In [1]:
# simple image compression by svd
using Images, FileIO, PyPlot
# read khulogo.jpg anto an array
img = Float64.(load("khulogo.jpg"));
# img: mxn image
m = size(img,1);
n = size(img,2);
# singular value decomposition
U, S, V = svd(img);
# take 15% of the most significant singular values
cutRatio = 0.15;
N = round(Int, cutRatio*length(S) );
# compression by taking N largest singular values
U_compressed = U[:,1:N];
S_compressed = S[1:N];
V_compressed = V[:,1:N];
# plot singular values
figure(1);
grid("True");
plot(S);
plot(S_compressed);
title("Singular values");
# image reconstruction
img_compressed = U_compressed*diagm(S_compressed)*V_compressed';
img_compressed = max(min(img_compressed,1.0),0);
save("khulogo_compressed.jpg",img_compressed);
# compare the image quality
figure(2);
subplot(1,2,1);
imshow(img, cmap="gray");
title("Original image");
subplot(1,2,2);
imshow(img_compressed, cmap="gray");
title("Compressed image");
# compute the compression ratio and print out the results
Nreal = m*n;
Nreal_compressed = N*(m+n+1);
println("$(Nreal) real numbers in the original image")
println("$(Nreal_compressed) real numbers in the original image")
println("Compression ratio: $(Nreal_compressed/Nreal)")
In [2]:
using Metalhead
# load the vgg-19 deep neural networks
vgg = VGG19();
# classify the images
class_original = classify(vgg, Gray.(Float32.(img)));
class_compressed = classify(vgg, Gray.(Float32.(img_compressed)));
println("Original logo: ",class_original);
println("Compressed logo: ",class_compressed);