Remote Sensing Lec 10
- 1. LSGI332 REMOTE SENSING CONTRAST ENHANCEMENT LILLESAND PP499-509, MATHER PP.100-109 1
- 2. CONTRAST ENHANCEMENT: DEFINITION • Techniques for increasing the visual distinction between features in a scene • Done by Spectral feature manipulation (as opposed to Spatial, which uses filters) • Consists of producing the ‘best’ image for a particular application • Comprises contrast stretching and level slicing 2
- 3. WHY IS CONTRAST ENHANCEMENT NECESSARY? • Image display and recording devices operate over a range of 0-255 grey levels • Sensor data in a single scene rarely extend over this whole range • Thus necessary to expand narrow range of brightness levels in any one scene to take advantage of the whole range of 0-255 display levels available • This maximizes contrast between features 3
- 4. NEEDS FOR CONTRAST ENHANCEMENT 4 Bob Warwick, University of Leicester
- 5. CONFIGURATION OF IMAGE MEMORY 5 Each layer of video memory can represent one of the 3 primary colours with 256 levels of intensity The lookup table (LUT) allows the intensity of each colour layer to be manipulated by the user but this does not affect the data file values stored on disk DAC= digital-to-analog converter
- 6. GRAPHICAL REPRESENTATION OF IMAGE HISTOGRAM: RAW DATA IMAGE 6 SPOT XS Band 3 (NIR)
- 7. 7 TYPES OF CONTRAST STRETCH Lillesand & Keifer Fig 7.13
- 8. LINEAR CONTRAST STRETCH • Translates the image pixel values from the observed range (Dmin to Dmax), scaling Dmin to 0 and Dmax to 255 • Intermediate values retain their relative positions so that e.g. median input pixel maps to 127 8 Algorithm for linear stretch: creates a Lookup Table
- 9. 9 LOOK-UP TABLE • A lookup table is produced which relates input to output values for the stretch • Efficient since each possible value for output DN computed only once • The image re-scaling is done in virtual display memory and does not affect data on disk
- 10. LOOKUP TABLE 10
- 11. GRAPHICAL REPRESENTATION OF LOOKUP TABLE 11
- 12. UNENHANCED IMAGE AND PLOT OF IMAGE HISTOGRAM 12
- 13. IMAGE AFTER A 0-100% (DEFAULT) LINEAR STRETCH 13
- 14. IMAGE AFTER A 5-95% LINEAR CONTRAST STRETCH 14
- 15. DISADVANTAGE OF LINEAR STRETCH - It assigns as many output display values to rarely occurring input values as it does to frequently occurring ones e.g. Refer to slide 5 image histogram • DN values 60-108 represent few pixels in image but linear stretch allocates DNvalues of 0-127 (half the output range) • Most of pixels in image confined to only half output range 15
- 16. HISTOGRAM EQUALISATION STRETCH 16 • Image values assigned to display levels based on frequency of occurrence in image • Refer to slide 5 – image range 109-158 now stretched over large portion of display levels (39-255) • Smaller portion (0-38) reserved for infrequently occurring values of 60-108 BUT - Relative brightness in original image not maintained - Number of levels used is reduced - Result may be unsatisfactory compared with linear stretch if few occupied LUT entries in output stretch
- 17. 17 After special linear stretch After histogram equalisation stretch
- 18. CALCULATION OF HISTOGRAM EQUALISATION LOOK-UP-TABLE 18 • Done by calculating target (t) no. of pixels to be placed in each histogram class, i.e. total no. of image pixels / 256 = nt • Next, histogram of input image converted to cumulative form, with no. of pixels in classes 0 to j represented by Cj • C represents cumulation • Histogram classes are labelled 0-255 thus Cj = n0+n1…..+nj • Output level for class j is simply Cj/nt where t = target This creates a LUT for ease of computation
- 19. RESULT OF HISTOGRAM EQUALISATION STRETCH 19 eg. 132,385/16384 = 8.1 (Cj/nt)
- 20. HISTOGRAM EQUALISATION IN ACTION Original Image Final Image Equalised Histogram Original Histogram Bob Warwick, University of Leicester
- 21. 21 Gaussian stretch • Fits observed histogram to a normal (gaussian) form • Normal distribution gives probability of observing a value if Mean and Standard Deviation are known - e.g. assume no. of pixels = 262,144 no. of quantisation levels = 16 - Then target no of pixels in each class for normal distribution = probability * 262144 (column iv) - Then cumulative no. of pixels at each level is calculated - Output pixel value determined by comparing v & vii - Once value of column viii determined, written to LUT
- 23. 23 I. Original pixel value II. Calculate the standard deviation above or below mean III. Calculate the probability of each class IV. Target number of pixel of each class, e.g. probability * total number of pixel V. Cumulative the target number of pixel VI. Observed number of pixels VII. Cumulative observed number of pixels VIII. New pixel value • Example: Original 0 class, Cumulative observed number of pixels is 1311, it is exceeded Cumulative the target number of pixel (530), so it is not belonged to class 0, the new class is class 1Probability eq: Mather 1999
- 24. RAW DATA BAND 3 – OUTPUT LIMITS SET TO INPUT LIMITS 24
- 25. LINEAR STRETCH OF ACTUAL DATA RANGE 25 SPOT XS Band 3 (NIR)
- 26. RESULT OF HISTOGRAM EQUALISATION 26 SPOT XS Band 3 (NIR)
- 27. RESULT OF GAUSSIAN TRANSFORM 27 SPOT XS Band 3 (NIR)
- 28. 28 SPECIAL STRETCH Specific features may be analysed in greater radiometric detail by assigning the display range exclusively to a particular range of image values. 28 For example, if water features were represented by a narrow range of values in a scene, characteristics in the water features could be enhanced by stretching this small range to the full display range
- 29. 29 EXAMPLE OF SPECIAL STRETCHING TO ENHANCE DETAIL IN SEA (DARK AREA)
- 30. 30 GREY LEVEL THRESHOLDING/ MASKING Segment an input image into two classes: One for those pixels having values below an analyst-defined grey level and Other for those above this value Normally used for creating binary bit mask
- 31. DENSITY SLICING • The mapping of a range of contiguous grey levels of a single band to a single level and colour • Each range of levels called a ‘slice’ • Range of 0-255 normally converted to several slices • Effective for highlighting different but homogeneous areas within image, but at expense of loss of detail • Effective if slice boundaries/colours carefully chosen 31 ‘Random colour selections may say more about the psychology of the perpetrator than about the information in the image… and may confuse, rather than enlighten’ (Mather p.110) BUT
- 32. 32 Density Slicing DNs along the x axis of an image histogram are divided into a series of analyst-specified intervals or "slices” DNs falling within a given interval in the input image are then displayed at a single DN in the output image. If six different slices are established, the output image contains only six different gray levels
- 33. 33 DENSITY SLICE FOR SURFACE TEMPERATURE VISUALISATION