FSMCNN-Based Accurate Fire Detection Using a Self-Modulative Convolutional Neural Network

1. Introduction

     Fire detection plays a crucial role in preventing disasters and minimizing damage to human life, property, and the environment. Traditional fire detection systems mainly rely on physical sensors such as smoke detectors and heat sensors. However, these methods may not perform effectively in large open environments such as forests, industrial areas, and public spaces.

     With recent advancements in computer vision and deep learning, image-based fire detection systems have become increasingly popular. Convolutional Neural Networks (CNNs) are widely used for image classification tasks due to their ability to automatically learn discriminative features from visual data.

     This work proposes a deep learning-based fire detection framework using a Self-Modulative Convolutional Neural Network (FSMCNN). The system learns fire-related visual features from images and classifies them into Fire or Neutral categories.

2. Problem Statement

Detecting fire from images presents several challenges:

• Fire and non-fire objects may have similar color patterns .

• Fire has irregular shapes and dynamic textures.

• Variations in illumination and environmental conditions affect detection accuracy.

• Traditional rule-based detection methods cannot handle complex scenes effectively.

To address these challenges, this work applies a deep convolutional neural network model capable of learning robust visual representations for fire detection.

3. System Overview

The proposed fire detection framework consists of several stages including data preprocessing, feature extraction, model training, and prediction.

The overall workflow includes:

1. Dataset loading

2. Data analysis

3. Image preprocessing

4. Dataset splitting

5. Data augmentation

6. CNN model training

7. Model evaluation

8. Fire prediction

4. Input and Output Specifications

Input

RGB Image

Size: 128 × 128 × 3

Output

The system produces:

• Fire / Neutral classification

• Prediction confidence score

5. Methodology

Step 1: Dataset Loading

The fire detection dataset is downloaded and loaded into the training environment. The dataset contains images belonging to two classes:

• Fire

• Neutral

The images represent real-world scenarios containing fire scenes and non-fire environments.

Step 2: Data Analysis

The dataset is analyzed to understand the distribution of fire and neutral images. Class distribution visualization is performed using bar graphs.

Step 3: Image Preprocessing

To prepare images for model training, several preprocessing operations are performed:

• Image resizing to 128 × 128 pixels

• RGB color conversion

• Pixel normalization (scaling values between 0 and 1)

These preprocessing steps ensure consistent input size and improve model performance.

Step 4: Dataset Splitting

The dataset is shuffled and divided into training and testing datasets.

Typical split:

• Training data: 80%

• Testing data: 20%

Label encoding is applied to convert class labels into numerical values.

Step 5: Data Augmentation

Data augmentation techniques are applied to increase the diversity of training images and improve generalization capability. Augmentation techniques include:

• Rotation

• Horizontal flipping

• Width shifting

• Height shifting

• Zoom transformation

• Shear transformation

Step 6: CNN Model Training

A convolutional neural network is trained to classify images into fire or neutral categories.

The network architecture includes:

• Convolution layers for feature extraction

• Batch normalization layers

• Max pooling layers

• Global average pooling layer

• Fully connected dense layers

• Dropout layer for regularization

The final output layer uses sigmoid activation for binary classification.

Step 7: Model Evaluation After training, the model performance is evaluated using different evaluation metrics to measure classification accuracy and reliability.

Step 8:

Fire Prediction The trained model is used to predict whether a given input image contains fire.

Prediction steps:

1. Load the trained model

2. Preprocess the input image

3. Pass the image through the trained model

4. Obtain prediction result (Fire or Neutral)

6. Expected Outcomes

 The proposed system is expected to achieve accurate fire detection and reduce false alarms in non-fire scenes. The model learns important visual characteristics such as flame patterns, brightness, and texture to distinguish fire from other objects.

7. Conclusion

 This work presents a deep learning-based fire detection framework using a convolutional neural network. By combining image preprocessing, data augmentation, and CNN-based feature learning, the system effectively classifies images into fire and neutral categories.

 The proposed framework can be integrated into real-time monitoring systems and web-based applications for automatic fire detection.

8. Reference

1. Patil, M. S., Chauhan, M. S., & Nikum, M. V. (2025). Fire Detection Using Machine Learning. International Journal of Advanced Research in Computer and Communication Engineering.

2. Ifeoma, N., Ekene, A., Obinna, O., & Kingsely, I. (2024). Development of a CNN-Based Smoke/Fire Detection System for High-Risk Environments. European Journal of Science, Innovation and Technology.

3. Zhang, Q., et al. (2020). Image Fire Detection Algorithms Based on Convolutional Neural Networks. Case Studies in Thermal Engineering.