Image Processing in Python

Tanmoy Dasgupta

thetdg@live.com | Assistant Professor | Department of Electrical Engineering | Techno India University, Kolkata

I colud not get the sepll checekr wroikng. So this ntoebook mgiht cnotain erorrs / toyps.

This tutorial is supposed to be an introduction to the different scientific packages in python that can be utilized to differnet tasks related to image processing. I have inherently assumed that the reader already has good exposure on core Python and some exposure on Numpy, Scipy and Matplotlib. It is also assumed that the reader knows how to start IPython Notebooks.

This notebook contains materails that are useful and new. However, things that are useful are not new and the things that are new are not always useful. Feel free to improve it and send me suggestions.

Packages You Need

Python 2.7 or Python 3.4

Choose Python 2.7 is you want to settle with the past. Choose 3.4 if you want to go with the future.

NumPy

NumPy is the fundamental package for scientific computing with Python. It contains among other things:

  1. a powerful N-dimensional array object
  2. sophisticated (broadcasting) functions
  3. tools for integrating C/C++ and Fortran code
  4. useful linear algebra, Fourier transform, and random number capabilities

SciPy

The SciPy library is one of the core packages that make up the SciPy stack. It provides many user-friendly and efficient numerical routines such as routines for numerical integration and optimization.

Matplotlib

Matplotlib is a python 2D and 3D plotting library which produces publication quality figures in a variety of hardcopy formats and interactive environments across platforms. matplotlib can be used in python scripts, the python and ipython shell (ala MATLAB® or Mathematica®), web application servers, and different graphical user interface toolkits.

IPython

IPython provides a rich architecture for interactive computing with:

  1. Powerful interactive shells (terminal and Qt-based).
  2. A browser-based notebook with support for code, rich text, mathematical expressions, inline plots and other rich media.
  3. Support for interactive data visualization and use of GUI toolkits.
  4. Flexible, embeddable interpreters to load into your own projects.
  5. Easy to use, high performance tools for parallel computing.

Python Imaging Library and Scikit-Image

These packages (among many others) have custom modules for image processing.

OpenCV

OpenCV (Open Source Computer Vision Library) is an open source computer vision and machine learning software library. OpenCV was built to provide a common infrastructure for computer vision applications and to accelerate the use of machine perception in the commercial products. The library has more than 2500 optimized algorithms, which includes a comprehensive set of both classic and state-of-the-art computer vision and machine learning algorithms. It has C++, C, Python, Java and MATLAB interfaces and supports Windows, Linux, Android and Mac OS.

Installing

Installing all these packages might seem a little too much! Don't worry. You can install all of the above (and many other packages) just by downloading and installing any of the following Python Distributions :

Continuum Anaconda Free (as in 'Freedom'). Both Python 2.7 and 3.4 are available. Availabe for Linux, Mac and Windows.

Enthought Canopy Free + Subscription based. Only Python 2.7. Availabe for Linux, Mac and Windows. Full featured product subscription is freely availabe for Academic use.

WinPython Portable. Free. Both Python 2.7 and 3.4 are available. Windows only.

Getting Started

Unlike almost all other IPython notebooks, I will NOT import Pylab Magic. But if you really want to, you can do it by uncommenting (removing the #) and running the following line.


In [ ]:
%pylab inline
from __future__ import division #Python 2.X and 3.X Compatibility
from __future__ import print_function #Python 2.X and 3.X Compatibility
import numpy as np
import scipy as sp
import matplotlib.pyplot as plt

In case you didn't have any error so far, you're good to go!

What is an image anyway?

Short answer It is just a 2 dimensional array (grayscale image) or a set of Three 2 dimensional arrays (colour image).

Long answer

Loading an Image as a Numpy array


In [ ]:
A = plt.imread('macaw.jpg')
print(A)

Check the data type and the size of the array $A$


In [ ]:
print(np.shape(A))
print(type(A))
print(A.dtype)

So, our image is a colour image and it has a resolution of $400 \times 267$. It is imported as an N-dimenstional array object available from numpy.


In [ ]:
plt.imshow(A)
plt.show()

Now, let us segrigate the Red, Green and the Blue channels


In [ ]:
A_red = A[:, :, 0]
A_green = A[:, :, 1]
A_blue = A[:, :, 2]

plt.figure()
plt.imshow(A_red, cmap=cm.gray) #For a single channel / grayscale image you need to mention the colourmap
plt.title('The RED channel')

plt.figure()
plt.imshow(A_green, cmap=cm.gray)
plt.title('The GREEN channel')

plt.figure()
plt.imshow(A_blue, cmap=cm.gray)
plt.title('The BLUE channel')

plt.show()

In [ ]:
plt.figure()
plt.imshow(A_red)
plt.show()

Now that we know how to read and display an image. Let us do this:


In [ ]:
#create a random uint8 array of size 300x300 with numbers from 0 to 255
x = np.random.randint(0, 256, (300, 400)).astype('uint8')
plt.imshow(x, cmap=cm.gray)

See! A random image! Now let us create a random color image!


