Research on Lightweight Fire Flame Detection Model Based on Convolution Neural Network

• Published in: 2023 3rd International Symposium on Computer Technology and Information Science (ISCTIS) pp. 393 - 398
• Main Authors: Gao, Bingwen, Li, Xi, Tang, Defeng, Liu, Yang, Zhao, Songpu
• Format: Conference Proceeding
• Language: English
• Published: IEEE 07.07.2023
• Subjects: adaptive multiscale; color space conversion; component; Convolution; convolutional neural network; Feature extraction; fire flame detection; Fires; Image color analysis; Neural networks; Prediction algorithms; secondary feature extraction; Stability analysis
• DOI: 10.1109/ISCTIS58954.2023.10213009

In order to improve the recognition and detection performance of fire and flame detection systems, a lightweight detection model based on deep convolutional neural networks is proposed by combining machine vision with deep learning technology. Firstly, the model utilizes multiple color space conversion algorithms to process input images to enrich fire and flame information in different scenes; Secondly, a low channel convolution strategy is used to extract coarse flame features from shallow to deep, and a spatial attention mechanism is designed to highlight key features for each proposed dimension of coarse features. Then, a feature fusion method is employed to extract object accurately by connecting the features in a sequential manner. Finally, an adaptive multi-scale fusion structure and an improved non-maximum suppression algorithm are introduced to enhance the recognition effect of fires at different stages. Experiments on multiple public datasets show that the proposed model effectively improves the detection effect of fire flames compared to existing similar detection methods, and has high stability and robustness, which can better prevent fire accidents.