Fire and Smoke Detection with Intensity Classification

Authors

• Md Tamzid Omor Bhuiyan Department of ICT, Bangladesh University of Professionals, Mirpur Cantonment, Dhaka
• Md Asifuzzaman Department of ICT, Bangladesh University of Professionals, Mirpur Cantonment, Dhaka
• Faizur Rahman Department of ICT, Bangladesh University of Professionals, Mirpur Cantonment, Dhaka

Abstract

Fire detection and smoke classification are among the most critical applications of machine learning and computer vision, leveraging advanced neural networks to ensure safety and reliability. However, there are still many challenges that need to be overcome in this field today including the application of these systems to a wider range than a limited dataset or improving generalizability. Moreover, most of the existing architectures have been set with restrictions and designed to work with specified fire data which limit their versatility. In this study we address these issues by presenting a two-stage framework for fire detection and classification based on modern Convolutional Neural Networks (CNN) architectures. For the purposes of this research, contact has been made with Fire-D data set, which enables the use of DenseNet201, MobileNetV2, InceptionV3, Xception, InceptionResNetV2 and NASNetMobile architectures. In addition, the cluttered dataset will allow us to resolve image classification and object detection tasks simultaneously. Focus of our model is on boundary box optimization and multi task learning to ensure a robust classification of 96.43% accuracy, with a precision of 96.45%. For object detection, the performance was below par with a Mean IoU of 60.06% but there is still room for improvement. Multiple configurations tested showed that the system is self efficient and scalable. This work demonstrates the feasibility of employing modern CNN s for a fire incident analysis as a step towards the usage in real environments with limited resources and minimum computational capacity.

Introduction

The rise in the intensity and the incidence of urban and nature fires is increasingly becoming a danger to people, the environment and even property. It is thus imperative to have effective fire detection systems. Traditional sensor-based systems have existed for a while but they are not able to keep up with the smoke and fire sensor requirements of low latency, flexibility and accuracy. On the other hand, computer vision, and more specifically CNNs are vastly more effective at fire and smoke detection.

This study focuses on the use of advanced smoke and flame identification and classification through the use of CNN models such as DenseNet201, MobileNetV2, and Inception-ResNetV2. The models were selected due to their proven ability for image-based tasks. For robust evaluation of these models, the varying fire and smoke scenarios provided by the D-Fire dataset was considered.

This new integrated method to classification and detection is a huge improvement compared to traditional methods which lack in both categories. This dual strategy enhances practical applications by improving response times and accuracy, paving the way for more effective fire control systems.

The methodology relies on cutting edge CNN architectures and a custom dataset, taking into account such problems as non-uniform fire strengths, their context and visual barriers. These efforts are focused on developing modular and scalable solutions for a wide range of environments for example, houses, industries, or the outdoors [1].

This work is a step further in the scope of AI and machine learning in disaster management by undertaking a comparative study to determine the most optimal detection techniques. As fire and smoke are challenged with many complexities this research aims to contribute through the use of modern methods enhanced technology towards achieving improved outcomes of protecting lives and assets with effective and efficient real-time solutions.

Literature Review

The creation of fire detection systems has become a fo-cal point for many researchers, especially considering the increasing demand for fire hazards that could be detected quickly. Many works have incorporated CNNs into fire and smoke detection systems as it has been demonstrated that such systems can ensure high accuracy due to their capacity to retrieve complex features from images.

Majid et al. [1] in 2022 used an attention-based CNN and recorded 95.40% recall and 97.61% accuracy in fire detection. In the same vein, Abdusalomov et al. [2], 2021, presented improvements in fire detection enhancement using YOLOv3, further increasing the robustness and decreasing the processing time. In a separated test dataset, Nishanth and Varaprasad

1. in 2023 were able to achieve optimal performance on it to support the efficiency of MobileNetV2. In the same year. Sathishkumar et al. [4] took a different approach, where they transferred the learning into their model and achieved 96.89% accuracy in detecting forest fires.

According to Venancioˆ et al. [6] in 2022, they were able to create a low-power CNN based system by tweaking the convolutional filters but still getting the intended accuracy. In the meantime, Daoud et al. [7] developed a FireClassNet and reached a remarkable accuracy of 99.73% on classification of PJF fire images. Li and Zhao [8] along the same lines pointed to the capability of CNNs for smoke and fire detec-tion in surveillance videos, achieving significant performance improvements.

