Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
97 tokens/sec
GPT-4o
53 tokens/sec
Gemini 2.5 Pro Pro
44 tokens/sec
o3 Pro
5 tokens/sec
GPT-4.1 Pro
47 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

Deep Anomaly Detection Using Geometric Transformations (1805.10917v2)

Published 28 May 2018 in cs.LG and stat.ML

Abstract: We consider the problem of anomaly detection in images, and present a new detection technique. Given a sample of images, all known to belong to a "normal" class (e.g., dogs), we show how to train a deep neural model that can detect out-of-distribution images (i.e., non-dog objects). The main idea behind our scheme is to train a multi-class model to discriminate between dozens of geometric transformations applied on all the given images. The auxiliary expertise learned by the model generates feature detectors that effectively identify, at test time, anomalous images based on the softmax activation statistics of the model when applied on transformed images. We present extensive experiments using the proposed detector, which indicate that our algorithm improves state-of-the-art methods by a wide margin.

User Edit Pencil Streamline Icon: https://streamlinehq.com
Authors (2)
  1. Izhak Golan (3 papers)
  2. Ran El-Yaniv (44 papers)
Citations (581)

Summary

  • The paper presents a geometric transformation-based method that significantly outperforms traditional anomaly detection techniques in deep neural networks.
  • It trains a multi-class classifier on self-labeled datasets derived from flips, translations, and rotations to learn intrinsic features of normal images.
  • Experimental results on CIFAR-10 and CatsVsDogs show AUROC improvements up to 67%, highlighting the method's effectiveness over existing benchmarks.

Deep Anomaly Detection Using Geometric Transformations

The paper "Deep Anomaly Detection Using Geometric Transformations" by Izhak Golan and Ran El-Yaniv addresses the challenge of anomaly detection within image datasets, particularly focusing on how deep learning models can be trained to effectively identify out-of-distribution instances. The core contribution of this paper is a novel detection technique that employs geometric transformations to enhance the learning capabilities of deep neural networks, thereby improving anomaly detection performance over existing methods.

Overview and Methodology

Anomaly detection is crucial for applications where unanticipated inputs may pose significant risks. The authors propose a method that circumvents traditional techniques such as autoencoders and generative adversarial networks (GANs), which typically rely on reconstruction errors or low-density reconstructions for anomaly scoring. Instead, the proposed approach utilizes geometric transformations to facilitate the learning of distinct features by deep neural models. This method involves training a multi-class classifier over a self-labeled dataset derived from transformed images of a given normal class.

The geometric transformations comprise compositions of flips, translations, and rotations. By training the classifier to recognize these transformations, the network implicitly learns features intrinsic to the normal class, thus enhancing its ability to differentiate anomalies. At test time, each image is subject to the learned transformations, and the normality score is calculated based on softmax activation statistics using a Dirichlet distribution model.

Experimental Evaluation

The authors conduct extensive experiments on benchmarks such as CIFAR-10, CIFAR-100, Fashion-MNIST, and the CatsVsDogs dataset. The proposed method demonstrates substantial improvements over state-of-the-art techniques, measured in terms of Area Under the Receiver Operating Characteristic Curve (AUROC) and Area Under the Precision-Recall Curve (AUPR). For instance, on CIFAR-10, the method increases the AUROC of the top-performing baseline by an average of 32%, while on the more complex CatsVsDogs dataset, it delivers a 67% improvement.

Implications and Future Prospects

The results indicate that the proposed transformation-based method consistently outperforms traditional anomaly detection approaches across varied image datasets. This signifies a promising direction for the development of more robust and adaptive machine vision systems, which can efficiently handle outlier detection in real-world applications.

The theoretical underpinnings of why certain transformations enhance feature learning remain an open area for exploration. Future work could focus on optimizing transformation sets relative to specific dataset characteristics or expanding the method to other domains such as text or audio. Additionally, the integration of this technique into broader AI systems could yield improvements in fields like open-world learning and uncertainty estimation, where reliable detection of unknown inputs is paramount.

In conclusion, the paper presents a well-articulated, experimentally validated method for improving anomaly detection in complex image datasets through innovative use of geometric transformations. This advancement holds significant promise for the continued evolution of intelligent systems capable of autonomously managing unforeseen inputs.