Crack detection has vital importance for structural health monitoring and inspection. We would like to train a network to detect Cracks, we will denote the images that contain cracks as positive and images with no cracks as negative.
Examples from the dataset:
We used different models as below:
- ResNet50
- VGG16
- ResNet18
You will find in the repositry each model & the directories will be divided as follows:
- /models: contains the saved model after training
- /notebooks: contains the notebooks used for training
In each of the architectures mentioned above the fully connected layer was changed to classify the images into two classes (positive & negative) & the activation function we used was softmax.
The dataset used is from IBM capstone project of the professional cerificate AI Engineering
Dataset can be found from the following links:
- Data for ResNet50 & VGG16: link
- Data for ResNet18:
The below table will show the training for each model with the number of epochs & the accuracy achieved.
| Model | Optimizer | Epochs | Accuracy | loss |
|---|---|---|---|---|
| ResNet50 | Adam | 2 | 0.9985 | 0.0061 |
| VGG16 | Adam | 2 | 0.9958 | 0.0188 |
| ResNet18 | Adam | 1 | 0.9943 | 23.88 |
The evaluation was done on ResNet50 & VGG16 models, the evaluation was done on 1000 images for each model
The ResNet50 model achieved an accuracy of 0.99906 & the VGG16 model achieved an accuracy of 0.99624 so overall the ResNet50 model performed better than the VGG16 model.
Special thanks to IBM for providing the dataset & the course for the capstone project.