An Optimized Deep-Learning-Based Network with an Attention Module for Efficient Fire Detection

Abstract

Globally, fire incidents cause significant social, economic, and environmental destruction, making early detection and rapid response essential for minimizing such devastation. While various traditional machine learning and deep learning techniques have been proposed, their detection performances remain poor, particularly due to low-resolution data and ineffective feature selection methods. Therefore, this study develops a novel framework for accurate fire detection, especially in challenging environments, focusing on two distinct phases: preprocessing and model initializing. In the preprocessing phase, super-resolution is applied to input data using LapSRN to effectively enhance the data quality, aiming to achieve optimal performance. In the subsequent phase, the proposed network utilizes an attention-based deep neural network (DNN) named Xception for detailed feature selection while reducing the computational cost, followed by adaptive spatial attention (ASA) to further enhance the model’s focus on a relevant spatial feature in the training data. Additionally, we contribute a medium-scale custom fire dataset, comprising high-resolution, imbalanced, and visually similar fire/non-fire images. Moreover, this study conducts an extensive experiment by exploring various pretrained DNN networks with attention modules and compares the proposed network with several state-of-the-art techniques using both a custom dataset and a standard benchmark. The experimental results demonstrate that our network achieved optimal performance in terms of precision, recall, F1-score, and accuracy among different competitive techniques, proving its suitability for real-time deployment compared to edge devices.