Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ATAC-Net: Zoomed view works better for Anomaly Detection

Shaurya Gupta, Neil Gautam, Anurag Malyala

TL;DR

This work proposes ATAC-Net, a framework that trains to detect anomalies from a minimal set of known prior anomalies, and introduces attention-guided cropping, which provides a closer view of suspect regions during the training phase.

Abstract

The application of deep learning in visual anomaly detection has gained widespread popularity due to its potential use in quality control and manufacturing. Current standard methods are Unsupervised, where a clean dataset is utilised to detect deviations and flag anomalies during testing. However, incorporating a few samples when the type of anomalies is known beforehand can significantly enhance performance. Thus, we propose ATAC-Net, a framework that trains to detect anomalies from a minimal set of known prior anomalies. Furthermore, we introduce attention-guided cropping, which provides a closer view of suspect regions during the training phase. Our framework is a reliable and easy-to-understand system for detecting anomalies, and we substantiate its superiority to some of the current state-of-the-art techniques in a comparable setting.

ATAC-Net: Zoomed view works better for Anomaly Detection

TL;DR

This work proposes ATAC-Net, a framework that trains to detect anomalies from a minimal set of known prior anomalies, and introduces attention-guided cropping, which provides a closer view of suspect regions during the training phase.

Abstract

The application of deep learning in visual anomaly detection has gained widespread popularity due to its potential use in quality control and manufacturing. Current standard methods are Unsupervised, where a clean dataset is utilised to detect deviations and flag anomalies during testing. However, incorporating a few samples when the type of anomalies is known beforehand can significantly enhance performance. Thus, we propose ATAC-Net, a framework that trains to detect anomalies from a minimal set of known prior anomalies. Furthermore, we introduce attention-guided cropping, which provides a closer view of suspect regions during the training phase. Our framework is a reliable and easy-to-understand system for detecting anomalies, and we substantiate its superiority to some of the current state-of-the-art techniques in a comparable setting.
Paper Structure (10 sections, 12 equations, 5 figures, 2 tables)

This paper contains 10 sections, 12 equations, 5 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Literature Review
  3. Methodology
  4. Weaker-Supervision Heads
  5. Deviation Loss
  6. Experimentation
  7. Datasets
  8. Training Details
  9. Results
  10. Conclusion And Future Works

Figures (5)

  • Figure 1: Grad-CAM grad-cam visualization of a dog class within the original image and zoomed view of the object. The density of the saliency map increases as the model receives a zoomed object
  • Figure 2: ATAC-Net, using attention augmentation module to learn and find anomalies while also providing interpretability of the anomaly via the saliency map. The network’s feature extraction works two-fold to find the best results for anomalies within the provided sample. The normal (un-cropped) and attention-cropped flows are used in train and inference time.
  • Figure 3: Histogram of anomalous and normal samples on a test dataset between SPADE and ATAC-Net
  • Figure 4: t-SNE plots between different training versions of a baseline ResNet-50. ATAC-Net shows the effect of attention-based cropping, which helps better distinguish between anomalies and normal samples.
  • Figure 5: Anomalous regions detected by the attention-cropping mechanism from ATAC-Net. The heatmaps are binarized using a threshold after normalizing for the coordinate points extraction