lorecioni / HEp-2-CellsClassification

Fisher tensors for classifying human epithelial HEp-2 cells implementation, based on SVM classifier.

#HEp-2 Cells Classification#

Matlab code implementing Fisher tensors for classifying human epithelial cells for ICPR 2014 Contest on HEp-2 Cells Classification. Here you can find the dataset.

###Configuration###

Before launching the program edit configuration file.

• extract_train for extracting features from train images
• extract_train for extracting features from test images

Single cell configuration

• train_path train image path
• train_labels train labels groundtruth
• test_path test images path
• test_labels test images groundtruth

Full image configuration

• train_path train images folder path.
• train_labels train images groundtruth (mat file).
• test_path test images folder path.
• test_labels test images groundtruth (mat file).

Patterns

• patterns maps patterns names into ids.

Feature extraction option

• Gabor_options Gabor filters settings.
• block_size sliding window size.
• delta sliding window step
• gray convert images in grayscale (if they are not).
• resize true/false, if true images will be resized.
• resizeTo if resize is true set the width of the resized images.

Classification options

• use_NN_classifier evaluate results with NN classifier.
• use_SVM_classifier evaluate results with SVM classifier.
• K number of gaussians in GMM.
• crossvalidate for evaluate dataset with crossvalidation.
• crossvalidate_SVM_parameters for tuning SVM parameters with crossvalidation (slow).
• kFolds number of folds for cross-validation.
• showConfusionMatrix display confusion matrix.

###Running###

Prepare first your training set. You can choose between:

• loadDataset loads dataset images and groundtruth. Creates a mat file for associating image filename, mask filename and pattern id.

This will show you a table containing image id, label and filename.

Run the code as follows:

• extractFeature extracts Covariance Descriptor from each image in dataset.
• runGMM execute GMM.
• saveSignatures save signatures for images, fisher tensors.
• runClassifier run SVM classifier.