[PYTHON] Image binarization using linear discriminant analysis

Goal of this article

Understand Fisher's linear discriminant analysis and be able to binarize images using it.

Environment to use

Fisher's linear discriminant analysis

Overview

• A method for finding the parameters of a linear discriminant that classifies data.
• The linear discriminant is represented by $ y = w ^ {T} x $. Where $ y $ is the class, $ w $ is the parameter, and $ x $ is the data.
• $ w $ is asked to "increase the mean distance between classes and reduce the variance within the class".
• It is also closely related to the Bayesian optimal classifier.
• In addition to binarizing images, it can also be applied to data dimension reduction.

Objective function

Find the parameter $ w $ that maximizes the function below.

\begin{equation}
J(w) = \frac{w^{T}S_{B}w}{w^{T}S_{W}w}
\end{equation}

• $ w $: Linear discriminant parameter.
• $ S_ {W} $: Intraclass covariance matrix.
• $ S_ {B} $: Interclass covariance matrix.
• $ T $: Transpose.

In-class covariance matrix

Each class when $ N $ data is classified into $ k $ classes so that the number of data in each class is $ N_ {i} (i = 0, ..., k-1) $ The mean of the covariance matrix of.

\begin{equation}
S_{W}=\sum_{i=0}^{k-1}\frac{N_{i}}{N}S_{i}
\end{equation}

• $ S_ {i} $: Covariance matrix of class $ i $.

Interclass covariance matrix

A covariance matrix of the mean vectors for each class.

\begin{equation}
S_{B}=\sum_{i=0}^{k-1}\frac{N_{i}}{N}(\mu_{i}-\mu)^{T}(\mu_{i}-\mu)
\end{equation}

• $ \ mu_ {i} $: Mean vector of class $ i $.
• $ \ mu $: The average vector of the entire data.

Relationship between covariance matrix of all data, intraclass covariance matrix, and interclass covariance matrix

The sum of the intraclass covariance matrix ($ S_ {W} ) and the interclass covariance matrix ( S_ {W} ) is equal to the covariance matrix ( S $) for all the data. Hereinafter, the covariance matrix of all the data will be referred to as a total covariance matrix.

\begin{align}
S_{W} + S_{B} &=\sum_{i=0}^{k-1}\frac{N_{i}}{N}S_{i}+\sum_{i=0}^{k-1}\frac{N_{i}}{N}(\mu_{i}-\mu)^{T}(\mu_{i}-\mu)\\
&=\frac{1}{N}\sum_{i=0}^{k-1}(N_{i}(\sum_{j \in C_{i}}\frac{x_{j}^{T}x_{j}}{N_{i}}-\mu_{i}^{T}\mu_{i})+N_{i}\mu_{i}^{T}\mu_{i})-\mu^{T}\mu\\
&=\frac{1}{N}\sum_{i=0}^{k-1}(\sum_{j \in C_{i}}x_{j}^{T}x_{j})-\mu^{T}\mu\\
&=\frac{1}{N}\sum_{l=0}^{N-1}x_{l}^{T}x_{l}-\mu^{T}\mu\\
&=S
\end{align} 

• $ C_ {i} $: i-th class (set of data contained in).
• $ x_ {j} $: The jth data.

Determining parameters by Lagrangian

Even if you replace $ w $ in the objective function $ J (w) = \ frac {w ^ {T} S_ {B} w} {w ^ {T} S_ {W} w} $ with $ \ alpha w $ Since the value does not change, it can be replaced with $ w ^ {T} S_ {W} w = 1 $, and the problem of finding the parameter $ w $ that maximizes $ J (w) $ is $ w ^ {T. We come down to the problem of finding $ w $ that maximizes $ w ^ {T} S_ {B} w $ under the constraint of} S_ {W} w = 1 $.

\begin{equation}
argmax_{w}(w^{T}S_{B}w),
\hspace{3mm}subject\hspace{3mm}to\hspace{3mm}
w^{T}S_{W}w=1
\end{equation}

Convert this to a minimization problem.

\begin{equation}
argmin_{w}(-\frac{1}{2}w^{T}S_{B}w),
\hspace{3mm}subject\hspace{3mm}to\hspace{3mm}
w^{T}S_{W}w=1
\end{equation}

Then, the Lagrangian becomes as follows.

\begin{equation}
L=-\frac{1}{2}w^{T}S_{B}w+\frac{1}{2}\lambda (w^{T}S_{W}w - 1)
\end{equation}

Differentiate $ L $ with $ w $ and set it as $ 0 $ to get the following equation. This is called the generalized eigenvalue problem.

\begin{equation}
S_{B}w=\lambda S_{W}w
\end{equation}

Assuming $ S_ {W} $ is regular, multiply both sides by its inverse matrix $ S_ {W} ^ {-1} $ from the left to get the usual eigenvalue problem.

\begin{equation}
S_{W}^{-1}S_{B}w=\lambda w
\end{equation}

The eigenvector corresponding to the largest of the eigenvalues obtained as a result of solving the generalized eigenvalue problem is adopted as $ w $. This selection criterion is obtained by considering the dual problem of this optimization problem.

