Flame Guardian: Developing a Deep Learning-Based Fire Detection System

Introduction

Imagine waking up to the smell of smoke, heart racing as you ensure your family’s safety. Early detection is crucial, and “Flame Guardian,” a deep learning-powered fire detection system, aims to make a life-saving difference. This article guides you through creating this technology using CNNs and TensorFlow, from data gathering and augmentation to model construction and fine-tuning. Whether you’re a tech enthusiast or a professional, discover how to leverage cutting-edge technology to protect lives and property.

Learning Outcomes

• Gain skills in preparing, organizing, and augmenting image datasets to optimize model performance.
• Learn how to construct and fine-tune convolutional neural networks for effective image classification tasks.
• Develop the ability to assess and interpret model performance using metrics and visualizations.
• Learn how to deploy and adapt DL(Deep Learning) models for practical applications, demonstrating their utility in real-world problems like fire detection.

This article was published as a part of the Data Science Blogathon.

Revolution of Deep Learning in Fire Detection

In recent times, the Deep Learning has revolutionized  colorful fields, from healthcare to finance, and now, it’s making strides in safety and disaster operations. One particularly  instigative operation of Deep Learning is in the realm of fire discovery. With the adding frequency and inflexibility of backfires worldwide, developing an effective and dependable fire discovery system is more pivotal than ever. In this comprehensive companion, we will walk you through the process of creating an important fire discovery system using convolutional neural networks( CNNs) and TensorFlow. This system, aptly named” Flame Guardian,” aims to identify fire from images with high delicacy, potentially abetting in early discovery and forestallment of wide fire damage.

Fires, whether wildfires or structural fires pose a significant threat to life, property, and the environment. Early detection is critical in mitigating the devastating effects of fires. Deep-Learning based fire detection systems, can analyze vast amounts of data quickly and accurately, identifying fire incidents before they escalate.

Challenges in Fire Detection

Detecting fire using Deep Learning presents several challenges:

• Data Variability: Fire images can vary greatly in terms of color, intensity, and surrounding environment. A robust detection system must be able to handle this variability.
• False Positives: It’s crucial to minimize false positives (incorrectly identifying non-fire images as fire) to avoid unnecessary panic and resource deployment.
• Real-Time Processing: For practical use, the system should be able to process images in real-time, providing timely alerts.
• Scalability: The system should be scalable to handle large datasets and work across different.

Dataset Overview

The dataset used for the Flame Guardian fire detection system comprises images categorized into two classes: “fire” and “non-fire.” The primary purpose of this dataset is to train a convolutional neural network (CNN) model to accurately distinguish between images that contain fire and those that do not.

Composition of Fire and Non-Fire Images

• Fire Images : These images contain various scenarios where fire is present. The dataset includes images of wildfires, structural fires, and controlled burns. The fire in these images may vary in size, intensity, and the environment in which it is present. This diversity helps the model learn the different visual characteristics of fire.
• Non-Fire Images : These images do not contain any fire. They include a wide range of scenarios such as landscapes, buildings, forests, and other natural and urban environments without any fire. The inclusion of diverse non-fire images ensures that the model does not falsely identify fire in non-fire situations.

You can download the dataset from here.

Setting Up the Environment

 First, we need to set up our terrain with the necessary libraries and tools. We will be using Google Collab for this design, as it provides a accessible platform with GPU support. We’ve formerly downloaded the dataset and uploaded it on drive. 

#Mount drive
from google.colab import drive
drive.mount('/content/drive')

#Importing necessary Libraries
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import seaborn as sns
import plotly.express as px
import plotly.graph_objects as go
from plotly.subplots import make_subplots
import os
import tensorflow as tf
from tensorflow.keras.preprocessing import image
from tensorflow.keras.preprocessing.image import ImageDataGenerator


#setting style grid 
sns.set_style('darkgrid')

Data Preparation

We require a dataset with pictures of fire and non-fire scripts in order to train our algorithm. A blank DataFrame and a function to add images from our Google Drive to it will be created.

# Create an empty DataFrame
df = pd.DataFrame(columns=['path', 'label'])

# Function to add images to the DataFrame
def add_images_to_df(directory, label):
    for dirname, _, filenames in os.walk(directory):
        for filename in filenames:
            df.loc[len(df)] = [os.path.join(dirname, filename), label]

# Add fire images
add_images_to_df('/content/drive/MyDrive/Fire/fire_dataset/fire_images', 'fire')

# Add non-fire images
add_images_to_df('/content/drive/MyDrive/Fire/fire_dataset/non_fire_images', 'non_fire')

# Shuffle the dataset
df = df.sample(frac=1).reset_index(drop=True)

Visualizing the Distribution of Images

Visualizing the distribution of fire and non-fire images helps us understand our dataset better. We’ll use Plotly for interactive plots.

