Source-linked AI summary
One-Class Convolutional Neural Network
Poojan Oza, Vishal M. Patel
TL;DR
One-class classification must identify target-class samples despite having no negative-class training data. OC-CNN introduces Gaussian pseudo-negative samples in feature space and trains a CNN discriminatively with binary cross-entropy. Across authentication, abnormality-detection, and novelty-detection datasets, it reports consistent improvements over competing approaches.
Problem
One-class classification must learn a target-class detector from target-class data alone because negative-class data are unavailable.
Method
OC-CNN appends zero-centered Gaussian pseudo-negative samples to extracted features and trains an end-to-end CNN classifier with binary cross-entropy.
Results
OC-CNN or OC-SVM+ achieves the best performance on all three datasets, with OC-CNN improving over OC-NN by approximately 4%, 10%, and 5% across the reported datasets.
Takeaways & Limitations
The approach functions both as a standalone one-class classifier and as a target-class feature extractor usable with a more sophisticated classifier.
Abstract
from arXiv · showhide
We present a novel Convolutional Neural Network (CNN) based approach for one class classification. The idea is to use a zero centered Gaussian noise in the latent space as the pseudo-negative class and train the network using the cross-entropy loss to learn a good representation as well as the decision boundary for the given class. A key feature of the proposed approach is that any pre-trained CNN can be used as the base network for one class classification. The proposed One Class CNN (OC-CNN) is evaluated on the UMDAA-02 Face, Abnormality-1001, FounderType-200 datasets. These datasets are related to a variety of one class application problems such as user authentication, abnormality detection and novelty detection. Extensive experiments demonstrate that the proposed method achieves significant improvements over the recent state-of-the-art methods. The source code is available at : github.com/otkupjnoz/oc-cnn.
I. INTRODUCTION
One-class classification learns to identify a target class using only target-class training data, making the unavailable negative class a central challenge. OC-CNN addresses this gap with an end-to-end discriminative CNN approach that avoids external negative data and performs well across one-class applications.
- Problem: One-class classification identifies target-class objects among all possible objects using training data containing only that class.The missing negative-class data makes learning the decision boundary difficult.
- Applications: One-class classification supports anomaly detection, novelty detection, and user authentication, including settings where novel-class data are unavailable beforehand.Novelty detection therefore trains only on target-class data.
- Prior Work: Prior approaches include statistical methods such as OC-SVM, SVDD, and single-class MPM, alongside generative and discriminative deep-learning methods.These methods differ in how they represent the target class and construct a decision boundary.
- Gap: Discriminative deep approaches remain relatively underexplored, and one prior method uses an external reference dataset as negative-class data.OC-CNN instead learns representations end-to-end without using negative-class data.
- Contribution: The paper proposes an end-to-end CNN-based one-class classifier and reports that it outperforms statistical and deep-learning alternatives while generalizing across applications.The claimed application scope includes authentication, abnormality detection, and novelty detection.
II. PROPOSED APPROACH
The proposed approach combines a feature extractor with a classifier and augments target-class features with pseudo-negative Gaussian samples. The classifier then assigns each representation to the target or pseudo-negative class.
- Architecture: The network consists of a feature extractor that embeds target-class images and a classifier that assigns confidence scores to feature representations.The classifier output is binary: 1 for target-class data and 0 for the other class.
- Pseudo-Negative Data: Extracted target-class features are appended with pseudo-negative samples generated from a zero-centered Gaussian in the feature space.The combined representations are passed to a fully connected classification network.
A. Feature Extractor
OC-CNN can use a pre-trained CNN as its feature extractor, with extracted features augmented by Gaussian pseudo-negative data before classification.
- Feature Extractor: Any pre-trained CNN can serve as the feature extractor, and the paper uses AlexNet and VGG16 after removing their softmax layers.The convolutional layers are frozen while the fully connected layers are trained.
- Gaussian Sampling: For D-dimensional features, pseudo-negative samples are drawn from N(¯µ, σ^2 · I), representing D independent one-dimensional Gaussians.Here, I is a D × D identity matrix, while ¯µ and σ specify the Gaussian mean and standard deviation.
B. Classification Network
The classification network distinguishes extracted features from Gaussian pseudo-negative samples using a two-output softmax classifier trained with binary cross-entropy.
- Classification Network: Appending pseudo-negative data creates paired classifier inputs, processed by a fully connected layer followed by a two-output softmax layer.The fully connected layer has the same dimension as the feature representation.
- Loss Function: Binary cross-entropy trains the network to distinguish feature-extractor inputs from Gaussian samples, with p denoting the softmax probability of y = 0.The labels are y = 0 for feature-extractor inputs and y = 1 for Gaussian samples.
