Comparative Analysis of CNN and Transformer Architectures for Real-Time Fire and Smoke Detection
Open Access
Abstract
Real-time automatic detection of fire and smoke is a critical component of modern safety systems in surveillance, industrial, and environmental-protection applications, where conventional sensor-based systems show fundamental limitations. In this paper, a comparative analysis of four deep-learning architectures—DETR, Faster R-CNN (ResNet-50-FPN), Faster R-CNN (MobileNetV2), and RetinaNet—was conducted on a heterogeneous corpus of 67,765 annotated images originating from four different datasets. All models shared the same data-preparation, augmentation, and evaluation pipeline, while each was trained with the default configuration of its reference implementation; the comparison therefore reflects each architecture as it is typically deployed rather than a comparison under a single unified training budget. Faster R-CNN with the ResNet-50-FPN backbone achieved the highest accuracy (mAP@0.50 = 78.7% on the Indoor set; 73.8% on the SmokeAndFire set) and the highest mean mAP@0.50 across all datasets (48.7%), obtained with an inference speed of 76.9 FPS and a latency of 13.5 ms on NVIDIA RTX 4090 hardware, which makes it a promising candidate for real-time fire-detection systems on comparable hardware. The main contribution of this work is a unified evaluation of four representative object-detection architectures across four heterogeneous fire-and-smoke datasets. The study further quantifies the performance asymmetry between fire and smoke detection through a normalized morphological asymmetry index, reflecting the consistently lower detection accuracy achieved for smoke owing to its diffuse and semi-transparent appearance, and provides practical guidelines for selecting an architecture according to real-time deployment requirements. Because each configuration was trained once with a fixed random seed and all speed measurements were obtained on a single desktop GPU, the reported differences are interpreted descriptively; their statistical validation, cross-dataset evaluation, and measurement on embedded hardware are identified as future work.Keywords
Funding Information
- Academy of Technical and Art Applied Studies (451-03-34)
- Ministry of Science, Technological Development, and Innovation (2026-03/200132 (University of Kragujevac, Faculty of Technical Sciences Cacak))
This publication has 31 references indexed in Scilit:
- FLAME: fire detection in videos combining a deep neural network with a model-based motion analysisNeural Computing and Applications, 2025
- Scale-aware token-matching for transformer-based object detectorPattern Recognition Letters, 2024
- Advances in the study of global forest wildfiresJournal of Soils and Sediments, 2023
- An improved feature pyramid network for object detectionNeurocomputing, 2022
- Wildfires and global changeFrontiers in Ecology and the Environment, 2021
- GridMix: Strong regularization through local context mappingPattern Recognition, 2021
- End-to-End Object Detection with Transformers2020
- Deep learningNature, 2015
- The Pascal Visual Object Classes (VOC) ChallengeInternational Journal of Computer Vision, 2009
- An experimental comparison of performance measures for classificationPattern Recognition Letters, 2009