Anomaly Detection

Overview

Welcome to the Anomaly Detection Research Project repository. This project aims to advance the field of anomaly detection in technical applications using innovative techniques, including transfer learning, SIFT-FLANN, and cosine similarity. The project's primary goal is to develop a dynamic and responsive framework for online adaptive anomaly detection.

Background

The field of anomaly detection is crucial for various technical applications, from manufacturing quality control to real-time surveillance. Traditional methods often rely on rule-based thresholds or supervised learning, which can be impractical due to the need for labeled data. This project explores unsupervised approaches, such as transfer learning, SIFT-FLANN, and cosine similarity, to overcome these limitations and enhance anomaly detection.

Research Goals

• Develop an online adaptive anomaly detection framework.
• Leverage transfer learning principles to enhance adaptability.
• Integrate SIFT-FLANN and cosine methods with transfer learning.
• Use of MVG method as normality model.
• Compare the proposed framework with the state of the art.
• Explore the robustness and adaptability of the framework in various technical environments.

Repository Structure

The repository is organized as follows:

• Cosine.py and SIFT_FLANN.py: Contains the source code for the anomaly detection methods.
• NN.py: Includes Neural networks used in experiments.
• Main.py: This is the main execution file to be executed.
• visualization.py: This file generates results after the algorithm is executed completely.
• requirements.txt: Contains supporting libraries for creating virtual environment.

Execution

To get started with this project, follow these steps:

1. Clone the repository to your local machine.

2. Create a virtual environment using requirements.txt

3. Datasets: The datasets used for the experimetns can be downloaded from the sources mentioned below:

   • Mockup
   • ISS Panel
   • MvTec
   • VisA

4. From the virtual environment execute the Main.py with following arguments:

   • -m, --anomaly_detection_method - Select either C for Cosine or SF for SIFT-FLANN method.
   • -f, --pretrained_features - Select either y to load pretrained features or n to compute train image dataset features
   • -d, --train_data_path - Provide the path containing images to be trained (/path/to/images/'.jpg', '.png', '.JPG')
   • -t, --test_data_path - Provide the path containing images to be tested for anoamly detection (/path/to/images/'.jpg', '.png', '.JPG')
   • -r, --result_path - Select the directory to save the generated results by proving the path to it.
   • (optional) -v, --visualize_detection_on_off - Choose either 0 == 'on' and 1 == 'off' to visualize the detections while execution of the algorithm.

   (venv)$python3 Main.py -m C -d /path/to/images/*.jpg -t /path/to/images/*.jpg -r /directory/to/save/the/results -v 0
   

(optional)

5. To visualize the generated results with classification report and compare true vs predictions made, run the visualization.py with following arguments:

• -t, --test_data_path - Provide the path containing images to be tested for anoamly detection (/path/to/images/'.jpg', '.png', '.JPG'
• -d, --true_value_path - Provide the path to .csv where true values are stored for the test data.
• -r, --results_path - Provide the path for the folder containing generated results.

   (venv)$python3 visualization.py -t /path/to/images/*.jpg -d /path/to/true/values/.csv -r /path/conatining/folder/of/results.

Testing

To check the correctness of the code, following conditions should be met:

• The executes without any errors.
• The Main.py when executed with correct arguments should display the following results:
  1. Result : Anomaly/ No-Anomaly
  2. Computation time per frame:
  3. Average % data saved from training for every frame:
  4. Average computation time per frame:

This also saves the results generated in the results folder. This folder will contain masked image results and predicted values in a csv.

• (optional) visualization.py when executed with correct arguments should display the following results:
  1. Classification report
  2. A window opens displaying the true image, detected image and the live graph of true vs predicted values.

License

The code is distributed under the 3-Clause BSD license

Citation

DOI: https://doi.org/10.1109/CASE59546.2024.10711376

@INPROCEEDINGS{10711376, author={Shete, Siddhant and Mronga, Dennis and Jadhav, Ankita and Kirchner, Frank}, booktitle={2024 IEEE 20th International Conference on Automation Science and Engineering (CASE)}, title={Online-Adaptive Anomaly Detection for Defect Identification in Aircraft Assembly}, year={2024}, volume={}, number={}, pages={4126-4133}, keywords={Training;Adaptation models;Accuracy;Computational modeling;Transfer learning;Benchmark testing;Feature extraction;Aircraft manufacture;Aircraft;Anomaly detection}, doi={10.1109/CASE59546.2024.10711376}}

Maintainer / Authors / Contributers

• Siddhant Shete "Online_Adaptive_Anomaly_Detection_for_Defect_Identification_in_Aircraft_Assembly" was initiated in SeMoSys project and is currently developed at the Robotics Innovation Center of the German Research Center for ArtificialIntelligence (DFKI) in Bremen.

"Online_Adaptive_Anomaly_Detection_for_Defect_Identification_in_Aircraft_Assembly" has been funded by German Federal Ministry of Economic Affairs and Climate Action (BMWK, grant number 20W1922F).

Copyright 2023, Siddhant Shete, DFKI RIC