530 likes | 714 Views
Lecture 5 Dr. Roger S. Gaborski. Fundamentals of Computer Vision. Quiz 1. In Class Exercise. Intensity image is simply a matrix of numbers. We can summary this information by only retaining the distribution if gray level values:. PARTIAL IMAGE INFO:.
E N D
Lecture 5 Dr. Roger S. Gaborski Fundamentals of Computer Vision Roger S. Gaborski
Quiz 1 Roger S. Gaborski
In Class Exercise Roger S. Gaborski
Intensity image is simply a matrix of numbers We can summary this information by only retaining the distribution if gray level values: PARTIAL IMAGE INFO: 117 83 59 59 68 77 84 94 82 67 62 70 83 86 85 81 71 65 77 89 86 82 76 67 72 90 97 86 66 54 68 104 121 107 85 46 58 89 138 165 137 91 38 80 147 200 211 187 138 40 80 149 197 202 187 146 56 76 114 159 181 160 113 An image shows the spatial distribution of gray level values Roger S. Gaborski
Image Histogram Plot of Pixel Count as a Function of Gray Level Value Pixel Count Gray Level Value Roger S. Gaborski
Histogram • Histogram consists of • Peaks: high concentration of gray level values • Valleys: low concentration • Flat regions Roger S. Gaborski
Formally, Image Histograms Histogram: • Digital image • L possible intensity levels in range [0,G] • Defined: h(rk) = nk • Where rk is the kth intensity level in the interval [0,G] and nk is the number of pixels in the image whose level is rk . • G: uint8 255 uint16 65535 double 1.0 Roger S. Gaborski
Notation • L levels in range [0, G] • For example: • 0, 1, 2, 3, 4, in this case G = 4, L = 5 • Since we cannot have an index of zero, • In this example, index of: Index 1 maps to gray level 0 2 maps to 1 3 maps to 2 4 maps to 3 5 maps to 4 Roger S. Gaborski
Normalized Histogram • Normalized histogram is obtained by dividing elements of h(rk) by the total number of pixels in the image (n): fork = 1, 2,…, L p(rk) is an estimate of the probability of occurrence of intensity level rk Roger S. Gaborski
MATLAB Histogram • h = imhist( f, b ) • h is the histogram, h(rk) • f is the input image • b is the number of bins (default is 256) • Normalized histogram Roger S. Gaborski
Color and Gray Scale Images Roger S. Gaborski
Background: Gray Image >> I = imread('Flags.jpg'); >> figure, imshow(I) % uint8 >> Im= im2double(I); % convert to double >> Igray = (Im(:,:,1)+Im(:,:,2)+Im(:,:,3))/3; >> figure, imshow(Igray) There is also the rgb2gray function that results in a slightly different image Roger S. Gaborski
Gray Scale Histogram Roger S. Gaborski
Plots • bar(horz, v, width) • v is row vector • points to be plotted • horz is a vector same dimension as v • increments of horizontal scale • omitted axis divided in units 0 to length(v) • width number in [0 1] • 1 bars touch • 0 vertical lines • 0.8 default Roger S. Gaborski
p= imhist(Igray)/numel(Igray); >> h1 = p(1:10:256); >> horz = (1:10:256); >> figure, bar(horz,h1) Review other examples in text and in MATLAB documentation Roger S. Gaborski
Chapter 3 www.prenhall.com/gonzalezwoodseddins Roger S. Gaborski
Color and Gray Scale ImagesRecall from Previous Slide Roger S. Gaborski
Gray Scale Histogram Roger S. Gaborski
Normalized Gray Scale Histogram >> p= imhist(Igray)/numel(Igray); >> figure, plot(p) Roger S. Gaborski
Normalized Gray Scale Histogram 256 bins 32 bins imhist(Igray)/numel(Igray); imhist(Igray,32)/numel(Igray) Roger S. Gaborski
Normalized Gray Scale Histogram >> p= imhist(Igray)/numel(Igray); >> figure, plot(p) probability Gray level values Roger S. Gaborski
Original Dark Light Roger S. Gaborski
Contrast enhancement • How could we transform the pixel values of an image so that they occupy the whole range of values between 0 and 255? Roger S. Gaborski
Gray Scale Transformation • How could we transform the pixel values of an image so that they occupy the whole range of values between 0 and 255? • If they were uniformly distributed between 0 and x we could multiply all the gray level values by 255/x • BUT – what if they are not uniformly distributed?? Roger S. Gaborski
Cumulative Distribution Function Histogram CDF Roger S. Gaborski
