Anomaly Detection Paper 모음

Patch SVDD

• ACCV 2020
• patch-wise detection method on image data (look at small sections of an image)
• self-supervised learning
• randomly initialized encoder of Deep SVDD is also used here, but taken further to study why the random features were separable
• Deep SVDD:
• Patch-wise Deep SVDD:
• self-supervised learning:
• hierarchical encoding:
• generating anomaly maps:

Deep SAD

• ICLR 2020
• deep semi-supervised anomaly detection
• end-to-end deep method for general semi-supervised AD
• generalization of unsupervised Deep SVDD method to the semi-supervised AD setting
• same loss term as Deep SVDD for unlabeled data
• for labeled data then, new loss term weighted by hyperparameter controlling the balance between the labeled and unlabeled term

A Two-Stage Autoencoder For Visual Anomaly Detection

• 2021 IEEE International Conference on Image Processing (ICIP)
• https://ieeexplore.ieee.org/abstract/document/9506538
• Deep Convolutional AutoEncoder (DCAE)

1. single RotNet (like encoder) (RotNet is a self-supervision approach that relies on predicting image rotations as the pretext task in order to learn image representations.) (https://paperswithcode.com/method/rotnet)
2. from decoder, get discriminative representations to train
3. 2 parallel decoders for image reconstruction
4. reconstruction error from 2 decoders are combined to be anomaly score

• this paper is for images, not on time series data
• does not consider temporal data

Choosing Effective Projections for Fast and Accurate Anomaly Detection

• ODD (Object Detection and Description workshop) 2021
• https://oddworkshop.github.io/assets/papers/2.pdf
• Chimera
• consensus-based approach
• select best subset of projections
• requires reliable uni-modal unsupervised statistical test, so there is S-Chimera that removes this requirement

Enhancing Unsupervised Anomaly Detection with Score-Guided Network

• https://arxiv.org/abs/2109.04684
• submitted to IEEE (Sept 2021)
• Score-Guided AutoEncoder (SG-AE)
• has official code
• challenges of unsupervised learning:
• this model has:
• can learn more representations with sufficient information
• not on time series data
• maybe the score function can be extended to time series but how? will calculating at every time step be time-consuming?
• does it work with CNN?

 

Learning and Evaluating Representations for Deep One-class Classification

• https://openreview.net/forum?id=HCSgyPUfeDj
• ICLR 2021
• 2 stages:
• distribution augmented contrastive learning
• data augmentation to prevent uniformity of contrastive representations

 

DOC3-Deep One Class Classification using Contradictions

• https://arxiv.org/abs/2105.07636
• learning from contradictions a.k.a. universum learning
• success depends on tuning hyperparameters
• 2 goals:
• effectiveness also depend on type of universum used

Contrastive Predictive Coding

• https://arxiv.org/abs/2107.07820
• ICML 2021 Workshop
• CPC
• solve the problem of lack of data
• self-supervised representation learning setting
• patch-wise contrastive loss directly become anomaly score
• for images
• patches within an image is contrasted
• can also create anomaly segmentation masks
• 2 adjustments:

Semi-Supervised Anomaly Detection Algorithm Using Probabilistic Labeling (SAD-PL)

• https://ieeexplore.ieee.org/abstract/document/9576706 
• IEEE 2021
• related to Deep SVDD, DROCC, and Deep SAD
• uses LeNet type CNN, just like Deep SVDD
• 2 steps:
• runs until rate of label changes is lower than threshold

DOCC: Deep one-class crop classification via positive and unlabeled learning for multi-modal satellite imagery

• https://www.sciencedirect.com/science/article/pii/S0303243421003056
• International Journal of Applied Earth Observation and Geoinformation 2021
• multi-modal time series satellite images
• deep one-class crop framework:
• input: only samples of 1 target class
• automatically extract features, without labeling based on prior knowledge
• beneficial for large-scale mapping when samples of multi-class are difficult to obtain