Performs transfer learning to classify images of concrete as having crack or not, using ResNet50 as base model. Dataset is obtained from Mendeley Data.

• Dataset is split into train, validation and test sets with ratio of 70:24:6.
• Data augmentation layer is applied to further diversify the dataset.
• Applied augmentation layers are RandomFlip, RandomRotation, RandomContrast and RandomZoom.
• Augmented data is then preprocessed by tensorflow.keras.applications.resnet50.preprocess_input layer.

• 3 additional layers (GlobalAveragePooling2D, Dropout, Dense) were appended after the base model layer:
  • GlobalAveragePooling2D layer to aggregate width and height into 1 dimension. Transforming the last layer of the base model from shape of (32, 5, 5, 2048) to (32, 2048).
  • Dropout layer with rate of 0.2 to reduce chances of overfitting.
  • Final Dense layer with 2 nodes as output layer.
• Base model layer was first freezed to train the above appended layers.
• Early stopping callback is applied during this initial training. Training stopped at epoch 2 with val_accuracy: 0.9982.
• Next, 1 out of 175 layers of base model were unfreezed for fine tuning.
• Again, early stopping was also applied to fine tuning. Training stopped at epoch 2 with no improvement in validation accuracy.

2018 – Özgenel, Ç.F., Gönenç Sorguç, A. “Performance Comparison of Pretrained Convolutional Neural Networks on Crack Detection in Buildings”, ISARC 2018, Berlin.

Lei Zhang , Fan Yang , Yimin Daniel Zhang, and Y. J. Z., Zhang, L., Yang, F., Zhang, Y. D., & Zhu, Y. J. (2016). Road Crack Detection Using Deep Convolutional Neural Network. In 2016 IEEE International Conference on Image Processing (ICIP). http://doi.org/10.1109/ICIP.2016.7533052