Abstract
Most automatic fire alarm systems detect fire through thermal, smoke, or flame sensors. A promising alternative is image-based detection, which is easily embeddable across various devices without specialized hardware. However, the high computational cost of deep learning methods remains a challenge for deployment on portable hardware. In this work, we propose KutralNet, a new deep learning architecture designed for fire recognition with minimal floating-point operations (FLOPs). By utilizing modern techniques like inverted residual blocks, depth-wise convolutions, and octave convolutions, we significantly reduce computational costs. Experiments on the FireNet and FiSmo datasets show that KutralNet maintains high accuracy while using 71% fewer parameters than FireNet and substantially fewer FLOPs than ResNet50, making it ideal for mobile and edge device implementation.
Introduction
The primary challenge in image-based fire detection is the resource-accuracy trade-off. While standard deep learning models like ResNet50 offer high accuracy, their massive memory and processing requirements lead to high energy consumption and slow inference on low-power hardware. Existing lightweight models often struggle to generalize or still demand too many FLOPs for real-time performance. KutralNet (named after the Mapuche word for fire) addresses this by focusing on reducing both model size (parameters) and operational complexity (FLOPs).
Methodology
The authors developed KutralNet through a series of architectural iterations aimed at maximizing efficiency:
- KutralNet (Baseline): A 5-layer CNN optimized for 84x84 RGB images, utilizing LeakyReLU and batch normalization.
- KutralNet Mobile: Replaces standard convolutions with Inverted Residual Blocks. These blocks use depth-wise separable convolutions to simplify signal processing and reduce the parameter count.
- KutralNet Octave: Implements Octave Convolutions, which split the image signal into high and low spatial frequencies. This reduces spatial redundancy by processing the low-frequency component at a lower resolution.
- KutralNet Mobile Octave (Hybrid): This variant combines inverted residuals with octave convolutions to achieve maximum efficiency.
Experimental Setup
The models were evaluated on the FireNet and FiSmo datasets. To ensure robustness, the authors introduced augmented datasets (FiSmoA/BA) containing purely black images labeled as "no-fire." This critical test verified that the models were learning structural features of flames and smoke rather than simply reacting to image brightness.
Results and Discussion
KutralNet demonstrates that frequency splitting is a powerful tool for fire recognition, as the model can focus on global features (low frequency) and edge details (high frequency) separately.
Quantitatively, the improvements are stark. While KutralNet Mobile Octave has slightly more parameters than the baseline, its FLOP count is nearly 70% lower. Specifically, it requires only 24.59M FLOPs, compared to 76.85M for the baseline and 4.13G for ResNet50. Despite this reduction, it achieved a mean test accuracy of 79.49% and an AUROC of 0.90, outperforming FireNet (64.27% accuracy) in cross-dataset testing.
| Model | Parameters | FLOPs | Test Acc (Mean) |
|---|---|---|---|
| ResNet50 | 31.91M | 4.13G | 70.26% |
| FireNet | 646.82K | - | 64.27% |
| KutralNet (Baseline) | 138.91K | 76.85M | 76.01% |
| KutralNet Mobile Octave | 185.25K | 24.59M | 79.49% |
Conclusion
KutralNet provides a highly efficient solution for fire recognition on the edge. While the current model focuses on classification, future work will aim to incorporate bounding box detection for fire localization and adapt the architecture for real-time video stream processing on autonomous mobile vehicles.
Summary updated