COMP 364: Computer Tools for Life Sciences

Transcription

1 COMP 364: Computer Tools for Life Sciences Introduction to image analysis with scikit-image (part one) Christopher J.F. Cameron and Carlos G. Oliver 1 / 27

2 Quiz #9 the penultimate quiz Key course information available Monday, November 27th (closes 11:59:59 pm) covers topics from the last two weeks HW5 available early next week due Thursday, December 7th, 2017 at 11:59:59 pm shorter than previous assignments Course evaluations available now at the following link: WWWLogin?ret_code=f 2 / 27

3 Why perform digital image analysis? Digital image analysis (DIA) The extraction of useful information from images important for good feature desgin emphasizes important traits while diluting noisy ones For example in machine vision, image preprocessing plays a huge role before extracting features from an digital image it s extremely useful to be able to augment it to highlight aspects that are important for the machine learning task to stand out 3 / 27

4 DIA in Python scikit-image module or (skimage) image processing module in Python holds a wide library of image processing algorithms: filters, transforms, point detection API We ll start with an example image using the io module basic I/O submodule of scikit-image API http: //scikit-image.org/docs/dev/api/skimage.io.html 4 / 27

5 5 / 27

6 Reading an image into memory 1 import skimage.io as io 2 3 # read image into memory 4 image = io.imread("./../images/monkey.jpg") 5 # print top-left pixel RGB values 6 print(image[0,0]) 7 # prints: [ ] 8 # write image to disk 9 io.imsave("./../images/monkey_copy.jpg",image) What are RGB values? 6 / 27

7 RGB colors Red green blue (RGB) An RGB color value is specified with: rgb(red, green, blue) Each parameter (red, green, and blue) defines the intensity of the color as an integer between 0 and 255 For example, rgb(0, 0, 255) is rendered as blue because the blue parameter is set to its highest value (255) the others are set to 0 RGB color picker/codes chart: 7 / 27

8 Handling colors Let s make copies of our image and increase intensity for each color intensity red, green, blue note: the format of our image object is image[ #ycoordinate, #xcoordinate, [red green blue]] top-left pixel is [0, 0, [red green blue]] 1 red, green, blue = image.copy(), 2 image.copy(), image.copy() 3 red[:,:,(1,2)] = 0 # NumPy indexing 4 green[:,:,(0,2)] = 0 5 blue[:,:,(0,1)] = 0 6 io.imsave("./../images/monkey_red.jpg",red) 7 io.imsave("./../images/monkey_green.jpg",green) 8 io.imsave("./../images/monkey_blue.jpg",blue) 8 / 27

9 red intensity green intensity blue intensity 9 / 27

10 Grayscaling Most image processing algorithms assume a 2D matrix not an image with a third dimension of color To bring the image into two dimensions we need to summarize the three colors into a single value this process is more commonly know as grayscaling where the resulting image only holds intensities of gray with values between 0 and 1 skimage submodule color has useful functions for this task API color.html 10 / 27

11 1 from skimage.color import rgb2gray 2 3 # read image into memory 4 image = io.imread("./../images/monkey.jpg") 5 # convert to grayscale 6 gray_image = rgb2gray(image) 7 io.imsave("./../images/monkey_grayscale.jpg",gray_image) 8 print(image[0,0]) 9 # prints: [ ] 10 print(gray_image[0,0]) 11 # prints: 1.0 After we view the grayscale image let s find a better way to view the transformation using histograms 11 / 27

12 12 / 27

13 Histogram of RGB intensities 1 import matplotlib.pyplot as plt 2 3 for index,label in zip([0,1,2],["red","green","blue"]): 4 #.flatten() converts 2D list to 1D 5 plt.hist(image[:,:,(index)].flatten(),50 6,label=label,edgecolor="k",linewidth=1, 7 facecolor=label[0],alpha=0.75) 8 plt.xlabel("pixel intensity",size=16) 9 plt.xlim([0,255]) 10 plt.ylabel("frequency",size=16) 11 plt.tight_layout() 12 plt.savefig("./../images/histogram_"+label+".png") 13 plt.close() 13 / 27

