looking for imaging processing expert

Hi

I have a problem with using autoencoder for recreating fingerprinting images

I am using MATLAB and Sparse autoencoder

I am using extractHOGFeatures for extracting the features to be trained for autoencoder

I got good results on the performance of training and mseError which reaching 0.0039 However when i use autoencoder with testing images i got black color on the whole Reconstructed images

I followed this example in MATLAB for autoencoder but I am still facing the problem for Reconstructed images that are with black color for whole images

Herein the code

% Data processing 
 
clear all;
close all;
clc; 
 
%%  Initalize the data 
 
dataDir= fullfile('Data/');
exts = {'.jpg','.png','.tif','BMP'};
 
 
imds = imageDatastore(fullfile(dataDir),...
'IncludeSubfolders',true,'FileExtensions','.jpg','LabelSource','foldernames');
countEachLabel(imds);
 
[TrainData, TestData] = splitEachLabel(imds, 0.5);
size(TrainData);
countEachLabel(TrainData);
 
%% reference 
 
figure(); 
imshow(TrainData.Files{1}); 
 
 
img = readimage(TrainData, 1); 
[hog_8x8, vis8x8] = extractHOGFeatures(img, 'CellSize', [8 8]); 
 
CellSize = [8 8]; 
hogFeatureSize = length (hog_8x8);
 
numImages = numel(TrainData.Files); 
trainingFeatures = zeros(numImages, hogFeatureSize, 'single'); 
 
for i = 1:numImages
    img = readimage(TrainData, i); 
    
    img  = imbinarize(img); 
    trainingFeatures(i, :) = extractHOGFeatures(img, 'CellSize', CellSize); 
    
end 
 %% Training Data
trainingLabels = TrainData.Labels; 
 
X  = trainingFeatures; 
 
%% train a spare autoencoder 
 
  
hiddenSize = 25;
autoenc = trainAutoencoder(X,hiddenSize,'MaxEpochs',1000,...
'DecoderTransferFunction','purelin','EncoderTransferFunction','satlin','L2WeightRegularization',0.004,'SparsityRegularization',4,'SparsityProportion',0.15);
    

    
 
 
Xreconstructed = predict(autoenc, X); 
 
mseError = mse(X - Xreconstructed);
 figure();
for i = 1:10
    subplot(4,5,i);
    imshow(Xreconstructed(i));
end
 
 
%% Test data   
 
 
 
img = readimage(TestData, 1); 
[hog_8x8, vis8x8] = extractHOGFeatures(img, 'CellSize', [8 8]); 
 
CellSize = [8 8]; 
hogFeatureSize = length (hog_8x8);
 
numImages = numel(TestData.Files); 
testingFeatures = zeros(numImages, hogFeatureSize, 'single'); 
 
for i = 1:numImages
    img = readimage(TestData, i); 
    
    img  = imbinarize(img); 
    testingFeatures(i, :) = extractHOGFeatures(img, 'CellSize', CellSize); 
    
end 
 
testingLabels = TestData.Labels;
testFeature = reshape(testingFeatures, [102 20736]); 
testFeature'; %#ok<VUNUS>
xReconstructed = predict(autoenc,testFeature);
 %% Test Images 
figure();
for i = 1:10
    subplot(4,5,i);
    imshow(TestData.Files{i});
end
 
 %% Reconstructed images from TestData
figure();
for i = 1:10
    subplot(4,5,i);
    imshow(xReconstructed(i));
end

Answer 1

[Walter Roberson]

    imshow(Xreconstructed(i));

What datatype is Xreconstructed(i) ? It looks to me from the code that it will be scalar double, in the range of 1 to the number of classes or perhaps 0 to 255. It looks to me as if you are asking imshow() to display one pixel at a time, and without scaling it either.

You should probably be using something more like

  imshow(Xreconstructed, [0 255])

[Abdussalam Elhanashi]

Hi Walter Roberson

This is the data of xReconstructed(i)

I am using fingerprint images 200 x 200

However I tried to make

imshow(Reconstructed, [0 255])

but i get like in this image

[Walter Roberson]

imshow(Xreconstructed, [])

[Abdussalam Elhanashi]

Hi Walter

Thank you for your reply

I used imshow(xreconstructed, []) but i get following result one line with dots

supposed to be as these or similar to fingerprints

[Walter Roberson]

No it should not be anything like that when you display that array. You are extracting HOG features from the image, building an autoencoder, and putting the extracted features through the autoencoder prediction. The result is the same size as the extracted features, not the same size as the image, so you should not expect that the visualization of it would resemble a fingerprint in any respect.