ENGN6528 - Computer Vision - Computer Lab - 1 - Solved

The goal of this lab is to help you become familiar with, and practice Python-based or Matlab-based basic image processing operations, binary image analysis, image filtering, template tracking and mathematical morphology.   

This is CLab-1 for ENGN6528. The objective is to help you become familiar with basic image I/O functions in either Matlab or Python. Note, in this lab, you are free to choose either of those two languages; Lab task descriptions are provided for both.  

If you have not used Matlab or Python before, this lab is an opportunity to get you to quickly familiar with basic language usages and relevant libraries for image processing and computer vision. Please note that Python is now increasingly used in computer vision, and we encourage you to practise it in this course.


C-Lab-1 Tasks
 

Task-1:    Matlab (Python) Warm-up. (2 marks):  

Describe (in words where appropriate) the result/function of each of the following commands of your preferred language in report. Please utilize the inbuilt help() command if you are unfamiliar with the these functions.  

Note: Different from Matlab, Python users need to import external libraries by themselves. And we assume you already know some common package abbreviations

(e.g. numpy = np). [You only need to complete one set of questions either in matlab or Python.] (0.2 marks each)

Matlab 
Python 
(1)     a = [2, 3, 4; 5, 2, 200] ; 

(2)     b = a(: ,2); 

(3)     f = randn(400, 1) + 3; 

(4)     g = f(find(f > 0 )) *3; 

(5)     x = zeros (1,100) + 0.45 ; 

(6)     y = 0.5 .* ones(1,length(x) ); 

(7)     z = x + y; 

(8)     a = [1:2: 500]; 

(9)     b = a([end: -2:1]); 

(10)  b(b > 50)=0; 
(1)     a = np.array([[2, 3, 4],[5, 2, 200]]) 

(2)     b = a[: , 1]  

(3)     f = np.random.randn(400,1)+3 

(4)     g = f[ f > 0 ]*3 

(5)     x = np.zeros (100) + 0.45  

(6)     y = 0.5 * np.ones([1, len(x)]) 

(7)     z = x + y 

(8)     a = np.linspace(1,499, 250, dtype=int) 

(9)     b = a[: :-2] 

(10)  b[b >50]=0 
 

Hint:  

Do the necessary typecasting (uint8 and double) when processing and displaying the image data in the following tasks. For Python, please be aware of the default datatype of different libraries (e.g. Image, matplotlib, cv2). An improper datatype of image will cause many troubles when you want to display the image.

Task-2:    Basic Coding Practice (1 marks)
Write functions to process an input grayscale image with following requirements, where you need to write a script to load the given image in the Lab package, apply each transformation to the input, and display the results in a figure. For Matlab, you can use subplot() function; for Python, we suggest to use matplotlib.pyplot.subplot(). Please notice that each subplot needs to be labelled with an appropriate title. (0.2 marks each)

1.      Load a grayscale image (such as Atowergray.jpg from wattle), and map the image to its negative image, in which the lightest values appear dark and vice versa. Display it side by side with its original version.  

2.      Flip the image horizontally (i.e, map pixels from right to left changed from left to right). 

3.      Load a colour image, swap the red and green colour channels of the input.   

4.      Average the input image with its horizontally flipped image (use typecasting).

5.      Add a random value between [0,127] to every pixel in the grayscale image, then clip the new image to have a minimum value of 0 and a maximum value of 255. (Note: the intensity values of the original grayscale image range from 0 to 255.) 

 

 

Task-3:    Basic Image I/O (2 marks)
Note: You need to download the image1.jpg, image2.jpg, image3.jpg from wattle.

In this task, you are asked to:   

1. Using image1.jpg, develop short computer code that does the following tasks:

a.       Read this image from its JPG file, and resize the image to 384 x 256 in columns x rows (0.2 marks).

b.      Convert the colour image into three grayscale channels, i.e., R,G,B images, and display each of the three channel grayscale images separately (0.2 marks for each channel, 0.6 marks in total).

c.       Compute the histograms for each of the grayscale images, and display the 3

histograms (0.2 marks for each histogram, 0.6 marks in total).

d.      Apply histogram equalisation to the resized image and its three grayscale channels, and then display the 4 histogram equalization images (0.15 marks for each histogram, 0.6 marks in total). (Hint: you can use inbuilt functions for implementing histogram equalisation. e.g. histeq() in Matlab or cv2.equalizeHist() in Python).

 

  

Fig, 2 (a) Colour Wheel                                         (b) Palettes

Task-4:    Colour space conversion (3 marks)
Use the two images in Fig.2 to study colour space conversion from RGB to YUV (you can download them from the Wattle site):

1.      Based on the formulation of RGB-to-YUV conversion, write your own function cvRGB2YUV()that converts the RGB image to YUV colour space (0.7 marks). Read in Fig.2(a) and convert it with your function, and then display the Y, U, V channels in your report (0.1 marks for each channel, 0.3 marks in total).