Creating a Pie Chart for Image Distribution

Let us now create a pie chart for image distribution.

# Create the scatter plot
fig = px.scatter(
    data_frame=df,
    x=df.index,
    y='label',
    color='label',
    title='Distribution of Fire and Non-Fire Images'
)

# Update marker size
fig.update_traces(marker_size=2)

fig.add_trace(go.Pie(values=df['label'].value_counts().to_numpy(), labels=df['label'].value_counts().index, marker=dict(colors=['lightblue','pink'])), row=1, col=2)

Displaying Fire and Non-Fire Images

Let us now write the code for displaying fire and non-fire images.

def visualize_images(label, title):
    data = df[df['label'] == label]
    pics = 6  # Set the number of pics
    fig, ax = plt.subplots(int(pics // 2), 2, figsize=(15, 15))
    plt.suptitle(title)
    ax = ax.ravel()
    for i in range((pics // 2) * 2):
        path = data.sample(1).loc[:, 'path'].to_numpy()[0]
        img = image.load_img(path)
        img = image.img_to_array(img) / 255
        ax[i].imshow(img)
        ax[i].axes.xaxis.set_visible(False)
        ax[i].axes.yaxis.set_visible(False)
visualize_images('fire', 'Images with Fire')
visualize_images('non_fire', 'Images without Fire')

By displaying some sample images from both fire and non-fire categories we would get a sense of what our model will be working with.

Enhancing Training Data with Augmentation Techniques

We’re going to apply image addition ways to ameliorate our training data. Applying arbitrary image adaptations, similar as gyration, drone, and shear, is known as addition. By generating a more robust and different dataset, this procedure enhances the model’s capacity to generalize to new images.

from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Conv2D, MaxPool2D, Flatten, Dense

generator = ImageDataGenerator(
    rotation_range= 20,
    width_shift_range=0.1,
    height_shift_range=0.1,
    shear_range = 2,
    zoom_range=0.2,
    rescale = 1/255,
    validation_split=0.2,
)
train_gen = generator.flow_from_dataframe(df,x_col='path',y_col='label',images_size=(256,256),class_mode='binary',subset='training')
val_gen = generator.flow_from_dataframe(df,x_col='path',y_col='label',images_size=(256,256),class_mode='binary',subset='validation')
class_indices = {}
for key in train_gen.class_indices.keys():
    class_indices[train_gen.class_indices[key]] = key
    
print(class_indices)

Visualizing Augmented Images

We can visualize some of the augmented images generated by our training set.

sns.set_style('dark')
pics = 6  # Set the number of pics
fig, ax = plt.subplots(int(pics // 2), 2, figsize=(15, 15))
plt.suptitle('Generated images in training set')
ax = ax.ravel()
for i in range((pics // 2) * 2):
    ax[i].imshow(train_gen[0][0][i])
    ax[i].axes.xaxis.set_visible(False)
    ax[i].axes.yaxis.set_visible(False)

Constructing the Fire Detection Model

Our model will correspond of several convolutional layers, each followed by a maximum- pooling subcaste. Convolutional layers are the core structure blocks of CNNs, allowing the model to learn spatial scales of features from the images. Max- pooling layers help reduce the dimensionality of the point maps, making the model more effective. We will also add completely connected( thick) layers towards the end of the model. These layers help combine the features learned by the convolutional layers and make the final bracket decision. The affair subcaste will have a single neuron with a sigmoid activation function, which labors a probability score indicating whether the image contains fire. After defining the model armature, we’ll publish a summary to review the structure and the number of parameters in each subcaste. This step is important to insure that the model is rightly configured. 

from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Conv2D, MaxPool2D, Flatten, Dense

model = Sequential()
model.add(Conv2D(filters=32,kernel_size = (2,2),activation='relu',input_shape = (256,256,3)))
model.add(MaxPool2D())
model.add(Conv2D(filters=64,kernel_size=(2,2),activation='relu'))
model.add(MaxPool2D())
model.add(Conv2D(filters=128,kernel_size=(2,2),activation='relu'))
model.add(MaxPool2D())
model.add(Flatten())
model.add(Dense(64,activation='relu'))
model.add(Dense(32,activation = 'relu'))
model.add(Dense(1,activation = 'sigmoid'))
model.summary()

Compiling the Model with Optimizers and Loss Functions

Next, we’ll compile the model using the Adam optimizer and the binary cross-entropy loss function. The Adam optimizer is widely used in deep learning for its efficiency and adaptive learning rate. Binary cross-entropy is appropriate for our binary classification problem (fire vs. non-fire).

