FireLite: Lightweight CNN for Fire Detection

• FireLite is a lightweight convolutional neural network engineered for efficient fire detection in resource-constrained embedded systems.
• It leverages a pre-trained MobileNet backbone with a streamlined classifier head, reducing parameters by over 18× compared to larger models.
• The model achieves near-perfect accuracy (99.18%) with minimal false positives, enabling real-time fire detection on transportation infrastructure.

FireLite is a lightweight convolutional neural network (CNN) model engineered for rapid and precise fire detection in resource-constrained embedded environments, notably within transportation systems equipped with IP cameras. FireLite distinguishes itself by its minimal parameter footprint—just 34,978 trainable parameters—combined with state-of-the-art precision and recall, achieved through transfer learning on a MobileNet backbone and a highly efficient downstream classifier. Its technical foundation, experimental validation, and targeted deployment scenarios position FireLite as a leading solution for embedded fire detection applications in safety-critical infrastructure (Hasan et al., 30 Sep 2024).

1. Architectural Design and Mathematical Formulation

FireLite builds upon a MobileNet backbone pretrained on ImageNet, exploiting its capacity for high-level feature extraction with low computational overhead. The classification head is streamlined for minimal footprint:

• Feature Extraction: The MobileNet base is truncated, i.e., classification layers are removed, and only convolutional feature extractors are retained. Most MobileNet layers are frozen during FireLite fine-tuning.
• Global Average Pooling: The feature tensor output by MobileNet, $F(x)$, undergoes spatial pooling, $g(x) = \text{GlobalAveragePooling2D}(F(x))$, dramatically reducing the number of trainable parameters relative to traditional flatten-and-dense architectures.
• Classifier Head: A single Dense layer ($$h(x) = \text{ReLU}(\text{BN}(W_1 \cdot g(x) + b_1))$$) with 32 units provides a compact, regularized transformation. Batch normalization and dropout (rate = 0.5) counteract overfitting.
• Prediction: The output is processed by a final Dense layer ($$y_{\text{pred}} = \text{Softmax}(W_2 \cdot h_{\text{drop}}(x) + b_2)$$) yielding probabilities over two classes: fire and non-fire.

This architecture reduces parameter count—over 18× fewer than FireNet-Tiny—without substantive loss in representational power, facilitating deployment on edge or embedded environments.

2. Transfer Learning Paradigm and Model Adaptation

FireLite’s methodology leverages transfer learning to enhance generalization and speed training, crucial for limited-data domains such as safety monitoring. The operational pipeline is as follows:

• The MobileNet backbone weights are retained, except for the top classifier layers, which are replaced by FireLite’s lightweight head.
• The backbone’s weights are frozen, and only the newly appended head layers are trainable, drastically diminishing the number of updateable parameters.
• Fine-tuning involves fitting only the classifier head to annotated fire/non-fire samples, adapting the pre-trained feature representations to fire-specific cues.

This transfer learning workflow drives sample efficiency and robustness even on datasets substantially smaller than ImageNet, reducing training time and data collection overhead.

3. Performance Metrics and Comparative Analysis

FireLite achieves high performance across several metrics when benchmarked on the FireNet dataset:

Model	Accuracy (%)	Precision/Recall/F1 (%)	Parameters
FireLite	99.18	99.18–99.19	34,978
FireNet-Micro	96.78	—	171,234
FireNet-Tiny	95.75	—	261,922
FireNet-v2	94.95	—	318,460
FireNet	93.91	—	646,818

FireLite reduces false positives (2 cases) and eliminates false negatives, evidenced by a validation loss of 8.74. In all summary measures (accuracy, precision, recall, F1), FireLite matches or outperforms larger fire detection models while maintaining competitive generalization.

4. Computational Efficiency and Deployment Strategy

FireLite’s design enables efficient deployment in environments characterized by severe resource limitations:

• Low Parameter Count: At 34,978 trainable parameters, FireLite imposes negligible memory demands, reducing storage and facilitating rapid model loading.
• Throughput: Model inference benefits from the minimal dense head, with the majority of computation confined to the pretrained MobileNet backbone, which can exploit optimized library implementations for ARM, x86, and embedded GPU architectures.
• Regularization: Batch normalization and dropout allow for aggressive reduction in model complexity without sacrificing stability or accuracy.
• Real-Time Use: These architectural decisions permit practical real-time fire detection on IP camera hardware, embedded vehicle controllers, and other microcontroller-class devices typical of transportation fleet installations.

A plausible implication is that FireLite’s minimal resource footprint enables large-scale, distributed deployment across multiple points of infrastructure without commensurate hardware upgrades.

5. Application Domains and System Integration

FireLite directly addresses fire safety needs in the transportation industry, particularly under conditions heightened by political unrest or where rapid threat response is essential:

• Transport Vehicles: Continuous monitoring in trains, buses, and cargo ships via IP camera streams, with onboard fire detection.
• Automated Safety Protocols: Integration with alert and suppression systems, enabling instant notification and actuator triggering.
• Scalability: The low hardware requirements support broad installation across fleet vehicles and remote infrastructure, even where bandwidth and compute are restricted.
• Cost-Efficiency: By streamlining model footprint, FireLite allows deployment without recourse to specialized server-side inference or high-end local hardware, supporting large deployments with minimal cost impact.

This facilitates a proactive approach to fire hazard mitigation, consistent with the system-level needs of transportation safety networks.

6. Future Directions and Research Opportunities

Authors identify several axes of further development for FireLite:

• Dataset Expansion: Augmenting the training corpus with additional images, including diverse environmental settings and scenarios, to further improve robustness and drive down the rare but nonzero false positive rate.
• Regularization Techniques: Exploring advanced methods (e.g., adaptive regularization, novel normalization schemes) to counteract domain shift and further bolster reliability in varied operational contexts.
• Latency Optimization: Investigating additional architectural compressions, possibly including quantization and pruning, to further reduce inference latency—especially for time-critical response systems.
• Multimodal Safety Networks: Fusion with ancillary sensors and surveillance modalities (thermal, chemical, ultrasonic) to assemble a comprehensive safety grid for fire hazard detection and remediation.

This suggests that FireLite’s evolution may benefit from the quantization, throughput-optimized kernels, and scaling/normalization strategies advanced by FireQ (2505.20839), laying groundwork for deeper integration between low-level model compression and high-level computer vision architectures in safety-critical embedded networks.

7. Formal Context and Comparative Positioning

FireLite represents an overview of contemporary advances in lightweight neural architectures, transfer learning pipelines, and regularized fine-tuning for safety-critical deployment. Unlike KutralNet (Ayala et al., 2020), which achieves ~84% accuracy with ~140K parameters, FireLite takes the additional step of leveraging deep transfer learning for an order-of-magnitude parameter reduction and near-perfect test performance.

Although Light-YOLOv8-Flame (Lan et al., 11 Apr 2025) addresses object-level flame detection with a heavy emphasis on parameter and computation reduction within the YOLOv8 detection pipeline (e.g., FasterNet Block, Partial Convolutions), FireLite’s classifier-centric design enables its unique suitability for embedded platforms and binary event classification.

A plausible implication is that FireLite’s lightweight, transfer-learned architecture constitutes a reference point for subsequent approaches seeking maximal efficiency—potentially incorporating further advances from quantization and kernel-level throughput optimization frameworks as exemplified in FireQ (2505.20839).

In sum, FireLite’s strategically compact yet high-performing transfer learning paradigm authentically advances fire detection methodologies for constrained operational environments, with significant implications for the deployment of automated safety systems across the global transportation sector.