Automated crack detection

An academic/engineering project focused on automated crack detection in industrial infrastructure (e.g., concrete surfaces) using deep learning.

About the project

To solve the binary classification problem (crack vs. no crack), a Convolutional Neural Network (CNN) was built and trained using the TensorFlow/Keras library.

The dataset was split into three independent parts:

• Training set (80%) - used to train the model.
• Validation set (10%) - used to monitor the training process and prevent overfitting.
• Test set (10%) - used for the final, objective evaluation of the model's performance on completely unseen data.

Dataset Samples

Here are examples of the images used in this project:

No Crack (Negative)	Crack (Positive)

Technologies

• Python
• TensorFlow / Keras
• Matplotlib & Seaborn (Data Visualization)
• Scikit-learn (Evaluation Metrics)

Model results

The model achieved outstanding accuracy on the test set. Below are the detailed metrics, training history, and the final confusion matrix.

Evaluation Metrics

Metric	No Crack (Negative)	Crack (Positive)
Precision	0.99	1.00
Recall (Sensitivity)	1.00	0.99
F1-Score	1.00	1.00

Overall Accuracy: ~99.5% Tested on: 4000 images

Training History (Accuracy and Loss)

Confusion Matrix

How to Run the Project locally

1. Clone this repository.
2. Download the dataset and place the images inside the dataset/ folder (using positive and negative subfolders).
3. Install the required dependencies:

   pip install -r requirements.txt