Abstract
We propose a deep learning-based solution for the problem of feature learning in one-class classification. The proposed method operates on top of a Convolutional Neural Network (CNN) of choice and produces descriptive features while maintaining a low intra-class variance in the feature space for the given class. For this purpose two loss functions, compactness loss and descriptiveness loss are proposed along with a parallel CNN architecture. A template matching-based framework is introduced to facilitate the testing process. Extensive experiments on publicly available anomaly detection, novelty detection and mobile active authentication datasets show that the proposed Deep One-Class (DOC) classification method achieves significant improvements over the state-of-the-art.
Introduction: The Challenge of One-Class Classification
One-class classification (OCC) is a fundamental problem in machine learning where the goal is to identify whether a new observation belongs to the same distribution as the training data, which consists solely of "normal" examples. Unlike standard multi-class classification, there are no examples of "abnormal" or "alien" classes during training.
Standard Deep Learning thrives on discriminative training—learning to identify a cat by seeing how it differs from a dog. When a Convolutional Neural Network (CNN) is trained on only one class, it often converges to a "trivial solution" (e.g., all outputs are zero) or produces features that are too generic to distinguish the target class from visually similar but different objects.
Methodology: The Deep One-Class (DOC) Framework
To overcome these limitations, the authors introduce Deep One-Class (DOC) classification. The core innovation is a One-class Transfer Learning strategy that utilizes external, unrelated multi-class data to guide the feature learning process.
The Dual-Network Architecture
The DOC framework employs a parallel CNN architecture with shared weights. This setup allows the model to learn from two different data streams simultaneously:
- Descriptiveness Loss ($l_D$): One branch processes a standard multi-class dataset (e.g., ImageNet). By optimizing a traditional cross-entropy loss, the network is forced to learn "descriptive" features that can distinguish between a wide variety of objects. This prevents the network from collapsing into a trivial state.
- Compactness Loss ($l_C$): The second branch processes the target "one-class" data. A novel compactness loss function minimizes the intra-class variance, forcing the features of the "normal" class to cluster tightly in the feature space.
Inference via Template Matching
At test time, the network extracts features from a small set of "normal" training samples to serve as templates. A test image is then classified based on its Euclidean distance to these templates; if the distance exceeds a threshold, it is flagged as "abnormal."
Experimental Evaluation
The researchers evaluated DOC across three distinct domains:
- Abnormal Image Detection: Using the 1001 Abnormal Objects Dataset, which contains categories like airplanes and chairs with "abnormal" counterparts.
- Novelty Detection: Using Caltech 256, treating one category as "normal" and others as "alien."
- Mobile Active Authentication: Using the UMDAA-02 dataset of mobile face images to distinguish a device owner from intruders.
External "world knowledge" for the descriptiveness loss was provided by datasets such as ILSVRC12 (ImageNet), Places365, and Oxford Flowers 102.
Key Results and Insights
The DOC method demonstrated significant improvements over state-of-the-art (SOTA) methods:
- Abnormal Detection: DOC (VGG16) achieved an AUC of 0.956, surpassing previous graphical models (0.911) and standard OCNN (0.888).
- Novelty Detection: On Caltech 256, DOC reached a near-perfect 0.999 AUC for specific classes like "American Flag."
- Active Authentication: On the UMDAA-02 dataset, DOC achieved 0.810 AUC, a substantial leap from the ~0.60 range typical of standard one-class SVMs.
A critical insight from the study is the necessity of a "world view." By using an unrelated multi-class dataset, the model gains the context needed to understand what makes the target class unique. Furthermore, the authors found that features from the final layers (e.g., fc8) are most effective for OCC, as they capture high-level semantic information.
Conclusion
Deep One-Class (DOC) classification provides a robust framework for feature learning when only one class is available. By balancing descriptiveness and compactness through a dual-loss parallel architecture, it avoids the pitfalls of traditional deep OCC methods. Future work may explore alternative distance metrics for template matching to further refine the classification boundaries.
Summary updated