Back in 2018, handcrafted features have been out hand covered by CNN models that were created by Muhammad et al. [9] when they worked on cv fire detection in surveillance videos. That very same year, they specialized in the design of a CNN model which minimalist in nature, having its roots in Squeeze Net, the model is highly efficient in detecting and localizing fires, all this while having considerably low memory and computational requirements [10].

Tsalera et al. [11] examined lightweight CNNs such as SqueezeNet and ShuffleNet and achieved 96% accuracy on the average for wildfire recognition. Faster R-CNN with multidimensional texture analysis was used by Barmpoutis et al. [12] and the true positive rate was high but the false alarms were considerably decreased. As elucidated by Sun et al. [13] (2023), LCNet is effective in smoke detection of forest fires in real time. It is also increased the feature extraction capabilities. Similarly, Jeon et al. [14] (2021) employed a multi-scale prediction of fires in order to ameliorate the false alarms which were reliant on CNN based techniques.

A vast majority of the surveys and research papers have been churned out for either one of the two, either object detection or classification and not both. This gap is filled in this study through advanced cutting edge CNN architectures with a dual task for apRing all the above-mentioned features in a fully scalable manner. The efficient combination of the models and the data set are specifically designed to remove the bottlenecks that currently exist and are required for real time processing.

Dataset Description Plus Images

There are several datasets available for smoke and fire classification such as the” Fire Dataset” [5] published during the NASA Space Apps Challenge. However, the D-Fire dataset [6], which is captured through live video streams, is more com-patible with contemporary models. The dataset comprises four class labels as illustrated in Table I, which are, fire, smoke, both and none, along with the bounding boxes delineated in Table II, the total number of class images exceeds 21, 000. This dataset aids in the improvement of a classification and an object detection task.

Table I

Images In Each Category Before Data Processing

Category	Images
Only Fire	1,164
Only Smoke	5,867
Fire And Smoke	4,658
None	9,838

Table I I

Number Of Bounding Boxes in Each Class

Class	Bounding Boxes
Fire	14,692
Smoke	11,865

Example image Original (a)

Separate Fire and Smoke boxes (b)

One box for both Fire and Smoke (c)

Figure. 1. Sample Images Detecting Fire and Smoke

Study Methodology

The challenges posed by the task of detecting fire and smoke were considered in the design and methodology by employing the latest cutting-edge architecture in CNN, which has a multi-task classifier and object detection architecture. The entire flow has been encapsulated in Fig. 1.

The D-Fire dataset was preprocessed for dual-task classification and object detection by standardizing image sizes to ensure uniformity, applying data augmentation techniques (rotation, flipping, color variation) to mitigate overfitting, and splitting the data into training (70%), validation (20%), and test (10%) sets to minimize bias. Six CNN archi-tectures—DenseNet201, MobileNetV2, InceptionV3, Xcep-tion, InceptionResNetV2, and NASNetMobile—were selected for their efficacy in image-based tasks. DenseNet201 lever-ages densely connected layers for robust feature reuse; Mo-bileNetV2 employs depth wise separable convolutions for

Figure. 2. Research Workflow

lightweight deployment on low-resource devices; InceptionV3 optimizes computational efficiency via factorized convolu-tions; Xception enhances feature extraction by replacing Incep-tion modules with depthwise separable convolutions; Incep-tionResNetV2 combines residual connections with Inception modules to mitigate gradient vanishing; and NASNetMo-bile utilizes Neural Architecture Search (NAS) for mobile-optimized precision. A unified framework enabled simulta-neous classification (into Fire, Fire with Smoke, Smoke, or None) and object detection (via bounding box regression), with cross-entropy loss for classification and mean squared error for localization. Adaptive learning rates ensured stable convergence. Performance was evaluated using accuracy, pre-cision, recall, F1 score, and mean Intersection over Union (IoU), with IoU specifically assessing bounding box accuracy. Challenges like class imbalance and detection inconsistencies were addressed through augmented data, hyperparameter tun-ing, and class-weighted loss functions, refining the models to maximize classification accuracy while maintaining robust detection capabilities.

