关键词: 林火识别, 烟雾识别, 目标检测, 卷积神经网络

Abstract:

【Objective】Since both flame and smoke are dynamic targets, it makes forest fire small targets and smoke detection in aerial view challenging. The construction of forest fire small target and smoke detection model based on convolutional neural network has positive significance in balancing the number of parameters and detection accuracy of forest fire small target and smoke target.【Method】For the real-time requirements, the high real-time based FasterNet is improved and the backbone network layer of FRSCnet was constructed. The neck network layer of FRSCnet was constructed using group shuffle convolution (GSConv). In the head network, the convolutional block attention module (CBAM) was introduced. Finally, receptive field block (RFB) feature extraction module was added to the first layer of the neck network. 【Result】FasterNet employs a unique partial convolution (PConv) technique that effectively reduce redundant computations and memory accesses, thus ensuring the high-speed detection performance of the model when processing aerial images.GSConv is able to balance the number of parameters of the model with the performance of the feature fusion, thus reducing the complexity of the model while ensuring the detection accuracy. GSConv is able to balance the number of parameters and feature fusion performance of the model, thus reducing the complexity of the model while ensuring detection accuracy. With the CBAM attention mechanism, the model is able to better focus on forest fire small targets and smoke targets to improve the detection performance while increasing the number of parameters as little as possible. To increase the receptive field of the model to better extract forest fire small targets and smoke targets in aerial view.RFB (receptive field block) feature extraction module, which can increase the receptive field of the model to better extract forest fire small targets and smoke targets in aerial view. The experimental results show that the proposed FRSCnet model achieves a performance of 89.2% in terms of average accuracy (mAP@0.5), which is 2.9% and 5.3% better than YOLOv7-tiny and YOLOv5-s, respectively, but the number of parameter counts is lower than YOLOv7-tiny and YOLOv5-s. 【Conclusion】The proposed FRSCnet forest fire small target and smoke target detection model strikes a good balance between the number of parameters and detection accuracy.

Key words: forest fire recognition, smoke recognition, target detection, convolutional neural network