Abstract:As deep learning technologies, specifically convolutional neural networks (CNNs), have advanced, object detection models have become more powerful. This has opened up new possibilities for using cameras to detect fire and smoke for fire warning purposes. However, the high computational demands of deep CNNs make them unsuitable for low computing power devices, which is a challenge for fire detection. A fire detection algorithm based on adaptive region decoupling knowledge distillation has been proposed to address this. Knowledge distillation is a model compression technique that transfers knowledge from a larger, more accurate teacher model to a smaller, less accurate student model. This allows the smaller model to achieve relatively high accuracy using fewer parameters and computational resources. In the proposed approach, the most informative region of the feature map is selected through adaptive region selection, and separate the foreground and background of this region, distilling them with different weights. This approach has been applied to state-of-the-art detectors, and experiments have been conducted on the D-fire dataset. The experiments demonstrate that this approach has improved the AP score from 37.9 to 38.9 for Gfocal with ResNet-50 while training each model for only 12 epochs, especially without incurring any additional cost. Additionally, the proposed approach has achieved similar AP scores as the teacher model but with lower model complexity, fewer parameters, and a faster detection speed of less than 50 milliseconds per image. Specifically, this approach addresses the challenge of deploying such models for fire detection and proposes a promising solution.

Fire Detection Based on Infrared Filter and Deep Learing

Deep Learning Based Fire Detection System For Surveillance Videos

Deepfirenet: A Real-Time Video Fire Detection Method Based On Multi-Feature Fusion

Fire Detection in Infrared Video Surveillance Based on Convolutional Neural Network and SVM

Multi-Feature Based Fire Detection in Video

A Fire-Alarming Method Based on Video Processing

Fire Detection with Day-night Infrared Camera

Deep Learning-Based Fire Detection for Enhanced Safety Systems

Real-time multi-feature based fire flame detection in video.

Video Fire Detection Methods Based on Deep Learning: Datasets, Methods, and Future Directions

Real Time Fire Detection using Deep Convolutional Neural Networks and Long-Short Term Memory in Video Surveillance

Multi-Scale Video Flame Detection for Early Fire Warning Based on Deep Learning

Early Stage Fire Detection System Based on Shallow Guide Deep Network

Image-Based Fire Detection Using Dynamic Threshold Grayscale Segmentation and Residual Network Transfer Learning

Fire Detection Method Based on Improved Deep Convolution Neural Network.

A Fire Detection Algorithm Based on Adaptive Region Decoupling Distillation

An efficient deep neural network with color-weighted loss for fire detection

A Robust Fire Detection Model via Convolution Neural Networks for Intelligent Robot Vision Sensing

Image-Based Real-Time Fire Detection using Deep Learning with Data Augmentation for Vision-Based Surveillance Applications

An efficient fire detection algorithm based on multi‐scale convolutional neural network

Initial Fire Detection Method for Intelligent CCTV Using Deep Learning Technology