we have high derivatives in the vertical direction, it can be detected by Hy vertical kernel
–Gradient magnitude is estimated from the gradient approximations
G = (Gx^2 + Gy^2)^0.5
Image enhancement Filter operations
Sobel edge detection note
Before applying Sobel edge detection, we apply smoothing filter (Gaussian filter) on the original image. Because Sobel edge detection is sensitive to any changes in the image