Version 0.1
https://share.streamlit.io/nedraki/concrete_cracks_detention/main/app.py
Identification of concrete cracks in images with the application of convolutional neural networks.
To identify cracks on images taken for concrete inspection in structures.
The construction of the model and training data is currently being evaluated with a main focused in the following datasets:
The current version has been trained with more than 40K images combining the datasets from Utah-State university and Kaggle. The first training achieved and accuracy of 86% on the validation test and no signs of overfitting were identified.
In order to improve the accuracy metrics, the following actions were applied to the first trained model:
- Data augmentation techniques: Rotation and zoom on available images
- Adding dropout layer: To avoid presence of overfitting
- Adjusting learning rate
- Transfer learning: Using mobilenetV2 pre-trained model.
-
Increase the pixels size of images used for training: The current version is trained reducing the images to 80x80 pixels (Due to limited access to computational power). Higher quality on images might lead to better results.
-
Improve your dataset: The dataset can be enriched with more data helpful to build a more generic model.
-
Tuning Epochs, Batch size and learning rate: After 30 epochs, the training did not report increases on the accuracy for a batch size = 128. However the tuning of this parameters will remain essential to achieved better results and balance use of computational resources.
-
Evaluate transfer learning with another pre-trained model and its influence in the accuracy.
-
Google collab is a good friend to run the training notebook.