Research on flame detection method based on improved SSD algorithm

Article type: Research Article

Authors: Zhan, Huawei^{a; b; c} | Pei, Xinyu^{a; b; c; *} | Zhang, Tianhao^{a; b; c} | Zhang, Linqing^{a; b; c}

Affiliations: [a] College of Electronics and Electrical Engineering, Henan Normal University, Xinxiang, China | [b] Henan Key Laboratory of Optoelectronic Sensing Integrated Application, Xinxiang, China | [c] Henan Engineering Laboratory of Additive Intelligent Manufacturing, Xinxiang, China

Correspondence: [*] Corresponding author. Xinyu Pei. E-mail: [email protected].

Abstract: A flame detection algorithm based on the improved SSD (Single Shot Multibox Detector) is proposed in response to the issues with the limited detection distance, delayed reaction, and high false alarm rate of previous flame detection systems. First, the ResNet-50-SPD model was added to the original backbone network to improve the detection of low resolution and tiny objects. After that, incorporate feature fusion between layers to improve the bond between contexts. Before the feature entered the prediction, the impact of channel number reduction was eliminated using the adaptive module AAM. According to experimental findings, the modified SSD algorithm’s mAP value on on the random division dataset and K-fold verification dataset reaches 87.89% and 89.63%, respectively, which is 3.97% and 5.17% higher than the original SSD, while the FPS remains at 64.9 f/s. It is helpful to improve the time of the fire alarm, find the ignition point in time, and better meet the actual engineering needs of fire monitoring.

Keywords: Flame detection, SSD, ResNet-50-SPD, feature fusion, AAM

DOI: 10.3233/JIFS-232645

Journal: Journal of Intelligent & Fuzzy Systems, vol. 45, no. 4, pp. 6501-6512, 2023