In [ ]:
X = np.random.randint(0, 256, (300, 400, 3)).astype('uint8')
plt.imshow(X)

Image Enhancement Techniques

The principal objective of Image Enhancement is to process an image so that the result is more suitable than the original image for a particular application.

However, there are mainly TWO approaches towards image enhancement. A saptial domain approach and a frequency domain approach.

Spatial Domain Approach

It refers to the image plane itself and involves direct manipuation of the pixels of an image.

Frequency Domain Technique

Frequency domain processing techniques are based on modifying the Fourier Transform of an image.

Spatial Domain Approaches

Image processing functions in the spatial domain are often of the form $ g(x,y) = \mathcal{T}[f(x, y)]$, where, $f(x, y)$ is the input image and $g(x, y)$ is the processed output image. $\mathcal{T}$ is an operation on $f$ defined over some neighbourhood of the pixel at the location $(x, y)$. Usually a neighbourhood of $3\times 3$ (or sometimes $1\times 1$) is assumed about the pixel at $(x, y)$.

Spatial domain techniques include Pint processing, Image subtraction, Spatial filtering, Image averaging, etc.

Point processing include Contrast stretching, Gray-level slicing, Bit-plane slicing, Histogram processing, etc.

Spatial filtering includes Low pass filtering, Median filtering, High-pass filtering, etc.

Histogram and Contrast

The word histogram in the context of an image simply means a histogram plot of the pixel intensity vs number of pixels.

Now, let us go back to the original image of the Macaw. Let us plot the histogram of the pixel intensity. First convert the original image into a grayscale image. This RGB to grayscale conversion would use the formula (more on this formula later) $$X_{Gray} = [0.299\quad 0.587\quad 0.144] \cdot \left[\begin{array}{c} X_{R}\\ X_{G}\\ X_{B} \end{array}\right].$$


In [ ]:
from PIL import Image #Python Imaging Library
A_gray = Image.open('macaw.jpg','r')
A_gray = A_gray.convert('L')

temp = np.asarray(A_gray.getdata(), dtype=np.float64).reshape((A_gray.size[1], A_gray.size[0]))
A_gr = np.asarray(temp, dtype=np.uint8)

In [ ]:
plt.imshow(A_gr, cmap=cm.gray)
plt.show()

In [ ]:
plt.hist(A_gr.flatten(), 256, range=(0, 255), fc='k', ec='k');

Let us now narrow the contrast the image

Contrast Stretching

The possible causes for low contrast images are

  1. poor illumination
  2. lack of dynamic range in imaging sensor
  3. wrong setting of the lens aperture during image acquisition.

Contrast stretching attempts to increase the dynamic range of the gray levels of the image being processed. For a neighbourhood of size $1 \times 1$, contast stretching is usually done by a gray level transformation of the form $s = \mathcal{W}[r]$, where $r$ is the gray level of $f(x, y)$ at $(x, y)$, $s$ is the gray level of $g(x, y)$ at $(x, y)$ and $\mathcal{W}$ is a graylevel transformation function.

Now, let us load a low contrast image.


In [ ]:
girl = plt.imread('low_contrast.jpg')
plt.imshow(girl)

In [ ]:
plt.hist(girl.flatten(), 256, range=(0, 255), fc='k', ec='k');

In [ ]:
maxi = np.amax(girl)
mini = np.amin(girl)
intensity_range = maxi - mini
print('lowest intensity:', mini, ', highest intensity:', maxi, ', spread:', intensity_range)

As we can see that the image has low contrast, we will use the following transformation function to stretch the contrast of the image. $$s=\mathcal{W}(r)=\begin{cases} 0, & r<103\\ \frac{255}{132}(r-103), & 103\le r\le235\\ 255, & r>235 \end{cases}$$


In [ ]:
r = np.arange(0, 256, 1)
s = np.zeros(shape(r))
s = (255/intensity_range)*(r - mini)
s[r<mini] = 0
s[r>maxi] = 255

plt.plot(r, s)
plt.axis([0, 260, -5, 260])
xlabel('r')
ylabel('s')
title('gray-level transformation function $\mathcal{W}$')
plt.grid()

In [ ]:
girl_high = ((girl.astype('float64') - mini) * 255 / intensity_range).astype('uint8')

In [ ]:
plt.imshow(girl_high, cmap=cm.gray)

In [ ]:
plt.hist(girl_high.flatten(), 256, range=(0, 255), fc='k', ec='k');

See that, now the pixel intensities are spread over wider range $[0, 255]$. This is known as linear contrast streatching.

Negative

If you subtract the pixel intensities of the original image from 255, what you get is essentially the negative of the image. $$ X_{neg} = \left[\begin{array}{ccc} 255 & \cdots & 255\\ \vdots & \ddots & \vdots\\ 255 & \cdots & 255 \end{array}\right] _{m\times n} - X_{m\times n}.$$

The transformation function in this case looks like this.