2.      Compute the average Y values of five colour regions in Fig. 2(b) with your function and the Matlab’s inbuilt function rgb2yuv(). Print both of them under the corresponding regions (0.1 marks for each value, 0.5 marks in total). You also need to explain how to distinguish and divide the five regions, and how to calculate the average Y value (1.5 marks, higher marks only for a smarter solution).

(Note: The elements of both colormaps are in the range 0 to 1.).

 

Task-5:    Image Denoising via a Gaussian and Bilateral Filter (9 marks)
In this task, you are asked to:   

1.      Read in image2.jpg. Crop a square image region corresponding to the central part of the image,  resize it to 512×512,  and save this square region to a new grayscale image.  Please display the two images.  Make sure the pixel value range of this new image is within [0, 255]. Add Gaussian noise to this new 512x512 image (Review how you generate random number in Task-1). Use Gaussian noise with zero mean, and standard deviation of 15.

Hint: Make sure your input image range is within [0, 255]. Kindly, you may need np.random.randn() in Python. While Matlab provides a convenient function imnoise(). Please check the default setting of these inbuilt function.

Display the two histograms side by side, one before adding the noise and one after adding the noise (0.5 marks).  

2.      Implement your own Matlab/python function that performs a 5x5 Gaussian filtering (1.5 marks). Your function interface is:  my_Gauss_filter() 

                        input: noisy_image, my 5x5 gausskernel                    output: output_image

3.      Apply your Gaussian filter to the above noisy image, and display the smoothed images and visually check their noise-removal effects, investigating the effect of modifying the standard deviation of the Gaussian filter. You may need to test and compare different Gaussian kernels with different standard deviations (0.5 marks).

Note: In doing this task, and the bilateral filter below you MUST NOT use any Matlab’s (or Python’s) inbuilt image filtering functions (e.g. imfilter(), filter2() in Matlab, or cv2.filter2D() in Python). In other words, you are required to code your own 2D filtering code, based on the original mathematical definition for 2D convolution. However, you are allowed to generate a 5x5 sized Gaussian kernel with inbuilt functions.  

4.      Compare your result with that by Matlab’s inbuilt 5x5 Gaussian filter (e.g. filter2(), imfilter() in Matlab, or conveniently cv2.GaussianBlur() in Python,). Please show that the two results are nearly identical (0.5 marks).

Further reading material: http://setosa.io/ev/image-kernels/

Image Denoising via a Bilateral (Challenge task: this task will be difficult for most students to complete.) (5 marks)
1.      Using your Gaussian filter as a base, implement your own Matlab/Python function that performs a 5x5 Bilateral filtering to gray-scale image. (1.5 marks) Your function interface must be:  

my_Bilateral_filter() 

                        input: noisy_image, my_5x5_gausskernel, colour_sigma                  output: output_image 

            where colour_sigma is the sigma applied to the intensity/gray scale part of the filter.

2.      Apply your Bilateral filter to the above noisy image (greyscale version) from the last task, and display the smoothed images and visually check their noiseremoval and bilateral edge preserving effects (0.5 marks in total).

In addition to the Gaussian filter, the range filter also has a standard deviation. You may need to test and compare different standard deviations for range (1.0 marks).

Note: Do not use in-built functions, the same as for task 4.

3.      Extend the Bilateral filter to colour images (eg. The color version of the previous grayscale image.). For this you may need to consider the CIE-Lab colour space as described in the paper. You will need to explore this for yourself. Namely, You need to generate the noisy color image and implement the bilateral filter to the color image (CIE-Lab). (2.0 marks) (Note this task is more difficult again).  

Tomasi, C; Manduchi, R (1998). Bilateral filtering for gray and color images (PDF). Sixth International Conference on Computer Vision. Bombay. pp. 839– 846. 

4.      In up to half a page, discuss the impact of smoothing on colour. What are the difficulties of smoothing in RGB colour space? Why is CIE-Lab space a good idea for this smoothing? (Compare images smoothed in CIE-Lab space vs RGB) Does the bilateral filter itself help? You may want to compare Gaussian smoothed colour images and bilateral filtered ones, and investigate filtering in different colour spaces. For this you can use your processed images as examples, possibly cropping and enhancing detail in your report to illustrate your discussion. (1 marks) 

Task-6:    Image Translation (3 marks)
Choose an image among (image1.jpg, image2.jpg, image3.jpg) and resize it to 512 x 512,

1.    Implement your own function my_translation() for image translation by any given number of pixels between [-100, +100], in both x and y. Note that this can be a real number (partial pixels). Display images translated by (2.0,4.0),(4.0,-6.0),(2.5, 4.5), (-0.9,1.7), (92.0,-91.0) (0.20 for each image, 1.0 mark in total).  

Note: positive for away from upwards and right from the bottom left hand corner.

2.    Compare forward and backward mapping and analyze their difference (1.0 mark).

3.    Compare different interpolation methods and analyze their difference (1.0 mark).  

Hint: When analyzing the difference, you can focus on: (1) visual results; (2) the principles in terms of formulation or others relevant; (3) advantages and drawbacks; (4) the computational complexity. You can also think about it from other aspects.