A Novel Vision-based Fire Detection Model via OctConv Network from Optical Images

Fire recognition plays an important role in protecting human life and property and preventing fire accidents. However, the shape and structure of fires vary due to the influence of combustible materials and conditions. In addition, complex backgrounds and different lighting conditions also present significant challenges for accurate fire recognition. Faced with challenging factors, this paper proposes a novel vision-based fire recognition model based on Residual Octave Convolution (OctConv). To obtain strong feature representational capacity, a residual OctConv block is built to build the backbone network, which could well learn features at multiple scales simultaneously. For the OctConv, the high-frequency components capture fine-grained details, while the low-frequency components capture more global context, which enables the network to better recognize objects, leading to improved accuracy. Meanwhile, to retain important feature information while reducing feature redundancy, a shuffle attention mechanism is further introduced into the residual OctConv block to improve the detection performance of small targets. Combined with a self-built dataset, the prediction accuracy of fire recognition has reached 97.87%. Meanwhile, compared to some typical classification networks, proposed model has achieved significant improvements, which could well prove the potential of proposed method in safety monitoring.