In [ ]:
r = np.arange(0, 256, 1)
s = np.zeros(shape(r))
s = 255 - r

plt.plot(r, s)
plt.axis([0, 260, -5, 260])
xlabel('r')
ylabel('s')
title('$\mathcal{W}$ for finding the negative of an image')
plt.grid()

In [ ]:
girl_neg = (255*np.ones(shape(girl_high)) - girl_high).astype('uint8')

In [ ]:
plt.imshow(girl_neg, cmap=cm.gray)

In [ ]:
plt.hist(girl_neg.flatten(), 256, range=(0, 255), fc='k', ec='k');

Compare the histogram of the negative with that of the original.

Dynamic Range Compression

to be done later

Power Law (Gamma) Transformations / Gamma Corrections

Gamma correction involves a nonlinear transformation of the form $s = \mathcal{W}[r] = 255\,c\,\left(\dfrac{r}{255}\right)^\gamma$, where, $c$ and $\gamma$ are positive constants. The following plot shows a $r$-vs-$s$ plot for different values of $\gamma$.


In [ ]:
c = 1
r = np.arange(0, 256)

for gamma in [0.04, 0.10, 0.20, 0.40, 0.67, 1, 1.5, 2.5, 5, 10, 25]:
    s = 255*c*(r/255)**gamma
    plt.plot(r, s)
plt.axis([0, 255, 0, 255])
plt.xlabel('r')
plt.ylabel('s')
plt.title('Gamma correction $s = 255\,c\,(r/255)^\gamma$')

In [ ]:
xray_orig = plt.imread('chestxray.jpg')
figure(figsize=(20,7))
subplot(1, 2, 1)
plt.imshow(xray_orig, cmap=cm.gray)
plt.title('original')
subplot(1, 2, 2)
plt.hist(xray_orig.flatten(), 256, range=(2, 255), fc='k', ec='k');

c = 1.0
gamma = 2.0
figure(figsize=(20,7))
subplot(1, 2, 1)
xray_gamma1 = (255*c*(xray_orig / 255)**gamma).astype('uint8')
plt.imshow(xray_gamma1, cmap=cm.gray)
plt.title('$c=1.0$, $\gamma = 2.0$')
subplot(1, 2, 2)
plt.hist(xray_gamma1.flatten(), 256, range=(2, 255), fc='k', ec='k');

c = 1.0
gamma = 0.5
figure(figsize=(20,7))
subplot(1, 2, 1)
xray_gamma2 = (255*c*(xray_orig / 255)**gamma).astype('uint8')
plt.imshow(xray_gamma2, cmap=cm.gray)
plt.title('$c=1.0$, $\gamma = 0.5$')
subplot(1, 2, 2)
plt.hist(xray_gamma2.flatten(), 256, range=(2, 255), fc='k', ec='k');

Gray-level Slicing

Sometimes we need to highlight a specific range of gray levels in a image. Possible application areas are finding masses of water in satellite imagery, enhancement of flaws in x-ray images, etc.

There are two basic approaches towards gray-level slicing.

  1. Binary Thresholding : all gray levels in the range of interest are displayed using a high value and the rest using a low value.
  2. Gradual Thresholding : desired range of gray levels are brightened but the background and the gray-level tonalities are preserved.

For exaple, let the range of interest be in $[120\,\, 180]$. Then the corresponding transformation functions would look like the following.


In [ ]:
r = np.arange(0, 256)
s = np.zeros(shape(r))
s[:] = 200
s[r<120] = 10
s[r>180] = 10
plt.subplot(1, 2, 1)
plot(r, s)
plt.axis([0, 255, 0, 255])
plt.title('Binary thresholding function')

r = np.arange(0, 256)
s = np.arange(0, 256)
a = r>120
b = r<180
s[a * b] = 200
plt.subplot(1, 2, 2)
plot(r, s)
plt.axis([0, 255, 0, 255])
plt.title('Gradual thresholding function')

Now let us use these ideas in real life!

Load Scikit-image. It has built-in image data sets. More at http://scikit-image.org/docs/dev/api/skimage.data.html


In [ ]:
from skimage import data
scanned = data.page()
plt.imshow(scanned, cmap=cm.gray)

In [ ]:
plt.hist(scanned.flatten(), 256, range=(0, 255), fc='k', ec='k');

By means of visual inspection we find that the texts in the scanned page has intensity values higher than $150$ and the background has intensity values lower than that.

(N.B. This is a very crude method! We will automate this later.)

So, we would consider a thresholding function that will search the image pixel by pixel. If the intensity of a pixel is greater than or equal to 150, it will be assigned a value of 255 and if its intensity falls below 150, a zero will be assigned in its place.


In [ ]:
thres = np.zeros(shape(scanned)).astype('uint8')
threshold = 150
thres[scanned<threshold] = 0
thres[scanned>=threshold] = 255
plt.imshow(thres, cmap=cm.gray)

Change the value of the threshold in the above programme and see the changes!