- Optimization: Training uses Adam with a 10^-4 learning rate and input image batches of 64.These are the reported optimization settings for the experiments.
- Training Stabilization: The Gaussian parameters are set to ¯µ = ¯0 and σ = 0.01, and instance normalization is applied before classification to stabilize training.The normalization is reported as useful for stabilizing the training procedure.
III. EXPERIMENTAL RESULTS
The experiments evaluate one-class classification across abnormality detection, face-based user authentication, and novelty detection using three datasets and compare OC-CNN with statistical and neural baselines. The comparison includes AlexNet and VGG16 features, with OC-SVM+ additionally assessed as a non-end-to-end variant.
- Experiments cover abnormality detection, face-based user authentication, and novelty detection using Abnormality-1001, UMDAA-02, and FounderType-200.The datasets are used for the three application problems, respectively.
- The evaluation compares OC-CNN with OC-SVM, BSVM, MPM, SVDD, OC-NN, and OC-SVM+ baselines.The listed methods include statistical and neural one-class classification approaches.
- OC-SVM and SVDD are trained on AlexNet and VGG16 features, while BSVM uses zero-centered Gaussian noise as negative data.
- OC-NN is evaluated with linear, Sigmoid, and ReLU activations using AlexNet and VGG16 feature extractors for fair comparison.The compared OC-NN configuration differs from its original auto-encoder-based feature-extractor methodology.
- OC-SVM+ applies OC-SVM to OC-CNN-extracted features but is not end-to-end trainable.
A. Abnormality Detection
The abnormality-detection experiment identifies abnormal instances using training data containing only normal instances. Abnormality-1001 supplies six object categories with corresponding normal images from PASCAL VOC and balanced test sets.
- Abnormality detection identifies instances dissimilar to target-class instances when abnormal examples are unavailable during training.
- Abnormality-1001 contains 1001 abnormal images across Chair, Car, Airplane, Boat, Sofa, and Motorbike classes.
- Normal images come from PASCAL VOC, with train-test splits constructed so abnormal and normal test counts are equal.
- Reported results are averaged across all six abnormality-detection classes.
B. User Active Authentication
The user active-authentication experiment treats unauthorized-user rejection as a one-class classification problem and uses face images from UMDAA-02. Each target-user model is trained on that user's training split and evaluated against the corresponding test data.
- Active authentication identifies an enrolled user from biometric data and can reject unauthorized users as abnormal behavior.
- The experiment uses UMDAA-02 face data containing 33209 images from 48 users.
- Face images exhibit substantial variation in pose, illumination, appearance, and occlusions.
- For each user, training and test sets maintain an 80/20 ratio, with the target-user train set used for model training.
C. Novelty Detection
The novelty-detection experiment uses FounderType-200 to distinguish known font classes from novel classes. Results are organized as mean performance across classes for AlexNet and VGG16 base networks.
- FounderType-200 contains 6763 images from 200 different Chinese font types.
- The first 100 font classes are treated as target classes, while the remaining 100 classes provide novel data.
- Each target class is split into equal-sized train and test sets, and each novel class contributes 50 images to the novel set.
- Tables I and II report mean performance across classes using AlexNet and VGG16, respectively, highlighting the best and second-best methods.
- Evaluation compares each target-class test set against the novel set and averages results over all 100 target classes.
IV. RESULTS AND DISCUSSION
OC-CNN or OC-SVM+ achieves the best performance across all three datasets, with learned representations substantially improving one-class classification results. VGG16 generally outperforms AlexNet, while OC-NN performance varies with evaluation conditions.
- OC-CNN or OC-SVM+ achieves the best performance on all three datasets.
- OC-NN performance is inconsistent, varying with the ratio of target-class to novel or abnormal images in the evaluation protocol.
- Approximately 5% improvements over OC-NN are reported for Abnormality-1001, UMDAA-02 Face, and FounderType-200.
- VGG16 gives an overall ∼2% improvement over AlexNet, although OC-CNN with AlexNet performs better on FounderType-200.
- OC-SVM+ performs ∼18% better than OC-SVM with AlexNet and ∼17% better with VGG16 on average across datasets.
V. CONCLUSION
The paper proposes a CNN-based one-class classification method using pseudo-negative Gaussian data and binary cross-entropy training. It reports consistent improvements across authentication, abnormality, and novelty detection datasets, while identifying centrally aligned objects as the experimental scope.
- The proposed method introduces pseudo-negative Gaussian data in feature space and trains a CNN with binary cross-entropy loss.
- The method functions both as a standalone one-class classifier and as a target-class feature extractor for OC-SVM+.
- Consistent performance improvements are reported across authentication, abnormality, and novelty detection datasets.
- The experiments use centrally aligned objects, and future work will investigate joint end-to-end detection and classification.