Poor classification accuracy when using PCA

1 次查看（过去 30 天）

Rayne 2015-11-1

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/252135-poor-classification-accuracy-when-using-pca

I'm trying to classify a set of images using PCA. The training set contains ~1800 images of 380 persons, each image with the person's unique label (ID). The testing set contains ~750 images of the same 380 persons (different images), each image with the person's corresponding label (ID). All images are 160 x 128 pixels.

My code is as follows.

img_train{i} contains the i-th original image in the training set, img_test{i} contains the i-th original image in the testing set.

euclide_dist = zeros(total_TrainImageFiles,1);
num_correct_labels = 0;
% Reshape 2D training images into 1D image vectors
train_img = zeros(irow*icol,total_TrainImageFiles);
for i = 1 : total_TrainImageFiles
    temp = reshape(img_train{i}',irow*icol,1);
    train_img(:,i) = temp;
end
% Calculate mean image vector
mean_face = mean(train_img,2);
% Subtract mean face from all training images
centred_data = train_img - repmat(mean_face, 1, total_TrainImageFiles);
% Determine eigenvectors and eigenvalues using SVD
[U, D, V] = svd(centred_data,0);
d = 10;
% Generate feature vectors from training set for subsequent classification
% Keep the top 'd' eigenvectors
eigenvectors = U(:,1:d);
% Project the training images into the facespace to generate the training feature vectors
train_features = eigenvectors' * centred_data;
% Classify testing images
for i = 1 : total_TestImageFiles
    % Reshape 2D test image into 1D image vectors
    test_img = reshape(img_test{i}',irow*icol,1);
    % Subtract mean face from test image
    centred_test_img = double(test_img) - mean_face;
    % Project test image onto the facespace
    test_features = eigenvectors' * centred_test_img;
      % Calculate the euclidian distance of all projected trained images from the projected test image
      for j = 1 : total_TrainImageFiles
          edist = (norm(train_features(:,j) - test_features))^2;
          euclide_dist(j) = edist;
      end
      % Find the minimum distance and compare class labels
      [euclide_dist_min,train_index] = min(euclide_dist);
      predicted_labels(i) = training_label(train_index);
      if training_label(train_index) == testing_label(i)
          num_correct_labels = num_correct_labels + 1;
      end 
  end
% Calculate accuracy
accuracy = double(num_correct_labels) / double(total_TrainImageFiles);