Now consider gradual thresholding. We consider that the region of interest lies between intensity values of 100 and 170.


In [ ]:
thres1 = copy(scanned)
threshold_hi = 170
threshold_lo = 100
thres1[scanned<threshold_lo] = 0
thres1[scanned>threshold_hi] = 255
plt.imshow(thres1, cmap=cm.gray)

Bit Plane Slicing

  1. Sometimes it is desirable to highlight the contribution made by specific bits to the total image appearance.
  2. The image can be imagined to be composed of Eight 1-bit planes -- Plane 0 for the LSB plane and Plane 7 for the MSB.
  3. The higher order bits contain visually significant data, the lower order plane contain more subtle details.

This can be acomplished by doing a Bitwise AND operation. For example, say the intensity of a pixel is 246 in decimal. In binary this would be $(1111\,0110)_2$. So in order to find the value of the 6th bit, one has to siimply do this $(1111\,0110)_2 \odot (0100\,0000)_2$. The result will simply produce the value of the 6th bit. In other words, $246 \odot 64$ will give you the value of the 6th bit.


In [ ]:
plane7 = A_gr & 128*np.ones(shape(A_gr)).astype('uint8')
plane6 = A_gr &  64*np.ones(shape(A_gr)).astype('uint8')
plane5 = A_gr &  32*np.ones(shape(A_gr)).astype('uint8')
plane4 = A_gr &  16*np.ones(shape(A_gr)).astype('uint8')
plane3 = A_gr &   8*np.ones(shape(A_gr)).astype('uint8')
plane2 = A_gr &   4*np.ones(shape(A_gr)).astype('uint8')
plane1 = A_gr &   2*np.ones(shape(A_gr)).astype('uint8')
plane0 = A_gr &   1*np.ones(shape(A_gr)).astype('uint8')

plt.figure(figsize=(20,7))
plt.subplot(2, 4, 1)
plt.imshow(plane7, cmap=cm.gray)
plt.title('Plane 7 (MSB)')
plt.subplot(2, 4, 2)
plt.imshow(plane6, cmap=cm.gray)
plt.title('Plane 6')
plt.subplot(2, 4, 3)
plt.imshow(plane5, cmap=cm.gray)
plt.title('Plane 5')
plt.subplot(2, 4, 4)
plt.imshow(plane4, cmap=cm.gray)
plt.title('Plane 4')
plt.subplot(2, 4, 5)
plt.imshow(plane3, cmap=cm.gray)
plt.title('Plane 3')
plt.subplot(2, 4, 6)
plt.imshow(plane2, cmap=cm.gray)
plt.title('Plane 2')
plt.subplot(2, 4, 7)
plt.imshow(plane1, cmap=cm.gray)
plt.title('Plane 1')
plt.subplot(2, 4, 8)
plt.imshow(plane0, cmap=cm.gray)
plt.title('Plane 0 (LSB)')

Spatial Filtering

Low pass filters attenuate or eliminate high frequncy components in the Fourier domain. Low pass filtering gives rise to image blurring.

High pass filters attenuate or eliminate low frequncy components in the Fourier domain. High pass filtering gives rise to sharpening of edges and other sharp details.

How to implement?

Utilize suitable 2D masks of suitable size, e.g. $3 \times 3$, $5 \times 5$ or $7 \times 7$. In most of our use cases, we shall restrict ourselves to $3 \times 3$ masks. The mask is applied to certain pixels. Upon application, the mask calculates a wighted sum of the neighbourhood of the concerned pixel and the result substitutes the original pixel. Here is how it works.

Consider that there is a pixel of intensity $z_5$. The $3 \times 3$ neighbourhood of the pixel can be seen as $\left[\begin{array}{ccc} z_{1} & z_{2} & z_{3}\\ z_{4} & z_{5} & z_{6}\\ z_{7} & z_{8} & z_{9} \end{array}\right]$. Now consider that the mask that is applied on the pixel with intensity $z_5$ looks like $\left[\begin{array}{ccc} w_{1} & w_{2} & w_{3}\\ w_{4} & w_{5} & w_{6}\\ w_{7} & w_{8} & w_{9} \end{array}\right]$.

Then, it will substitute $z_5$ by $w_1 z_1 + w_2 z_2 + \cdots + w_9 z_9$. Thus, $$z_{5\{new\}}=\sum_{i=1}^9 w_i z_i.$$

Thus, $$\left[\begin{array}{ccc} z_{1} & z_{2} & z_{3}\\ z_{4} & z_{5} & z_{6}\\ z_{7} & z_{8} & z_{9} \end{array}\right] \otimes \left[\begin{array}{ccc} w_{1} & w_{2} & w_{3}\\ w_{4} & w_{5} & w_{6}\\ w_{7} & w_{8} & w_{9} \end{array}\right] = \left[\begin{array}{ccc} z_{1} & z_{2} & z_{3}\\ z_{4} & \sum_{i=1}^9 w_i z_i & z_{6}\\ z_{7} & z_{8} & z_{9} \end{array}\right].$$

It does not do any changes to any other pixel anyway. The mask is centred on the image pixel whose new intensity value is tobe calculated. This calculation is performed for each pixel seperately by moving the mask to centre it on the pixel under consideration.

