Defects localization in images using deep learning-based classification with CAM output

This repository includes code models used in IDAACS'2023 conference article 'Defects localization in images using deep learning-based classification with CAM output' by Rytis Augustauskas, Lukas Zabulis, Arūnas Lipnickas, Simas Jokubauskas

Modified image classification architectures are used with multihead output to predict and roughly segment images without pixel-wise annotation. It is making predictions in a single iteration.

The model is trained as binary classified, then converted to a 2-output model in which one output is for classification score, the other for CAM explainability.

The following CAM explainability map (used for segmentation) can be generated using this approach:

Dataset	CAM visualization
BSD
Oliena
PCB

Check the article for more details!

Launch

1. Install requirements (Tensorflow 2.12, OpenCV 4.7.0.72, Matplotlib 3.7.1, Albumentations 1.3.0). Optionally, use the following command:

pip install -r requirements.txt

2. Check the provided Jupyter Notebook for training, model conversion, inference and plotting routines. At the end of the training (overfit) on sample data (Oliena), some images will be rendered:

Image	Image+defectCAM

3. Use your data to get new results!

Aknowledgement

Research sponsored and conference expenses are covered by Agmis

Citing

@INPROCEEDINGS{10348813,
  author={Augustauskas, Rytis and Zabulis, Lukas and Lipnickas, Arūnas and Jokubauskas, Simas},
  booktitle={2023 IEEE 12th International Conference on Intelligent Data Acquisition and Advanced Computing Systems: Technology and Applications (IDAACS)}, 
  title={Defects Localization in Images Using Deep Learning-Based Classification with CAM Output}, 
  year={2023},
  volume={1},
  number={},
  pages={487-492},
  doi={10.1109/IDAACS58523.2023.10348813}}

The article ('Defects localization in images using deep learning-based classification with CAM output' by Rytis Augustauskas, Lukas Zabulis, Arūnas Lipnickas, Simas Jokubauskas) was presented at the 12th IEEE International Conference on Intelligent Data Acquisition and Advanced Computing Systems: Technology and Applications, in Dortmund, Germany on September 7-9th, 2023