One-Class Convolutional Neural Network (OC-CNN)

This text is study for arxiv.org/pdf/1901.08688.pdf

 

Overview

• Present a novel cnn based approach for one class classification, which is end-to-end trainable
• Use a zero centered gaussian noise in the latent space as the pseudo-negative class
• Train the network using cross-entropy loss
  • To learn a good representation
  • To find a decision boundary
• Any pre-trained CNN can be used as the base network for one class classification

 

Proposed Approach

• Overall order
  • Consists of feature extractor and classifier
  • Extracted features are then appended with the pseudo-negative class data, generated from a zero centered gaussian in the feature space
  • Classifier outputs 1 for target class, 0 for negative class
• Feature Extractor
  • Any pre-trained CNN can be used
  • In this paper, pre-trained AlexNet and VGG16 is used by removing the softmax regression layers
  • Training
    • Freeze convolution layers and only train fully-connected layers
    • The features are appended with the pseudo-negative data generated from Gaussian
• Classification Network
  • A simple fully-coneected layer followed by a softmax regression layer is used as the classifier network
• Loss Function
  • Binary-cross entropy