Histogram Equalization(HE) • HE generates an image with equally likely intensity values • Transformation function: Cumulative Distribution Function (CDF) • The intensity values in the output image cover the full range, [0 1] • The resulting image has higher dynamic range • The values in the normalized histogram are approximately the probability of occurrence of those values Roger S. Gaborski
Histogram Equalization • Let pr(rj), j = 1, 2, … , L denote the histogram associated with intensity levels of a given image • Values in normalized histogram are approximately equal to the probability of occurrence of each intensity level in image • Equalization transformation is: k = 1,2,…,L sk is intensity value of output rk is input value Sum of probability up to k value Roger S. Gaborski
Histogram Equalization Example • g = histeq(f, nlev) where f is the original image and nlev number of intensity levels in output image Roger S. Gaborski
Original Image INPUT Roger S. Gaborski
Transformation x255 Output Gray Level Value Input Gray Level Value Roger S. Gaborski
Equalization of Original Image OUTPUT Roger S. Gaborski
Simple Histogram Equalization Example Roger S. Gaborski
Histogram Equalization Input Image Output Image Roger S. Gaborski
Adaptive Equalization • g = adapthisteq(f, parameters..) • Contrast-limited adaptive histogram equalization • Process small regions of the image (tiles) individually • Can limit contrast in uniform areas to avoid noise amplification Roger S. Gaborski
>> help adapthisteq adapthisteq Contrast-limited Adaptive Histogram Equalization (CLAHE). adapthisteq enhances the contrast of images by transforming the values in the intensity image I. Unlike HISTEQ, it operates on small data regions (tiles), rather than the entire image. Each tile's contrast is enhanced, so that the histogram of the output region approximately matches the specified histogram. The neighboring tiles are then combined using bilinear interpolation in order to eliminate artificially induced boundaries. The contrast, especially in homogeneous areas, can be limited in order to avoid amplifying the noise which might be present in the image. J = adapthisteq(I) Performs CLAHE on the intensity image I. J = adapthisteq(I,PARAM1,VAL1,PARAM2,VAL2...) sets various parameters. Parameter names can be abbreviated, and case does not matter. Each string parameter is followed by a value as indicated below: 'NumTiles' Two-element vector of positive integers: [M N]. [M N] specifies the number of tile rows and columns. Both M and N must be at least 2. The total number of image tiles is equal to M*N. Default: [8 8]. Roger S. Gaborski
Adaptive Histogram Equalization Default, 8x8 tiles Roger S. Gaborski
Adaptive Equalization Roger S. Gaborski
Chapter 3 www.prenhall.com/gonzalezwoodseddins Roger S. Gaborski
Chapter 3 www.prenhall.com/gonzalezwoodseddins Roger S. Gaborski
Histograms • Histogram is nothing more than mapping the pixels in a 2 dimensional matrix into a vector • Each component in the vector is a bin (range of gray level values) and the corresponding value is the number of pixels with that gray level value
Similarity between Histograms • Similarity between histogram bins: • Assuming both histograms have ∑nj j=1…B pixels M. Swain and D. Ballard. “Color indexing,”International Journal of Computer Vision, 7(1):11–32, 1991.
Histogram Intersection • A simple example: • g = [ 17, 23, 45, 61, 15]; (histogram bins) • h = [ 15, 21, 42, 51, 17]; • in=sum(min(h,g))/min( sum(h),sum(g)) • in = 0.9863
>> g = [17,23,45,61,15]; >> h = [15,21,42,51,17]; >> min(g,h) ans= 15 21 42 51 15 >> N = sum(min(g,h)) N = 144 >> D=min(sum(h),sum(g)) D = 146 >> intersection = N/D intersection = 0.9863 Roger S. Gaborski
If Histograms Identical • g = 15 21 42 51 17 • h = 15 21 42 51 17 • >> in=sum(min(h,g))/min( sum(h),sum(g)) • in = 1
Different Histograms • h = 15 21 42 51 17 • g = 57 83 15 11 1 • >> in=sum(min(h,g))/min( sum(h),sum(g)) • in = 0.4315
Region and Histogram Similarity with itself: >>h = hist(q(:),256); >> g=h; >> in=sum(min(h,g))/min( sum(h),sum(g)) in = 1
>> r=236;c=236; >> g=im(1:r,1:c); >> g= hist(g(:),256); >> in=sum(min(h,g))/min( sum(h),sum(g)) in = 0.5474