Weakly supervised surface crack localization using Pytorch

This approach uses a traditional CNN for crack classification and Grad-CAM for crack localization. The whole procedure can be found in this Medium article.

Requirements

• Pytorch (>=1.9)
• torchvision (>=0.4)
• cv2
• pil
• matplotlib

We will create class activation maps to highlight the crack location. To create class activation map using the grad-cam method you need to install the package simply by typing:

pip install grad-cam

Getting started

At first clone this repository and install the required dependencies.
To train the network:

python train.py

An example inference command:

python inference.py

To generate grad-cam visual explanation (heat maps) run the following:

python xai.py

Customizing dataset

The procedure for using your own dataset is very simple. Just prepare the dataset in the same way shown inside the directory folder data/. You can also increase the number of classes as you need. All you need to change the final layer of the architecture.

Sample classification Results

Sample grad-cam visualization / Crack localization