Image Enhancement
Download
Report
Transcript Image Enhancement
Chapter
Image Enhancement
Analysis and applications of remote
sensing imagery
Instructor: Dr. Cheng-Chien Liu
Department of Earth Sciences
National Cheng Kung University
Last updated: 7 July 2015
Introduction
Image enhancement
• Mind excellent interpreter
• Eye poor discriminator
• Computer amplify the slight differences to make
them readily observable
Categorization of image enhancement
• Point operation
• Local operation
Order
• Restoration noise removal enhancement
Contrast manipulation
Gray-level thresholding
• Segment
• Fig 7.11
(a) TM1
(b) TM4
(c) TM4 histogram
(d) TM1 brightness variation in water areas only
Level-slicing
• Divided into a series of analyst-specified slices
• Fig 7.12
Contrast manipulation (cont.)
Contrast stretching
• Accentuate the contrast between features of interest
• Fig 7.13
(a) Original histogram
(b) No stretch
(c) Linear stretch
Fig 7.14: linear stretch algorithm, look-up table (LUT) procedure
(d) Histogram-equalized stretch
(e) Special stretch
• Fig 7.15: Effect of contrast stretching
(a) Features of similar brightness are virtually indistinguishable
(b) Stretch that enhances contrast in bright image areas
(c) Stretch that enhances contrast in dark image areas
• Non-linear stretching: sinusoidal, exponential, …
• Monochromatic color composite
Spatial feature manipulation
Spatial filtering
• Spectral filter Spatial filter
• Spatial frequency
Roughness of the tonal variations occurring in an image
High rough
e.g. across roads or field borders
Low smooth
e.g. large agricultural fields or water bodies
• Spatial filter local operation
Low pass filter (Fig 7.16b)
Passing a moving window throughout the original image
High pass filter (Fig 7.16c)
Subtract a low pass filtered image from the original, unprocessed image
Spatial feature manipulation (cont.)
Convolution
• The generic image processing operation
Spatial filter convolution
• Procedure
Establish a moving window (operators/kernels)
Moving the window throughout the original image
• Example
Fig 7.17
(a) Kernel
Size: odd number of pixels (3x3, 5x5, 7x7, …)
Can have different weighting scheme (Gaussian distribution, …)
(b) original image DN
(c) convolved image DN
Pixels around border cannot be convolved
Spatial feature manipulation (cont.)
Edge enhancement
• Typical procedures
Roughness kernel size
Rough small
Smooth large
Add back a fraction of gray level to the high frequency component image
High frequency exaggerate local contrast but lose low frequency brightness information
Contrast stretching
• Directional first differencing
Determine the first derivative of gray levels with respect to a given direction
Normally add the display value median back to keep all positive values
Contrast stretching
Example
Fig 7.20a: original image
Fig 7.20b: horizontal first difference image
Fig 7.20c: vertical first difference image
Fig 7.20d: diagonal first difference image
Fig 7.21: cross-diagonal first difference image highlight all edges
Spatial feature manipulation (cont.)
Fourier analysis
• Spatial domain frequency domain
• Fourier transform
Quantitative description
Conceptual description
Fit a continuous function through the discrete DN values if they were plotted along each
row and column in an image
The “peaks and valleys” along any given row or column can be described mathematically
by a combination of sine and cosine waves with various amplitudes, frequencies, and
phases
• Fourier spectrum
Fig 7.22
Low frequency center
High frequency outward
Vertical aligned features horizontal components
Horizontal aligned features vertical components
Spatial feature manipulation (cont.)
Fourier analysis (cont.)
• Inverse Fourier transform
Spatial filtering (Fig 7.23)
Noise elimination (Fig 7.24)
Noise pattern vertical band of frequencies wedge block filter
• Summary
Most image processing spatial domain
Frequency domain (e.g. Fourier transform) complicate
and computational expensive
Multi-image manipulation
Spectral ratioing
• DNi / DNj
• Advantage
Convey the spectral or color characteristics of image features,
regardless of variations in scene illumination conditions
Fig 7.25
deciduous trees coniferous trees
Sunlit side shadowed side
Example: NIR/Red stressed and nonstressed vegetation quantify
relative vegetation greenness and biomass
• Number of ratio combination: Cn2
Landsat MSS: 12
Landsat TM or ETM+: 30
Multi-image manipulation (cont.)
