daxiahuang/unsupervised_defect_segmentation

Improving_Unsupervised_Defect_Segmentation

This is Keras code from "Improving Unsupervised Defect Segmentation by Applying Structural Similarity to Autoencoders".
https://arxiv.org/abs/1807.02011
I tried for my computer vision research.
That's is amazing method for unsupervised defect segmentation using AutoEncoder with SSIM.

Usage

0. Install Library

keras >= 2.0
tensorflow >= 1.6
scikit-learn
PIL
matplotlib

1. Use AutoEncoder

You can use your images with AutoEncoder.ipynb.
Please set your Image Path and automatically be resized on this code 128×128×1.
minimum 10 images required

2. Use SSIM

Next time, you can compare Inpue Image and Decoded Image.

If you use SSIM method, you have to pip install SSIM-PIL.

unsupervised_defect_segmentation's Issues

Why do I reproduce the image in black? The original is defect detection. Why don't you see your results?

Question about the train images or dataset

Hi, I've done a research about defect detection recently, and I've noticed your code for the paper. In your code, you changed the 'img_dir' to 'img_dir = ("./textures/texture_1/train/good/")', do you have any dataset about the defect detection？

Looking forward to your reply. Thank you！