Smoothig spatial filters

Low pass spatial filtering

Examples of low pass spatial filter masks are

$$\mathcal{L}_1 = \dfrac{1}{9}\left[\begin{array}{ccc} 1 & 1 & 1\\ 1 & 1 & 1\\ 1 & 1 & 1 \end{array}\right]; \quad\quad \mathcal{L}_2 = \dfrac{1}{16}\left[\begin{array}{ccc} 1 & 2 & 1\\ 2 & 4 & 2\\ 1 & 2 & 1 \end{array}\right]$$
  1. A low pass filter must have all positive coefficients.
  2. For a low pass spatial filter mask shown as $\mathcal{L}_1$, the operation is also popularly termed as neighbourhood averaging. This averaging causes blurring and loss of sharpness.
  3. For a filter mask shown in $\mathcal{L}_2$, it is called weighted averaging.

Median filtering

Median filters are nonlinear (why?) filters employed with an objective of noise reduction, withot bluring.

$$\underset{\textrm{image section under consideration}}{\underbrace{\left[\begin{array}{ccc} z_{1} & z_{2} & z_{3}\\ z_{4} & z_{5} & z_{6}\\ z_{7} & z_{8} & z_{9} \end{array}\right]}} \underset{\textrm{median filtering}}{\Rightarrow} \underset{\textrm{result of median filtering}}{\underbrace{\left[\begin{array}{ccc} z_{1} & z_{2} & z_{3}\\ z_{4} & \textrm{med}\{z_{5}\} & z_{6}\\ z_{7} & z_{8} & z_{9} \end{array}\right]}},$$

where $\textrm{med}\{z_{5}\}$ is the median of $z_1, z_2, \cdots, z_9$. Median value can be easily calculated by arranging $z_1, z_2, \cdots, z_9$ in ascending order of maginitude and then finding the value that is in the middle position.

This filter is most effective when the noise pattern consists of niose-like components and it is of utmost importance to preserve edge sharpness.

Now we shall use scikit-image. There are many built in filters. Check http://scikit-image.org/docs/stable/api/skimage.filters.html for more.


In [ ]:
#mean filtering
from scipy import ndimage
lena_noisy = plt.imread('lena_noisy.png')
mask1 = (1/9)*np.array([[1, 1, 1], [1, 1, 1], [1, 1, 1]]) #LP mask
mask2 = (1/16)*np.array([[1, 2, 1], [2, 1, 2], [1, 2, 1]]) #LP mask
result1 = ndimage.convolve(lena_noisy, mask1, mode='constant', cval=0.0)
result2 = ndimage.convolve(lena_noisy, mask2, mode='constant', cval=0.0)
figure(figsize=(15, 5))
subplot(1, 3, 1)
imshow(lena_noisy, cmap=cm.gray)
title('Original')
subplot(1, 3, 2)
imshow(result1, cmap=cm.gray)
title('Filtered with LP mask $\mathcal{L}_1$')
subplot(1, 3, 3)
imshow(result2, cmap=cm.gray)
title('Filtered with LP mask $\mathcal{L}_2$')

See that the low pass filter significantly reduces the noise level.

Now let us apply median filtering to the same image.


In [ ]:
from skimage.morphology import disk #needed for the mask
from skimage.filters.rank import median
lena_med = median(lena_noisy, disk(3))
figure(figsize=(10, 5))
subplot(1, 2, 1)
imshow(lena_noisy, cmap=cm.gray)
title('Original')
subplot(1, 2, 2)
imshow(lena_med, cmap=cm.gray)
title('Median filtering')

As one can easily see, an image corrupted with speckle noise can be better denoised with mean filtering. Now check an image with salt and pepper noise.


In [ ]:
salt = plt.imread('saltandpepper.jpg')
mask = (1/9)*np.array([[1, 1, 1], [1, 1, 1], [1, 1, 1]])
salt_lp = ndimage.convolve(salt, mask, mode='constant', cval=0.0)
salt_med = median(salt, disk(2))
figure(figsize=(15, 5))
subplot(1, 3, 1)
imshow(salt, cmap=cm.gray)
title('Original')
subplot(1, 3, 2)
imshow(salt_lp, cmap=cm.gray)
title('Filtered with LP mask $\mathcal{L}_1$')
subplot(1, 3, 3)
imshow(salt_med, cmap=cm.gray)
title('Median filtering')

See the difference!

Sharpening spatial filters

Derivative filters

  1. The differentiation operation is expected to sharpen an image.
  2. One can use either first derivative or second derivative information.

