MLZoomcamp 2022 Capstone Project

A Convolutional Neural Network to detect cracks on a surface

Description:

Concrete is one of the most used and important materials in civil construction. Our houses, our offices, the bridges that connect cities and a lot of other types of constructions, are made of concrete, or if concrete is not the main material, at least is used in some part of the construction. The major defect detected in civil structures are the surface cracks on concrete. It is important to make inspections in the structure to evaluate the rigidity and the tensile, once these concrete surface cracks can compromise the structure and cause big problems and disasters. Then it is very important to make inspections regularly in the structures in order to detect these cracks in the early stage to avoid major problems in the future. A convolutional neural network could be used to do a classification and an analysis if there is a surface crack in the photos taken during the process of structure inspection.

How to run the application:

1. Docker

Step 1: Clone the GitHub repository with the project.

git clone https://github.com/jvscursulim/mlzoomcamp_capstone_project

Step 2: Access the GitHub repository folder.

cd mlzoomcamp_capstone_project

Step 3: Create the docker image.

docker build -t surface_crack_detection .

Step 4: Run the application with docker.

docker run -p 4242:4242 surface_crack_detection

Step 5: Put some images that you want to classify in imgs folder

Step 6: Change the value associated with the key image_path of the dictionary assigned with the variable img inside test.py file. The new value must be the path of one of the images inside imgs folder. (Please refer to section "How to send data for the application")

2. Without Docker

Step 1: Clone the GitHub repository with the project.

git clone https://github.com/jvscursulim/mlzoomcamp_capstone_project

Step 2: Access the GitHub repository folder.

cd mlzoomcamp_capstone_project

Step 3: Create a virtual environment.

python -m venv env

Step 4: Activate your virtual environment.

• Linux: Activation of the virtual environment.

source env/bin/activate

• Windows: Activation of the virtual environment.

env/Scripts/Activate.ps1

Step 5: Install pipenv.

pip install pipenv

Step 6: Install the packages required for this application using the command below.

pipenv install

Step 7: Run the application with gunicorn.

gunicorn --bind=0.0.0.0:4242 predict:app

Step 8: Put some images that you want to classify in imgs folder

Step 9: Change the value associated with the key image_path of the dictionary assigned with the variable img inside test.py file. The new value must be the path of one of the images inside imgs folder. (Please refer to section "How to send data for the application")

Observation: If you want to train a model

• https://www.kaggle.com/datasets/arunrk7/surface-crack-detection

• Observation: Download the dataset using the link above and extract the files in data/dataset before start the training.

Access script folder.

cd script

Make your changes in train.py and run the file using the command below.

python train.py

After training process, you can build the new application following the instructions in the sections 1. Docker and 2. Without Docker.

How to send data for the application:

For instance the test.py file.

Code snippet

import requests

img = {"image_path": "/workspaces/mlzoomcamp_capstone_project/imgs/00001.jpg"}

url = "http://localhost:4242/predict"
print(requests.post(url, json=img).json())

References:

1. MLZoomcamp Course Material - Session #8 - Deep Learning
2. Surface Crack Detection Dataset