Face recognition using Eigen face using Pytorch
Face Recognition using Eigenfaces**
We will be using FEI Face Database for our face recognition purpose.
Datasets consist of faces of 200 people and each person has two frontal images (one with a neutral expression and the other with a smiling facial expression), there are 400 full frontal face images manually registered and cropped.
We will use normalized, equalized and cropped frontal face images.
Add dataset in your Colab using:
!wget <link>
! wget https://fei.edu.br/~cet/frontalimages_manuallyaligned_part1.zip
! wget https://fei.edu.br/~cet/frontalimages_manuallyaligned_part2.zip
--2019-12-12 05:17:05-- https://fei.edu.br/~cet/frontalimages_manuallyaligned_part1.zip
Resolving fei.edu.br (fei.edu.br)... 200.232.90.210
Connecting to fei.edu.br (fei.edu.br)|200.232.90.210|:443... connected.
HTTP request sent, awaiting response... 200 OK
Length: 6209322 (5.9M) [application/x-zip-compressed]
Saving to: ‘frontalimages_manuallyaligned_part1.zip’
frontalimages_manua 100%[===================>] 5.92M 3.23MB/s in 1.8s
2019-12-12 05:17:07 (3.23 MB/s) - ‘frontalimages_manuallyaligned_part1.zip’ saved [6209322/6209322]
--2019-12-12 05:17:11-- https://fei.edu.br/~cet/frontalimages_manuallyaligned_part2.zip
Resolving fei.edu.br (fei.edu.br)... 200.232.90.210
Connecting to fei.edu.br (fei.edu.br)|200.232.90.210|:443... connected.
HTTP request sent, awaiting response... 200 OK
Length: 6168379 (5.9M) [application/x-zip-compressed]
Saving to: ‘frontalimages_manuallyaligned_part2.zip’
frontalimages_manua 100%[===================>] 5.88M 3.02MB/s in 1.9s
2019-12-12 05:17:13 (3.02 MB/s) - ‘frontalimages_manuallyaligned_part2.zip’ saved [6168379/6168379]
Unzip your zip files using:
!unzip <name_of_your_zip_file>
!unzip frontalimages_manuallyaligned_part1.zip
!unzip frontalimages_manuallyaligned_part2.zip
Archive: frontalimages_manuallyaligned_part1.zip
inflating: 178b.jpg
inflating: 179a.jpg
inflating: 179b.jpg
inflating: 180a.jpg
inflating: 180b.jpg
inflating: 181a.jpg
inflating: 181b.jpg
# importing all the necessary packages
import numpy as np
from matplotlib import pyplot as plt
from PIL import Image
###Vectorization
We will use Python Imaging Library (PIL) to load images from files.
PIL.Image.open() Opens and identifies the given image file.Note Open into greyscale, or L(Luminance) mode:
img = Image.open(f'{i}a.jpg').convert('L') # i ranges from 1 to 200
img = img.resize((width, height),Image.ANTIALIAS) # for speed
Rearrange each Face image N x N into a column vector N2 x 1. After Vectorization the shape of face matrix should be 200 x (width x height)
# neutral face
neutral = []
for i in range(200):
i += 1
img = Image.open(f'{i}b.jpg').convert('L')
img = img.resize((58,49), Image.ANTIALIAS)
img2 = np.array(img).flatten() # vectorization
neutral.append(img2)
# Check face_matrix.shape
faces_matrix = np.vstack(neutral)
faces_matrix.shape
(200, 2842)
# Find mean_face
mean_face = np.mean(faces_matrix, axis=0)
mean_face.shape
(2842,)
plt.imshow(mean_face.reshape(height,width),cmap='gray');
plt.title('Mean Face')
plt.imshow(mean_face.reshape(49,58),cmap='gray')
<matplotlib.image.AxesImage at 0x7f6fe62d8e80>
Normalization
#Perfrom Normalization
faces_norm = faces_matrix - mean_face
faces_norm.shape
(200, 2842)
# Calculate covariance matrix
#np.cov expects features as rows and observations as columns, so transpose
face_cov = np.cov(faces_norm.T)
face_cov.shape
(2842, 2842)
#Find Eigen Vectors and Eigen Values, you can use SVD from np.linalg.svd
eigen_vecs, eigen_vals, _ = np.linalg.svd(face_cov)
eigen_vecs.shape
(2842, 2842)
Plot the first 10 Eigenfaces
fig, axs = plt.subplots(1,3,figsize=(15,5))
for i in np.arange(10):
ax = plt.subplot(2,5,i+1)
img = <eigen_vectors>[:,i].reshape(height,width)
plt.imshow(img, cmap='gray')
fig.suptitle("First 10 Eigenfaces", fontsize=16)
#Plot the first 10 Eigenfaces
fig, axs = plt.subplots(1,3,figsize=(15,5))
for i in np.arange(10):
ax = plt.subplot(2,5,i+1)
img = eigen_vecs[:,i].reshape(49,58)
plt.imshow(img, cmap='gray')
fig.suptitle("First 10 Eigenfaces", fontsize=16)
Text(0.5, 0.98, 'First 10 Eigenfaces')
Reconstruction with increasing Eigenfaces
fig, axs = plt.subplots(2,5,figsize=(15,6))
for k, i in zip([0,1,9,19,39,79,159,199,399,799],np.arange(10)):
# Reconstruct the first picture '1a.jpg' whose index is 0.
# Get PC scores of the images (wights)
# Reconstruct first face in dataset using k PCs (projected_face)
ax = plt.subplot(2,5,i+1)
ax.set_title("k = "+str(k+1))
plt.imshow(projected_face.reshape(height,width)+mean_face.reshape(height,width),cmap='gray');
# Reconstruct with increasing Eigenfaces
fig, axs = plt.subplots(2,5,figsize=(15,6))
for k, i in zip([0,1,9,19,39,79,159,199,399,799],np.arange(10)):
# Reconstruct the first picture '1a.jpg' whose index is 0.
weight = faces_norm[0,:].dot(eigen_vecs[:,:k]) # Get PC scores of the images
projected_face = weight.dot(eigen_vecs[:,:k].T) # Reconstruct first face in dataset using k PCs
ax = plt.subplot(2,5,i+1)
ax.set_title("k = "+str(k+1))
plt.imshow(projected_face.reshape(49,58)+mean_face.reshape(49,58),cmap='gray');
fig.suptitle(("Reconstruction with Increasing Eigenfaces"), fontsize=16)
Text(0.5, 0.98, 'Reconstruction with Increasing Eigenfaces')
Written on January 18, 2020