Digital approximation of first derivative : $\dfrac{\partial f(x,y)}{\partial x} = f(x+1, y) - f(x, y), $ and $\dfrac{\partial f(x,y)}{\partial y} = f(x, y+1) - f(x, y).$

Constraints : The response of a first derivative filter must be

  1. zero in areas of constant intensity,
  2. must be non-zero at the onset of an intensity step or ramp,
  3. nonzero along ramps.

Digital approximation of second derivative : $\dfrac{\partial ^2f(x,y)}{\partial x^2} = f(x+1, y) - 2 f(x, y) +f(x-1, y),$ and $\dfrac{\partial ^2f(x,y)}{\partial y^2} = f(x, y+1) - 2 f(x, y) +f(x, y-1).$

Constraints : The response of a second order derivative filter must be

  1. zero in areas of constant intensity,
  2. non-zero at the onset and the end of an intensity step or ramp,
  3. zero along ramps of constant slope.

Implementing a first derivative filter for image sharpening

A first derivative image sharpening filter can be implemented by by applying the Gradient function. The gradient of a function $f(x, y)$ at coordinates $(x, y)$ is defined as the 2D column vector $$\nabla f(x,y)\equiv\mathrm{grad}(f)\equiv \mathbf{g} \equiv \left[\begin{array}{c} g_{x}(x,y)\\ g_{y}(x,y) \end{array}\right]=\left[\begin{array}{c} \frac{\partial f(x,y)}{\partial x}\\ \frac{\partial f(x,y)}{\partial y} \end{array}\right].$$ The magnitude (length) of vector $\nabla f$ is given by $M(x, y) = ||\nabla f||= \sqrt{\mathbf{g}^T \mathbf{g}} = \sqrt{g_x^2 + g_y^2}$. In image processing, this last expression is often approximated as $|g_x| + |g_y|$.

$M(x, y)$ is an image of the same size of the original and called the gradient image. The computation of this gradient is the basis of various approaches to develop first derivative filter.

If $\left[\begin{array}{ccc} z_{1} & z_{2} & z_{3}\\ z_{4} & z_{5} & z_{6}\\ z_{7} & z_{8} & z_{9} \end{array}\right]$ is the image section under consideration, then, according to the above theory, after the application of first derivative filtering, the new value of $z_5$ will be $M(x,y) = [(z_8 - z_5)^2 + (z_6 - z_5)^2]^{1/2} \approx |z_8 - z_5| + |z_6 - z_5|$. Another implementation involves cross-differences. $M(x,y) = [(z_9 - z_5)^2 + (z_8 - z_6)^2]^{1/2} \approx |z_9 - z_5| + |z_8 - z_6|$. But there is one problem : masks of even size are awkward to implement (why?). Hence an approximation with $3 \times 3$ neighbourhood is preferred. The mostly used first order derivative masks are Sobel masks and Prewitt masks. They are like follows :

$$\mathcal{S}_{y}=\underset{\textrm{Sobel horizontal derivative}}{\underbrace{\left[\begin{array}{ccc} -1 & -2 & -1\\ 0 & 0 & 0\\ 1 & 2 & 1 \end{array}\right]}}; \quad \mathcal{S}_{x}=\underset{\textrm{Sobel vertical derivative}}{\underbrace{\left[\begin{array}{ccc} -1 & 0 & 1\\ -2 & 0 & 2\\ -1 & 0 & 1 \end{array}\right]}}; \quad \mathcal{P}_{y}=\underset{\textrm{Prewitt horizontal derivative}}{\underbrace{\left[\begin{array}{ccc} -1 & -1 & -1\\ 0 & 0 & 0\\ 1 & 1 & 1 \end{array}\right]}}; \quad \mathcal{P}_{x}=\underset{\textrm{Prewitt vertical derivative}}{\underbrace{\left[\begin{array}{ccc} -1 & 0 & 1\\ -1 & 0 & 1\\ -1 & 0 & 1 \end{array}\right]}} .$$

Now, lets get to work!


In [ ]:
from skimage import filters, data
camera = data.camera()

#apply sobel gradient
sobel_camera = filters.sobel(camera)

#apply prewitt gradient
prewitt_camera = filters.prewitt(camera)

figure(figsize=(15, 5))
subplot(1, 3, 1)
imshow(camera, cmap=cm.gray)
title('Original')
subplot(1, 3, 2)
imshow(sobel_camera, cmap=cm.gray)
title('Result of a Sobel gradient')
subplot(1, 3, 3)
imshow(prewitt_camera, cmap=cm.gray)
title('Result of a Prewitt Gradient')

Another Example


In [ ]:
from skimage import filters, data
xray = plt.imread('chestxray.jpg')

#apply sobel gradient
sobel_xray = filters.sobel(xray)

#apply prewitt gradient
prewitt_xray = filters.prewitt(xray)

figure(figsize=(15, 5))
subplot(1, 3, 1)
imshow(xray, cmap=cm.gray)
title('Original')
subplot(1, 3, 2)
imshow(sobel_xray, cmap=cm.gray)
title('Result of a Sobel gradient')
subplot(1, 3, 3)
imshow(prewitt_xray, cmap=cm.gray)
title('Result of a Prewitt Gradient')