Spectral ratioing (cont.)
• Fig 7.26: ratioed images derived from Landsat TM data
(a) TM1/TM2: highly correlated low contrast
(b) TM3/TM4:
Red: road, water lighter tone
NIR: vegetation darker tone
(c) TM5/TM2:
Green and MIR: vegetation lighter tone
But some vegetation looks dark discriminate vegetation type
(d) TM3/TM7
Red: road, water lighter tone
MIR: low but varies with water turbidity water turbidity
• False color composites twofold advantage
Too many combination difficult to choose
Landsat MSS: C(4, 2)/2 = 6, C(6, 3) = 20
Landsat TM: C(6, 2)/2 = 15, C(15, 3) = 455
Optimum index factor (OIF)
Variance && correlation OIF
Best OIF for conveying the overall information in a scene may not be the best OIF for conveying the
specific information need some trial and error
Multi-image manipulation (cont.)
Spectral ratioing (cont.)
• Intensity blind troublesome
Hybrid color ratio composite: one ratio + another band
• Noise removal is an important prelude
Spectral ratioing enhances noise patterns
• Avoid mathematically blow up the ratio
DN΄ = R arctan(DNx/DNy)
arctan ranges from 0 to 1.571. Typical value of R is chosen to be 162.3
DN΄ranges from 0 to 255
Multi-image manipulation (cont.)
Principal and canonical components
• Two techniques
Reduce redundancy in multispectral data
Extensive interband correlation problem (Fig 7.49)
Prior to visual interpretation or classification
• Example: Fig 7.27
DNI = a11DNA + a12DNB
DNII = a21DNA + a22DNB
Eigenvectors (principal components)
The first principal component (PC1) the greatest variance
• Example: Fig 7.28 Fig 7.29 (principal component)
(A) alluvial material in a dry stream valley
(B) flat-lying quanternary and tertiary basalts
(C) granite and granodiorite intrusion
Multi-image manipulation (cont.)
Principal and canonical components (cont.)
• Intrinsic dimensionality (ID)
Landsat MSS: PC1+PC2 explain 99.4% variance ID = 2
PC4 depicts little more than system noise
PC2 and PC3 illustrate certain features that were obscured by the more
dominant patterns shown in PC1
Semicircular feature in the upper right portion
• Principal Canonical
Little prior information concerning a scene is available Principal
Information about particular features of interest is known Canonical
Fig 7.27b
Three different analyst-defined feature types (D, □, +)
Axes I and II maximize the separability of these classes and minimize the variance
within each class
Fig 7.30: Canonical component analysis
Multi-image manipulation (cont.)
Vegetation components
• AVHRR
VI (vegetation index)
VI Ch2 Ch1
NDVI (normalized difference vegetation index)
• Landsat MSS
NDVI
Ch2 Ch1
Ch2 Ch1
Tasseled cap transformation (Fig 7.31)
Brightness soil reflectance
Greenness amount of green vegetation
Wetness canopy and soil moisture
TVI (transformed vegetation index)
DN NIR DN red
TVI
0.5
DN NIR DN red
1/ 2
Fig 7.32, Fig 5.8, Plate 14
100
TVI green biomass
Precision crop management, precision farming, irrigation water, fertilizers, herbicides, ranch
management, estimation of forage, …
GNDVI (green normalized difference vegetation index)
Same formulation as NDVI, except the green band is substituted for the red band
Leaf chlorophyll levels, leaf area index values, the photosynthetically active radiation absorbed by a
crop canopy
• MODIS
EVI (enhanced vegetation index)
Multi-image manipulation (cont.)
