Surface Crack Detection

About

This project is a deep learning model to detect cracks on civil engineering building elements. The model is based on the U-Net architecture and SAM (Segment Anything Model) loss function. The dataset used to train the model is the Concrete Crack Images for Classification dataset.

Getting Started

To run the project, you need to follow the steps below:

Installation

    $ git clone
    $ cd surface_crack_detection

Prerequisites

What things you need to have to be able to run:

• Python 3.11 +
• Pip 3+
• VirtualEnvWrapper is recommended but not mandatory

Requirements

    $ pip install -r requirements.txt

If you want to use Segmentation Anything Model (SAM), you must create another virtual environment:

    $ pip install -r requirements-torch.txt

Running the project

Training the model

You can train your own model (classification or segmentation) by running the script below. Each script is associated with a different model.

Type of model	Model	Script
Classification	Resnet50	python surface_crack_detection/models/resnet.py
Classification	VGG16	python surface_crack_detection/models/vgg.py
Classification	InceptionV3	python surface_crack_detection/models/inception.py
Segmentation	U-Net	python surface_crack_detection/crack_segmentation/classes/train_evaluate.py
Segmentation and Classification	U-Net-Resnet50	python surface_crack_detection/models/unet_resnet50.py
Segmentation and Classification	U-Net-Mobilnet	python surface_crack_detection/models/unet_mobilenet.py
Segmentation and Classication	SAM-Resnet50	python surface_crack_detection/models/sam_resnet50.py

We have more three models that classify a crack image in isolated or disseminated:

• CNN model:

    $ python surface_crack_detection/classification/models/cnn.py

• InceptionV3 model:

    $ python surface_crack_detection/classification/models/inception.py

• ResNet50 model:

    $ python surface_crack_detection/classification/models/resnet.py

Getting prediction

U-Net-MobileNet

If you want to segment and classify an image with our trained model:

1. You must set the input directory that contains the images.
2. You can change the output directory, but by default the images will save in surface_crack_detection/image_output directory. (optional)
3. Run the script:

    $ python surface_crack_detection/models/model_predictions.py

By default, we use U-Net-Mobilenet model. The output of this script will save the segmented image on your device and classify it as either having a crack or not.

Classification models

You can also input an image and see whether it has a crack (positive) or not (negative) by running the script below:

    $ python ./surface_crack_detection/predictions.py <model_name> <image_path>

Models available: cnn, inception, resnet50 and vgg.

Publications related to this project

H. C. Dantas, L. M. G. Morais, P. H. A. Bezerra and R. C. B. Rego, "Concrete Crack Detection Using Embedded Machine Learning," 2024 8th International Symposium on Instrumentation Systems, Circuits and Transducers (INSCIT), Joao Pessoa, Brazil, 2024, pp. 1-6, doi: 10.1109/INSCIT62583.2024.10693377.

Bezerra, P. H. A., H. C. Dantas, L. M. G. Morais, and R. C. B. Rego. "A Deep Learning Artificial Intelligence Algorithm to Detect Cracks on Civil Engineering Building Elements." In: XX International Conference on Building Pathology and Constructions Repair, 2024, Fortaleza. XX International Conference on Building Pathology and Constructions Repair. Fortaleza/CE, 2024. v. 1.