Laplacian filter

A second order derivative filter can be implemented by employing a Laplacian mask. The Laplacian of an image function $f(x, y)$ of two variables is defined as $\nabla ^2 f(x, y) = \dfrac{\partial ^2f(x,y)}{\partial x^2} + \dfrac{\partial ^2f(x,y)}{\partial y^2}.$

Thus, $\nabla ^2 f(x, y) = f(x+1, y) + f(x-1, y) + f(x, y+1) + f(x, y-1) - 4f(x,y)$. So, a Laplacian mask would look like

$$\nabla^2_\perp=\underset{\textrm{Laplacian mask}}{\underbrace{\left[\begin{array}{rrr} 0 & 1 & 0\\ 1 & -4 & 1\\ 0 & 1 & 0 \end{array}\right]}}; \quad \nabla^2_\odot=\underset{\textrm{Omnidirectional Laplacian mask}}{\underbrace{\left[\begin{array}{rrr} 1 & 1 & 1\\ 1 & -8 & 1\\ 1 & 1 & 1 \end{array}\right]}}.$$

The second one, here, considers four directions 1. horizontal, 2. vertical, 3. +45$^\circ$ and 4. -45$^\circ$, whereas, the first one only considers the horizontal and vertical directions.

However, there is a problem regarding the direct implementation of a Laplacian mask. Being a second derivative operation, it highlights intensity discontinuities in an image, and in the process de-emphasizes image regions having slow variations in intensity profile. SO, in order to preserve the original background features and yet perform sharpening operation, the Laplacian operator is utilized in the following manner: $$g(x,y) = f(x, y) + c\left[\nabla^2 f(x, y)\right],$$ where, $c=-1$ for the operators we considered.

Here is a comparison of Laplacian filter with Sobel filters. For the first time, we are going to use OpenCV.


In [ ]:
import cv2 #this is OpenCV
import numpy as np
from matplotlib import pyplot as plt

img = cv2.imread('dave.jpg',0) #import the image as grayscale

laplacian = cv2.Laplacian(img,cv2.CV_64F)
sobelx = cv2.Sobel(img,cv2.CV_64F,1,0,ksize=5)
sobely = cv2.Sobel(img,cv2.CV_64F,0,1,ksize=5)

plt.figure(figsize=(15, 10))
plt.subplot(2,2,1)
plt.imshow(img, cmap = 'gray')
plt.title('Original')
plt.subplot(2,2,2),plt.imshow(laplacian,cmap = 'gray')
plt.title('Laplacian')
plt.subplot(2,2,3)
plt.imshow(sobelx,cmap = 'gray')
plt.title('Sobel X')
plt.subplot(2,2,4)
plt.imshow(sobely,cmap = 'gray')
plt.title('Sobel Y')

plt.show()

You can also use Scikit-image to achieve the same goals.


In [ ]:
from skimage import img_as_ubyte, img_as_int
from scipy import ndimage
import cv2
moon = cv2.imread('blurry_moon.jpg',0)
img = img_as_int(moon)
laplacian1 = np.array([[0, 1, 0], [1, -4, 1], [0, 1, 0]], dtype='float64')  #Laplacian 1 mask
laplacian2 = np.array([[1, 1, 1], [1, -8, 1], [1, 1, 1]], dtype='float64') / 3.0 #Laplacian 2 mask
sobelx = np.array([[1, 0, -1], [2, 0, -2], [1, 0, -1]], dtype='float64') / 4.0 #Sobel x mask
sobely = np.array([[1, 2, 1], [0, 0, 0], [-1, -2, -1]], dtype='float64') / 4.0 #Sobel y mask

out_laplacian1 = img_as_ubyte(ndimage.convolve(img, laplacian1, mode='constant', cval=0.0))
out_laplacian2 = img_as_ubyte(ndimage.convolve(img, laplacian2, mode='constant', cval=0.0))
out_sobelx = img_as_ubyte(ndimage.convolve(img, sobelx, mode='constant', cval=0.0))
out_sobely = img_as_ubyte(ndimage.convolve(img, sobely, mode='constant', cval=0.0))
img_out_laplacian1 = img_as_ubyte(img - out_laplacian1)
img_out_laplacian2 = img_as_ubyte(img - out_laplacian2)
img_out_sobelx = img_as_ubyte(img - out_sobelx)
img_out_sobely = img_as_ubyte(img - out_sobely)