Tested Models and Results

1. Tables for Results

Table III displays performance evaluation metrics of the six CNN models on four different metrics, which are accuracy, precision, recall and F1 Score. It can be observed that the InceptionResNetV2 model out performs all other models with accuracy and F1 score of 96.43%, when classifying the data into fire, smoke or both. While the DenseNet201 and NASNet-Mobile models also performed considerably well in terms of accuracy as they achieved over 96%, proving their efficiency

in performing classification-based tasks. On the other hand, although MobileNetV2 produced the worst results, it was still able to sustain decent levels of precision and recall.

Table IV Measures the object detection performances of the 6 models using Mean Intersection over union, Precision, Recall and F1 Score. NASNetMobile however surpassed all other models with a Mean IoU of 60.06%, precision of 61.08%, recall of 61.16% and F1 score of 60.75%. This shows that this particular model has proven to be effective in detecting and localizing smoke and fire regions. Despite this though, it is important to note that all the models still face difficulties performing the task of object detection as shown by the lower Mean IoU scores.

Table I II

Classification Results

	Model			Performance Metrics
			Accuracy	Precision		Recall		F1 Score
	DenseNet201		96.27%	96.31%	96.26%		96.28%
	MobileNetV2		95.37%	95.50%	95.36%		95.40%
	InceptionV3		95.93%	95.99%	95.93%		95.95%
	Xception		95.40%	95.55%	95.40%		95.44%
	InceptionResNetV2		96.43%	96.45%	96.43%		96.43%
	NASNetMobile		96.03%	96.11%	96.03%		96.05%
	Table IV Object Detection Results
	Model		Mean IoU	Precision		Recall		F1 Score
	DenseNet201		50.12%	51.22%		51.05%		50.77%
	MobileNetV2		49.77%	50.97%		50.72%		50.43%
	InceptionV3		56.98%	58.06%		58.08%		57.66%
	Xception		51.14%	52.52%		51.79%		51.80%
	InceptionResNetV2		52.58%	53.84%		53.52%		53.32%
	NASNetMobile		60.06%	61.08%		61.16%		60.75%

B. Result Analysis of DenseNet201

Figure. 3. DenseNet-201 Classification

Figure. 4. DenseNet-201 Loss Graph

C. Result Analysis of MobileNetV2

Figure. 5. MobileNet-V2 Classification

Figure. 6. MobileNet-V2 Loss Graph

Over in MobileNetV2, 130” Fire” images, 580” Fire and Smoke”, 218” Smoke” images, and 1933” None” images were incorporated into the model. There was noticeable improve-ment within the training loss graph, which is treated as an upper bound, indicatingthat the model is positively adjust-ing. This was followed by Mixed Outcome, which displayed greater fluctuations than validation loss. Contrary to mixed, classification loss recorded stable figures over the training period, while mixed started at a higher figure and tried moving down to the unchanged point.

Figure. 7. Inception-V3 Classification

Figure. 8. Inception-V3 Loss Graph

InceptionV3 managed to attain more than decent scores for” Fire” 126 images,” Fire and Smoke” 582 images,” Smoke” 216 images and” None” 1954 images. The classification loss showed sharp descent and altered after epoch 5, which caused training loss graph to suffer. The mixed loss improved slightly, showing minor shock.

E. Result Analysis of Xception

Xception identified images: 129 labeled as “Fire,” 578 as “Fire and Smoke,” 220 as “Smoke,” and 1935 as “None.” The bounding box loss graph during the training phase de-picts initial decrease in the oscillatory nature of the graph. Validation loss experienced the same phenomenon but with lowered ranges. Training classification loss steadily decreased and plateaued, while validation loss remained unchanged after some minor shifts.

F. Result Analysis of InceptionResNetV2

Figure. 9. Xception Classification

Figure. 10. Xception Loss Graph

Figure. 11. InceptionResNet-V2 Classification

Figure. 12. InceptionResNet-V2 Loss Graph

G. Result Analysis of NASNetMobile

Figure. 14. NASNetMobile Loss Graph

The NAS Net Mobile model classified 129 images as Fire, 582 as Fire and Smoke, 221 as Smoke, and 1949 as None. The bounding box loss graph indicates a significant decline at the start of training in conjunction with the validation loss, which stabilizes during the slight movement changes. Validation loss is unable to track the changes and remains at a poorer level, indicating that the model had more problems with the validation set. In the classifying task, training loss drops too fast and validation loss shows many more oscillations, indicating that the model is not consistently performing across the validation batches.

The results depict that all models were able to classify with great accuracy and exhibited relatively the same loss figures during the training and validation stages.

Conclusion

Our study presents a unified framework for fire/smoke detection and intensity classification, evaluating six CNNmodels (DenseNet201, MobileNetV2, InceptionV3, Xception, InceptionResNetV2, NASNetMobile) on the Fire-D dataset. Achieving 95.37%+ accuracy, the framework classifies Fire, Fire with Smoke, Smoke, and None, combining detection and intensity classification—a novel approach addressing gaps in single-task systems. Optimized bounding box localization enhances spatial precision, while future work focuses on lightweight architectures for real-time processing in resource-limited settings, hybrid datasets (thermal + visual data), and temporal feature integration to automate safety alerts in low-infrastructure environments. The goal of these modifications is to create intelligent, lazy systems to deploy in situations where real-time fire incident detection is needed and greatly help in automation for safety system

References:

1. S. Majid, F. Alenezi, S. Masood, M. Ahmad, E. S. Gund¨uz,¨ and K. Polat, “Attention-based CNN model for fire detection localization in real-world images,” Expert Systems with Applications, vol. 189, p. 116114, 2022.
2. A. Abdusalomov, N. Baratov, A. Kutlimuratov, and T. K. Whangbo, “An improvement of the fire detection and classification method using YOLOv3 for surveillance systems,” Sensors, vol. 21, no. 19, p. 6519, 2021.
3. P. Nishanth and G. Varaprasad, “Forest fire detection and classification using deep learning concepts,” in International Conference on Data Management, Analytics & Innovation, pp. 37–43, Springer, 2023.
4. V. E. Sathishkumar, J. Cho, M. Subramanian, and O. S. Naren, “Forest fire and smoke detection using deep learning-based learning without forgetting,” Fire Ecology, vol. 19, no. 1, pp. 1–17, 2023.
5. G. Ahmed, A. Ahmed, S. Heba, and S. Ahmed, “Fire dataset,” 2018.
6. A. V. B. Pedro Vin´ıcius Almeida Borges de Venancioˆ and A. Chaves Lisboa, “An automatic fire detection system based on deep convolutional neural networks for low-power, resource-constrained devices,” 2022.
7. Z. Daoud, A. Ben Hamida, and C. Ben Amar, “FireClassNet: A deep convolutional neural network approach for PJF fire images classifica-tion,” Neural Computing and Applications, pp. 1–17, 2023.
8. P. Li and W. Zhao, “Image fire detection algorithms based on convolu-tional neural networks,” Case Studies in Thermal Engineering, vol. 19, p. 100625, 2020.
9. K. Muhammad, J. Ahmad, I. Mehmood, S. Rho, and S. W. Baik, “Convolutional neural networks-based fire detection in surveillance videos,” IEEE Access, vol. 6, pp. 18174–18183, 2018.
10. K. Muhammad, J. Ahmad, Z. Lv, P. Bellavista, P. Yang, and S. W. Baik, “Efficient deep CNN-based fire detection and localization in video surveillance applications,” IEEE Transactions on Systems, Man, and Cybernetics: Systems, vol. 49, no. 7, pp. 1419–1434, 2018.
11. E. Tsalera, A. Papadakis, I. Voyiatzis, and M. Samarakou, “CNN-based, contextualized, real-time fire detection in computational resource-constrained environments,” Energy Reports, vol. 9, pp. 247–257, 2023.
12. P. Barmpoutis, K. Dimitropoulos, K. Kaza, and N. Grammalidis, “Fire detection from images using Faster R-CNN and multidimensional texture analysis,” in ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 8301–8305, IEEE, 2019.
13. B. Sun, Y. Wang, and S. Wu, “An efficient lightweight CNN model for real-time fire smoke detection,” Journal of Real-Time Image Processing, vol. 20, no. 4, p. 74, 2023.
14. M. Jeon, H.-S. Choi, J. Lee, and M. Kang, “Multi-scale prediction for fire detection using convolutional neural network,” Fire Technology, vol. 57, no. 5, pp. 2533–2551, 2021.