A recent study published in Fire suggests an efficient pre-processing image detection approach for flames and smoke produced during a fire. The color conversion and corner detection approach detect the flame zone. The optical flow method identifies the smoke region.
Study: A Study on a Complex Flame and Smoke Detection Method Using Computer Vision Detection and Convolutional Neural Network. Image Credit: LuYago/Shutterstock.com
Pre-processing images remove background regions and enable the selection of regions relevant to fire using a deep-learning-based convolutional neural network (CNN). When a fire is detected using this method, the detection accuracy increases by 5.5 to 6%.
Image-Based Fire Detection Method
Engineering strategies used in a fire include compartmentalization, dilution, ventilation, and pressurization to stop the spread of smoke and flames. These techniques are crucial for putting out fires when they start. These strategies do not detect a fire before it starts to develop.
An image-based fire detection method can address this issue. It specifically attempts to detect any possible flame and smoke that may appear in the early stages of the fire to respond appropriately. It is crucial to identify smoke in a fire and flames, particularly because smoke harms people's health more frequently than flames do. Due to its high temperatures, oxygen shortage, and carbon monoxide content, this type of fire smoke harms a person's health. Reduced visibility and the subsequent psychological fear harm evacuation behavior in addition to these direct reasons.
Limitations of Current Image-Based Fire Detection Methods
An artificial intelligence-based fire detection method "you only look once" (YOLO) model based on Tensorflow is used in an existing deep-learning computer vision-based flame detection technique to identify flames without separately filtering the input images. This method can only detect flames and not smoke.
Using a convolutional neural network (CNN) based on a multi-scale prediction framework fire detection, an accuracy of 97.9% can be achieved. Using a UAV-based object detection system, a 92.7% accuracy in detecting fire can be achieved,
Image Pre-Processing for Fire Detection
To filter out non-fire-related elements in advance, Ryu et al. suggested an efficient image pre-processing method for flame and smoke from the input image. Image pre-processing increases accuracy because the unwanted background region can be deleted beforehand, lowering the number of false negatives.
A flame can be produced when a flammable gas from pyrolysis of a solid combustible substance, such as wood, is combined with air and burned. Evaporative combustion can also produce a flame when the combustible liquid evaporates and burns. Pre-processing was used to identify the flame using features linked to its appearance when such burning occurs. Corner detection and hue saturation value (HSV) color conversion were employed during image pre-processing to detect flame.
The first step in the HSV color conversion process is detecting a color zone where a flame can be present. The flame is a crisp texture, resulting in many corners, among the objects that remain after HSV color conversion.
A convolutional neural network (CNN) was employed to identify fire with increased precision and dependability for the candidate smoke and pre-processed flame regions. Images of flames and smoke were gathered and accommodated as a training dataset, and the CNN model Inception-V3 was utilized for inference.
Research Findings
Ryu et al. proposed a suitable pre-processing technique to identify flames and smoke that appear during the early stage of a fire. Color-based and optical flow approaches were applied to achieve this, and a deep-learning-based CNN was employed to conclude the candidate region. It was feasible to increase fire detection accuracy while lowering false alarms brought on by obtrusive background regions.
Compared to image detection models without independent pre-processing, the proposed flame detection method's tests revealed an accuracy gain of 5.5%. Dark channel feature and optical flow were used for the smoke detection approach suggested in this work, and accuracy increased by 6% compared to previous object detection models.
Future of Image Pre-Processing for Fire Detection
Future research on image pre-processing can build or enhance a CNN to accurately detect irregularly shaped items, including fires and smoke. Creating an intelligent fire detector that can be used with low-specification systems and easily perform real-time detection by enhancing pre-processing techniques can also be researched. It is possible to create a technique that identifies fire accurately even from minute details in images and create a fire detection model with higher reliability.
Reference
Ryu, J., & Kwak, D. (2022). A Study on a Complex Flame and Smoke Detection Method Using Computer Vision Detection and Convolutional Neural Network. Fire, 5(4), 108. https://www.mdpi.com/2571-6255/5/4/108
Disclaimer: The views expressed here are those of the author expressed in their private capacity and do not necessarily represent the views of AZoM.com Limited T/A AZoNetwork the owner and operator of this website. This disclaimer forms part of the Terms and conditions of use of this website.