Intensity-Hue-Saturation color space transform
• Fig 7.33: RGB color cube
28 28 28 =16,777,216
Gray line
True color composite (B, G, R) false color composite (G, R, NIR)
• Fig 7.34: Planar projection of the RGB color cube
• Fig 7.35: Hexcone color model (RGB IHS)
Intensity
Hue
Saturation
• Fig 7.36: advantage of HIS transform
• Data fusion: Plate 19 (merger of IKONOS data)
1m panchromatic I΄
4m multispectral RGB HIS
Histogram matching I and I΄
I΄HS R΄G΄B΄
Tutorial: mosaicking images
Mosaicking (鑲嵌)
• The art of combining multiple images into a single
composite image
No-georeferenced images
Georeferenced images
• Feathering
Edge feathering
The edge is blended using a linear ramp that averages the two images across the specified
distance
Specified distance = XX pixels, top image = XX%, bottom image = XX%
Cutline feathering
The annotation file must contain a polyline defining the cutline that is drawn from edgeto-edge and a symbol placed in the region of the image that will be cut off.
Tutorial: mosaicking images (cont.)
Pixel-Based Mosaicking
• Map → Mosaicking → Pixel Based
Pixel Based Mosaic dialog
Import → Import Files
avmosaic directory
File: dv06_2.img.
Mosaic Input Files dialog
File: dv06_3.img.
Mosaic Input Files dialog, hold down the Shift key and click on the dv06_2.img and
dv06_3.img filenames to select them.
• Select Mosaic Size dialog
X Size: 614
Y Size: 1024
Pixel Based Mosaic dialog, click on the dv06_3.img filename.
YO: 513
File → Apply
• Create a virtual mosaic
File → Save Template
Output Mosaic Template
• Display the mosaicked image
Tutorial: mosaicking images (cont.)
Pixel-Based Mosaicking (cont.)
• Positioning two images into a composite mosaic image
Options→Change Mosaic Size
Select Mosaic Size dialog
X Size 768
Y Size 768
Left-click within the green graphic outline of image #2
Drag the #2 image to the lower right hand corner of the diagram.
Right-click within the red graphics outline of image #3 and select Edit Entry
Data Value to Ignore: 0
Feathering Distance: 25
Repeat the previous two steps for the other image.
File → Save Template
Load Band
No feathering is performed when using virtual mosaic.
File → Apply
Background Value of 255
Display
Compare the virtual mosaic and the feathered mosaic using image linking and dynamic overlays
Tutorial: mosaicking images (cont.)
Map Based Mosaicking
• Map → Mosaicking → Georeferenced
File → Restore Template
File: lch_a.mos
• Optionally Input and Position Images
Images will automatically be placed in their correct geographic locations The location and
size of the georeferenced images will determine the size of the output mosaic.
• View the Top Image, Cutline and Virtual, Non-Feathered Mosaic
Load Band: lch_01w.img
Right-click to display the shortcut menu and select Toggle → Display Scroll Bars to turn
on scroll bars
Overlay → Annotation
File → Restore Annotation
File: lch_01w.ann
Load Band: lch_02w.img
• File → Open Image File
File: lch_a.mos
• Create the Output Feathered Mosaic
File → Apply
• Compare
Tutorial: mosaicking images (cont.)
Color Balancing During Mosaicking
• Create the Mosaic Image without Color Balancing
Map → Mosaicking → Georeferenced
Import → Import Files
Open File: avmosaic directory, File: mosaic1_equal.dat
Open File: avmosaic directory, File: mosaic_2.dat
select the mosaic_2.dat file, then hold down the Shift key and select the
mosaic1_equal.dat file
Show RGB color composites of these multispectral images
Edit Entry
Mosaic Display, choose RGB.
For Red choose 1, for Green choose 2, and for Blue choose 3
Repeat
Two images are stretched independently
Tutorial: mosaicking images (cont.)
Color Balancing During Mosaicking (cont.)
• Output the Mosaic Without Color Balancing
File → Apply
The seams between the two images are quite obvious
• Output the Mosaic With Color Balancing
mosaic1_equal.dat
Edit Entry.
Color Balancing: Adjust.
mosaic_2.dat
Edit Entry.
Color Balancing: Fixed
File → Apply.
Color Balance using
stats from overlapping regions/
stats from complete files
Display
The seams between the two images are much less visible
Tutorial: Data fusion
Data Fusion
• The process of combining multiple image layers into a single
composite image
• Enhance the spatial resolution of multispectral datasets using
higher spatial resolution panchromatic data or singleband
SAR data.