Image binarization using Fisher's discriminant analysis

Overview

• Convert grayscale images to black and white images.
• Set the threshold $ t $ and classify those with pixel values smaller than $ t $ into black and those with larger pixel values into white.
• Determine the threshold $ t $ based on Fisher's linear discriminant analysis. Then, the optimum threshold can be automatically selected in that sense.
• This method assumes an image in which the original image is originally expressed in black and white.

Objective function

The objective function $ J (w) $ of Fisher's discriminant in the previous chapter is the following equation using the relationship of intraclass variance, interclass variance, and total variance, noting that the grayscale image data is one-dimensional. To get. Note that the parameter $ w $ is one-dimensional.

\begin{equation}
J(w) = \frac{\sigma_{B}}{\sigma_{W}}=\frac{\sigma_{B}}{\sigma -\sigma_{B}}
\end{equation}

• $ \ sigma_ {W} $: In-class distribution.
• $ \ sigma_ {B} $: Distribution between classes.
• $ \ sigma $: Fully distributed.

Threshold determination

Since the total variance $ \ sigma $ is constant, we can determine the threshold $ t $ that makes up $ \ sigma_ {B} $ that maximizes $ J (w) $.

$ \ sigma_ {B} $ can be calculated by the following equation.

\begin{equation}
\sigma_{B}=\frac{N_{0}}{N}(\mu_{0}-\mu)^2 + \frac{N_{1}}{N}(\mu_{1}-\mu)^{2}
\end{equation}

• $ N $: Number of pixels in the entire image
• $ N_ {0} $: Number of pixels classified as black
• $ N_ {1} $: Number of pixels classified as white
• $ \ mu $: Average pixel value of the entire image
• $ \ mu_ {0} $: Average pixel value of pixels classified as black
• $ \ mu_ {1} $: Average pixel value of pixels classified as white

If the minimum and maximum pixel values of a grayscale image are $ x_ {min} $ and $ x_ {max} $, respectively, then $ x_ {min} \ leq t \ leq x_ {max} $. Calculate $ \ sigma_ {B} $ for all such $ t $, find the one that maximizes $ J (w) $, and use it as the threshold.

program

# This file is part of "Junya's self learning project about digital image processing."
#
# "Junya's self learning project about digital image processing"
# is free software: you can redistribute it and/or modify
# it under the terms of the GNU General Public License as published by
# the Free Software Foundation, either version 3 of the License, or
# (at your option) any later version.
#
# "Junya's self learning project about digital image processing"
# is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
# GNU General Public License for more details.
#
# You should have received a copy of the GNU General Public License
# along with Foobar.  If not, see <http://www.gnu.org/licenses/>.
#
# (c) Junya Kaneko <[email protected]>

from PIL import Image
import numpy as np
from matplotlib import pyplot as plt


def discriminant_analysis(image):
    _image = image.reshape(image.shape[0] * image.shape[1])
    ave = np.average(_image)
    t = j0 = 0
    for _t in range(np.min(_image) + 1, np.max(_image) + 1):
        blacks = _image[_image < _t]
        whites = _image[_image >= _t]
        sigma_b = len(blacks) * np.power(np.average(blacks) - ave, 2) /len(_image) + len(whites) * np.power((np.average(whites) - ave), 2) / len(_image)
        j1 = sigma_b / (np.var(_image) - sigma_b)
        if j0 < j1:
            t = _t
            j0 = j1
    return t


if __name__ == '__main__':
    image = np.array(Image.open('img/test.jpg').convert('L'), dtype=np.uint8)
    bin_image = np.array(image)
    t = discriminant_analysis(image)

    for i in range(image.shape[0]):
        for j in range(image.shape[1]):
            bin_image[i, j] = 255 if image[i, j] >= t else 0

    plt.subplot(1, 2, 1)
    plt.title('Original')
    plt.imshow(image, cmap='Greys_r')

    plt.subplot(1, 2, 2)
    plt.title('Binary')
    plt.imshow(bin_image, cmap='Greys_r')

    plt.show()

The source code can be found on Github.

The execution result is as shown in the figure below. The original image is [Dictionary Wiki](http://ja.characters.wikia.com/wiki/%E3%83%95%E3%82%A1%E3%82%A4%E3%83%AB: % E4% B8% 89% E4% BD% 93% E7% BF% 92% E5% AD% 97% E3% 83% BB% E6% A5% B7_-_% E6% 88% 91.jpg) used.

References

• Digital Image Processing (CG-ARTS Association)
• Dimensionality Reduction (Javier Hernandez Rivera)
• Penalized linear discriminant analysis (Daniela M. Witten et al.)
• Fisher's Linear Discriminant Analysis (Max Welling)
• A Threshold Selection Method from Gray-Level Histograms (Nobuyuki Otu)
• Introduction to Linear Optimization (Dimitris Bertsimas et al.)
• Engineering basic optimization and its application (new engineering mathematics) (Hiroshi Yabe) = 1211 & creativeASIN = 4901683349 & linkCode = as2 & tag = mpsamurai-22)