14 Red Green Blue 14 / 27

15 15 / 27

16 Image enhancement - histogram equalization Histogram equalization (HE) Take a grayscale image attempt to distribute intensities more evenly along the range of possible values (0 to 1) pixels still rank the same a pixel with a higher value than another will still have a higher value after the transform...but the image as a whole becomes far more contrasted and normalized We ll use the submodule exposure to perform HE API exposure.html 16 / 27

17 HE with Python s skimage module 1 from skimage.exposure import equalize_hist 2 3 gray_image = rgb2gray(image) 4 print(gray_image[0,0]) 5 # prints: equalized_image = equalize_hist(gray_image) 7 print(equalized_image[0,0]) 8 # prints: io.imsave("./../images/monkey_he.jpg",equalized_image) Based on what you have learned about HE why does the top-left most pixel s value not change? 17 / 27

18 Grayscale Histogram equalization Why does the image become more contrasted? pixels that started with similar intensity values which were relatively hard to differentiate are now more distinctly separated Let s look at the histograms for both images 18 / 27

19 1 # plot hist of HE pixel intensities 2 plt.hist(equalized_image[:,:].flatten(),50,label="he", 3 edgecolor="k",linewidth=1,facecolor="blue", 4 alpha=0.75) 5 # plot hist of grayscale pixel intensities 6 plt.hist(gray_image[:,:].flatten(),50, 7 label="grayscale",edgecolor="k",linewidth=1, 8 facecolor="red",alpha=0.75) 9 plt.xlim([0,1]) 10 plt.xlabel("pixel intensity",size=16) 11 plt.ylabel("frequency",size=16) 12 plt.legend(loc="best") 13 plt.tight_layout() 14 plt.savefig("./../images/histogram_he.png") 15 plt.close() 19 / 27

20 20 / 27

21 Image enhancement - binarizing and blurring Sometimes, it helps to simplify an image even further instead of grayscale, binarize the image results in each pixel hold only one of two values more commonly recognized as a pure black and white image The objective is to separate the foreground from the background to make feature generation even easier A simple way of doing this is to just choose a threshold every pixel above that threshold is set to 1 every pixel below it to 0 21 / 27

22 Binarizing and blurring In our case, we ll select the mean of our grayscale image as the threshold every pixel above the mean is set to white (1.0) those below are set to black (0.0) 1 import numpy as np 2 3 gray_image = rgb2gray(image) 4 #print(gray_image[0,0]) 5 # prints: binary_image = np.where(gray_image > np.mean(gray_image) 7,1.0,0.0) 8 io.imsave("./../images/monkey_binary.jpg",binary_image) 9 print(binary_image[0,0]) 10 # prints: / 27

23 23 / 27

24 Image enhancement - blurring/smoothing Binary images may capture more detail than is helpful for example, the objective is to identify prominent features of the image monkey s hands and fur, foliage etc. the position for every piece of fur (or leaf) isn t necessary blurring/smoothing the image is a reasonable alternative Scikit-image s Gaussian filter (filter submodule) takes a weighted average of surrounding pixels so individual pixels incorporate local intensities into their own this produces a pretty recognizable blur/smoothing effect 24 / 27

25 1 from skimage.filters import gaussian 2 3 equalized_image = equalize_hist(gray_image) 4 for sigma,name in zip([3,6],["blurred","really_blurred"]): 5 blurred_image = gaussian(equalized_image,sigma=sigma) 6 fig,ax = plt.subplots() 7 ax.imshow(blurred_image,vmin=0, vmax=1) 8 # remove ticks 9 ax.set_xticks([]) 10 ax.set_yticks([]) 11 # remove spines 12 for spine in ["top","bottom","right","left"]: 13 ax.spines[spine].set_visible(false) 14 plt.savefig("./../images/monkey_"+name+".jpg") 15 plt.close() 25 / 27

26 Sigma = 3 Sigma = 6 The equalized image (rather than grayscale) has been used maintains a high level of contrast through the filtering as sigma increases, so does the blurring 26 / 27

27 More digital image analyses! edge and corner detection maybe more? Next time in COMP / 27