We’ll also specify additional metrics, such as accuracy, recall, and area under the curve (AUC), to evaluate the model’s performance during training and validation.

Adding Callbacks for Optimal Training

Callbacks are a powerful feature in TensorFlow that allows us to monitor and control the training process. We’ll use two important callbacks:

• EarlyStopping: Stops training when the validation loss stops improving, preventing overfitting.
• ReduceLROnPlateau: Reduces the learning rate when the validation loss plateaus, helping the model converge to a better solution.

#Compiling Model
from tensorflow.keras.metrics import Recall,AUC
from tensorflow.keras.utils import plot_model

model.compile(optimizer='adam',loss='binary_crossentropy',metrics=['accuracy',Recall(),AUC()])

#Defining Callbacks
from tensorflow.keras.callbacks import EarlyStopping, ReduceLROnPlateau
early_stoppping = EarlyStopping(monitor='val_loss',patience=5,restore_best_weights=True)
reduce_lr_on_plateau = ReduceLROnPlateau(monitor='val_loss',factor=0.1,patience=5)

Model Fitting: Training the Convolutional Neural Network

Model fitting refers to the process of training a machine learning model on a dataset. During this process, the model learns the underlying patterns in the data by adjusting its parameters (weights and biases) to minimize the loss function. In the context of deep learning, this involves several epochs of forward and backward passes over the training data.

model.fit(x=train_gen,batch_size=32,epochs=15,validation_data=val_gen,callbacks=[early_stoppping,reduce_lr_on_plateau])

Evaluating the Model

After training, we’ll evaluate the model’s performance on the validation set. This step helps us understand how well the model generalizes to new data. We’ll also visualize the training history to see how the loss and metrics evolved over time.

eval_list = model.evaluate(val_gen,return_dict=True)
for metric in eval_list.keys():
    print(metric+f": {eval_list[metric]:.2f}")
   
eval_list = model.evaluate(val_gen,return_dict=True)
for metric in eval_list.keys():
    print(metric+f": {eval_list[metric]:.2f}")

Example Usage: Predicting Fire in New Images

Finally, we’ll demonstrate how to use the trained model to predict whether a new image contains fire. This step involves loading an image, preprocessing it to match the model’s input requirements, and using the model to make a prediction.

Downloading and Loading the Image

We’ll download a sample image from the internet and load it using TensorFlow’s image processing functions. This step involves resizing the image and normalizing its pixel values.

Making the Prediction

Using the trained model, we’ll make a prediction on the loaded image. The model will output a probability score, which we’ll round to get a binary classification (fire or non-fire). We’ll also map the prediction to its corresponding label using the class indices.

# Downloading the image
!curl https://static01.nyt.com/images/2021/02/19/world/19storm-briefing-texas-fire/19storm-briefing-texas-fire-articleLarge.jpg --output predict.jpg
#loading the image
img = image.load_img('predict.jpg')
img

img = image.img_to_array(img)/255
img = tf.image.resize(img,(256,256))
img = tf.expand_dims(img,axis=0)

print("Image Shape",img.shape)

prediction = int(tf.round(model.predict(x=img)).numpy()[0][0])
print("The predicted value is: ",prediction,"and the predicted label is:",class_indices[prediction])

Conclusion

Developing an Deep Learning-based fire detection system like “Flame Guardian” exemplifies the transformative potential of Deep Learning in addressing real-world challenges. By meticulously following each step, from data preparation and visualization to model building, training, and evaluation, we have created a robust framework for detecting fire in images. This project not only highlights the technical intricacies involved deep learning but also emphasizes the importance of leveraging technology for safety and disaster prevention.

As we conclude, it’s evident that DL Model can significantly enhance fire detection systems, making them more efficient, reliable, and scalable. While traditional methods have their merits, the incorporation of Deep Learning introduces a new level of sophistication and accuracy. The journey of developing “Flame Guardian” has been both enlightening and rewarding, showcasing the immense capabilities of modern technologies.

Key Takeaways

• Understood Data handling and Visualization techniques.
• Understood proper data collection and augmentation ensure effective model training and generalization.
• Implemented Model Building and Model Evaluation.
• Understood Callbacks like EarlyStopping and ReduceLROnPlateau to optimize training and prevent overfitting.
• Learnt Building Convolutional Neural Network For Fire Detection using CNN.

Frequently Asked Questions

The media shown in this article is not owned by Analytics Vidhya and is used at the Author’s discretion.

Atharva Chandak

I'm Atharva Chandak, a Data Science Enthusiast working as a Junior Data Scientist at Kamjritztex Technologies and a student at Savitribai Phule University. I'm passionate about data analysis and machine learning. Excited to share insights and learn from this community!