Landsat TM and SPOT data fusion
• File → Open External File → IP Software → ER Mapper
Subdirectory: lontmsp
File: lon_tm.ers
Load RGB to display a true-color Landsat TM image
• File → Open External File → IP Software → ER Mapper
Subdirectory: lontmsp
File: lon_spot.ers
Load Band to display the gray scale SPOT image
Tutorial: Data fusion (cont.)
Landsat TM and SPOT data fusion (cont.)
• Resize Images to Same Pixel Size
Check spatial dimensions (2820 x 1569) and (1007 x 560)
The Landsat data: 28 meters
The SPOT data: 10 meters
The Landsat image has to be resized by a factor of 2.8 to create 10 m data that matches the SPOT data
Basic Tools → Resize Data (Spatial/Spectral)
choose the lon_tm image
Resize Data Parameters
Enter a value of 2.8 into the xfac text box
Enter a value of 2.8009 into the yfac text box
Tools → Link → Link Displays
• Perform Manual HSI Data Fusion
• Forward HSV Transform
Transform → Color Transforms → RGB to HSV
Select the resized TM data as the RGB image from the Display
Display the Hue, Saturation, and Value images as gray scale images or an RGB.
Create a Stretched SPOT Image to Replace TM Band Value
Basic Tools → Stretch Data
File: lon_spot
Data Stretching
Output Data: 0 for the Min and 1.0 for the Max
Tutorial: Data fusion (cont.)
Landsat TM and SPOT data fusion (cont.)
• Inverse HSV Transform
Transform → Color Transforms → HSV to RGB
Select the transformed TM Hue and Saturation bands as the H and S bands
Choose the stretched SPOT data as the V band
Display Results
• ENVI Automated HSV Fusion
Transform → Image Sharpening → HSV from the ENVI main menu.
Select Input RGB Input Bands dialog
Choose the TM image RGB bands
High Resolution Input File dialog
Choose the SPOT image
HSV Sharpening Parameters dialog
File: lontmsp.img
Display Results, Link and Compare
• Color Normalized (Brovey) Transform
Try the same process using
Transform → Image Sharpening → Color Normalized (Brovey)
Tutorial: Data fusion (cont.)
SPOT PAN and XS fusion
• File → Open Image File
Subdirectory: brestsp
File: s_0417_2.bil
Load RGB to display a falsecolor infrared SPOT-XS image with 20 m spatial resolution
• File → Open Image File
File: s_0417_1.bil
Load Band to display the SPOT Panchromatic data.
• Resize Images to Same Pixel Size
Check spatial dimensions (2835 x 2227) and (1418 x 1114)
The SPOT-XS image has to be resized by a factor of 2.0
Basic Tools → Resize Data (Spatial/Spectral)
Choose the SPOTXS image (s_0417_2.bil)
Resize Data Parameters dialog
Enter a value of 1.999 into the xfac
Enter a value of 1.999 into the yfac
Tools → Link → Link Displays
• Fuse Using ENVI Methods
Transform → Image Sharpening → HSV
Select Input RGB Input Bands
High Resolution Input File
HSV Sharpening Parameters dialog
Display and Compare Results
Tutorial: Data fusion (cont.)
Landsat TM and SAR Data Fusion
• Read and Display Images
File → Open Image File
Subdirectory: rometm_ers
File: rome_ers2
Load Band
File → Open Image File
File: rome_tm
Load RGB to display a false-color infrared Landsat TM image with 30m spatial resolution
• Register the TM images to the ERS image
Map → Registration → Select GCPs: Image-to-Image
Base Image: Display #1 (the ERS data)
Warp Image: Display #2 (the TM data)
File → Restore GCPs from ASCII
Ground Control Points Selection dialog
GCP file: rome_tm.pts
Options → Warp File
File: rome_tm
Registration Parameters dialog
Change Output Parameters
Enter 1 for the Upper Left Corner (XO),
Enter 1 for the Upper Left Corner (YO)
Enter 5134 for the Number of Samples
Enter 5549 for the Number of Lines
Tutorial: Data fusion (cont.)
Landsat TM and SAR Data Fusion
(cont.)
• Perform HSI Transform to Fuse Data
Transform → Image Sharpening → HSV
Select Input RGB Input Bands
High Resolution Input File dialog
Choose the ERS-2 image
• Display and Compare Results