plt.figure(figsize=(15, 15))
plt.subplot(3,3,1)
plt.imshow(img, cmap = 'gray')
plt.title('Original')
plt.subplot(3,3,2)
plt.imshow(out_laplacian2, cmap = 'gray')
plt.title('Laplacian 1')
plt.subplot(3,3,3)
plt.imshow(out_laplacian2, cmap = 'gray')
plt.title('Laplacian 2')
plt.subplot(3,3,4)
plt.imshow(out_sobelx, cmap = 'gray')
plt.title('Sobel X')
plt.subplot(3,3,5)
plt.imshow(out_sobely, cmap = 'gray')
plt.title('Sobel Y')
plt.subplot(3,3,6)
plt.imshow(img_out_laplacian1, cmap = 'gray')
plt.title('Original - Laplacian 1')
plt.subplot(3,3,7)
plt.imshow(img_out_laplacian2, cmap = 'gray')
plt.title('Original - Laplacian 2')
plt.subplot(3,3,8)
plt.imshow(img_out_sobelx, cmap = 'gray')
plt.title('Original - Sobel X')
plt.subplot(3,3,9)
plt.imshow(img_out_sobely, cmap = 'gray')
plt.title('Original - Sobel Y')

Now, check out Python Imaging Library (fork: pillow). It contains many built in spatial filter modules.


In [ ]:
from PIL import Image
from PIL import ImageFilter
    
im0 = Image.open('blurry_moon.jpg')

figure(figsize=(15,15))
subplot(3,4,1)
plt.imshow(im0)
plt.title('Original')
subplot(3,4,2)
im2 = im0.filter(ImageFilter.CONTOUR)
plt.imshow(im2)
plt.title('Contour')
subplot(3,4,3)
im3 = im0.filter(ImageFilter.DETAIL) 
plt.imshow(im3)
plt.title('Detail')
subplot(3,4,4)
im4 = im0.filter(ImageFilter.EDGE_ENHANCE) 
plt.imshow(im4)
plt.title('Laplacian 1')
subplot(3,4,5)
im5 = im0.filter(ImageFilter.EDGE_ENHANCE_MORE)  
plt.imshow(im5)
plt.title('Laplacian 2')
subplot(3,4,6)
im6 = im0.filter(ImageFilter.EMBOSS)  
plt.imshow(im6)
plt.title('Emboss')
subplot(3,4,7)
im7 = im0.filter(ImageFilter.FIND_EDGES)
plt.imshow(im7)
plt.title('Sobel')
subplot(3,4,8)
im8 = im0.filter(ImageFilter.SMOOTH)  
plt.imshow(im8)
plt.title('Low Pass 1')
subplot(3,4,9)
im9 = im0.filter(ImageFilter.SMOOTH_MORE) 
plt.imshow(im9)
plt.title('Low Pass 2')
subplot(3,4,10)
im10 = im0.filter(ImageFilter.SHARPEN)
plt.imshow(im10)
plt.title('Sharpen')
subplot(3,4,10)
im1 = im0.filter(ImageFilter.BLUR)
plt.imshow(im1)
plt.title('Blur')

#Custom mask

size = (3, 3)
kernel1 = [1, 1, 1, 0, 0, 0, -1, -1, -1]
ker1 = ImageFilter.Kernel(size, kernel1, scale=None, offset=0)
subplot(3,4,11)
im11 = im0.filter(ker1)
plt.imshow(im11)
plt.title('Custom 1')

kernel2 = [1, 0, -1, 1, 0, -1, 0, 0, -1]
ker2 = ImageFilter.Kernel(size, kernel2, scale=None, offset=0)
subplot(3,4,12)
im12 = im0.filter(ker2)
plt.imshow(im12)
plt.title('Custom 2')

An important point to note : For a high pass spatial filter mask , whether utilizing first derivative or second derivative, the sum of the mask coefficients is always zero.

Unsharp masking and high boost filtering

This approach sharpens an image using kind of a 'back-calculation' method! Let the original image function be $f(x, y)$. First, a blurred version of the image is created. Let this version be $\bar{f}(x,y)$. Then this blurred version is subtracted from the original image. This creates a mask like $g_{mask}(x,y) = f(x, y) - \bar{f}(x,y)$. Then this mask is added to the original image resulting $g(x,y) = f(x,y) + k\,g_{mask}(x,y)$, where, $k$ is a constant.

If $k=1$, this process is called unsharp masking. When $k>1$, it is called high boost filtering.


In [ ]:
from skimage import data
from skimage import img_as_ubyte, img_as_int
from scipy import ndimage

coins = img_as_int(data.coins())
low_pass = (1/9)*np.array([[1, 1, 1], [1, 1, 1], [1, 1, 1]], dtype='float64') #LP mask
blurred = ndimage.convolve(coins, low_pass, mode='constant', cval=0.0)
unsharp_mask = img_as_ubyte(coins - blurred)
k = 1
sharpened = img_as_ubyte(coins + k*unsharp_mask)

figure(figsize=(15,5))
subplot(1, 3, 1)
title('Original')
imshow(coins, cmap=cm.gray)
subplot(1, 3, 2)
title('Unsharp mask')
imshow(unsharp_mask, cmap=cm.gray)
subplot(1, 3, 3)
title('Unsharp masked image')
imshow(sharpened, cmap=cm.gray)

In [ ]: