About the Application of Autoencoders for Visual Defect Detection

Visual defect detection is a key technology in modern industrial manufacturing systems. There are many possible appearances of product defects, including distortions in color, shape, contamination, missing or superfluous parts. For the detection of those, besides traditional image processing techniques, convolutional neural networks based methods have also appeared to avoid the usage of hand-crafted features and to build more efficient detection mechanisms. In our article we deal with autoencoder convolutional networks (AEs) which do not require examples of defects for training. Unfortunately, the manual and/or trial-and-error design of AEs is still required to achieve good performance, since there are many unknown parameters of AEs which can greatly influence the detection abilities.

Datasets

MVTec Texture 1 and 2 datasets can be downloaded from the following link: https://www.mydrive.ch/shares/46066/8338a11f32bb1b7b215c5381abe54ebf/download/420939225-1629955758/textures.zip

Our dataset can be downloaded from here: https://drive.google.com/file/d/1LOFXTxGGufXJWHQdBbxTfBAzhwpGmL0j/view?usp=drive_link

Requirement

Make sure you have the following dependencies installed:

colorlog>=6.7.0
colorthief>=0.2.1
kornia>=0.7.0
matplotlib>=3.7.1
numpy>=1.23.5
opencv-python>=4.5.5.64
pandas>=2.0.0
sklearn>=0.0.post1
torch>=2.0.0+cu117
torchsummary>=1.5.1
torchvision>=0.15.1+cu117
tqdm>=4.65.0

Installation

First, clone/download this repository. In the const.py file you will find this:

root_mapping = {
    'ricsi': {
        "PROJECT_ROOT": 'D:/AE/storage/',
        "DATASET_ROOT": 'D:/mvtec/'
    }
}

• Update the designated username ('ricsi') to reflect the username associated with your logged-in operating system.
• Utilize PROJECT_ROOT as the central repository for storing essential data such as weights and images, including ROC plots, within your project.
• Employ DATASET_ROOT as the designated directory for managing datasets integral to the functioning of the project.
• const.py will create all the necessary folders.
• Download the datasets and place them into the appropropriate folders.

Usage

In the config.py file, key parameters and settings crucial for the training, testing, and data augmentation processes are centrally stored. These configurations provide a streamlined and organized approach to manage various aspects of the project, ensuring adaptability and ease of customization.

If you wish to train the network with cropped images, the following steps must be executed:

• In config.py, in the ConfigAugmentation() class, set do_augmentation to True, select the desired dataset_type, and set up the augmentation parameters.
• Run augmentation.py.
• If you wish to train the denoising autoencoder, run draw_rectangles.py as well. This script will create images with rectangles.
• Back to const.py, in the ConfigTraining() class, set up the desired parameters. Make sure that crop_it is set to True.
• Run train_ae.py or train_dae.py.
• After training, set up the parameters in ConfigTesting(), and run test.py

On the other hand, to execute the workflow without cropped images, skip do_augmentation and set crop_it to False.

Screenshots

Below are some pictures of how the program works.

ROC curve:

Training visualization:

Reconstruction visualization:

More info

Our paper is accessable here: http://wscg.zcu.cz/WSCG2021/FULL/I79.pdf

References

Our article is mainly inspired by the following article: @inproceedings{inproceedings, author = {Bergmann, Paul and Löwe, Sindy and Fauser, Michael and Sattlegger, David and Steger, Carsten}, year = {2019}, month = {01}, pages = {372-380}, title = {Improving Unsupervised Defect Segmentation by Applying Structural Similarity to Autoencoders}, doi = {10.5220/0007364503720380} }

The programs are based on this repo: https://github.com/plutoyuxie/AutoEncoder-SSIM-for-unsupervised-anomaly-detection-/tree/master