PCA (Data Augmentation) with Scikit-Image

pca.py

import numpy as np
from skimage import io, transform

# Using one image
image_filenames = 'image.jpg'

# Read the image
image = io.imread(images_filenames)

# Resize image (256 x 256 x 3) and turn the image matrix of m x n x 3 to lists of rgb 
images_resized = transform.resize(image, (256, 256, 3))

# Center the image data and then find the covariance
image_reshaped = images_resized.reshape(-1,3) # (256*256, 3)
image_centered = image_reshaped - image_reshaped.mean(axis=0)

# Find 3 x 3 covariance martix
image_cov = np.cov(image_centered, rowvar=False)

# Get eigen values and eigen vectors
eigen_vals, eigen_vecs = np.linalg.eigh(img_cov)

# Get principal components[p1, p2, p3]
sort = eigen_vals[::-1].argsort()
eigen_vals[::-1].sort()     
eigem_vecs = eigen_vecs[:, sort]
pc = np.column_stack((eigen_vecs))

# 3x1 matrix of eigen values multipled by a random variable drawn from Gaussian(0, 0.1)
m = np.zeros((3,1))
for i in range(eigen_vals.shape[0]):
    m[i][0] = np.random.normal(0, 0.1)*eigen_vals[i]

# Add value[p1, p2, p3][α1λ1, α2λ2, α3λ3]T
add_v = np.matrix(pc)*np.matrix(m)

# Scale the image
image_scaled = image/255.0

# Adding the value to create a augmented image
for i in range(image_scaled.shape[0]):
    for j in range(image_scaled.shape[1]):
        for k in range(image_scaled.shape[2]):
            image_scaled[i, j, k] = min(abs(float(image_scaled[i, j, k]) + float(add_v[k])), 1)