Adaptive linear feature-reuse network for rapid forest fire smoke detection model

Current mainstream smoke detection methods have some problems, such as missing report, large deviation of detection box and slow detection speed. Therefore, we propose an Adaptive Linear Feature-Reuse Network for Rapid Forest Fire Smoke Detection (ALFRNet). Firstly, we designed Double Linear Feature-Reuse Module(DLFR Module) to reduce information loss in the process of the acquisition of smoke images；and Hybrid Attention-Guided Module (HAG Module) was proposed to reduce the interference caused by blurred image and to emphasize the expression of smoke characteristics. At the convolutional layer of the proposed network, a novel Adaptive Depthwise Convolution Module (ADC Module) which can effectively solve the problem of difficulty in recognition caused by too small smoke targets in images was used. Besides, we adopt Cluster NMS (CNMS) in purpose of avoiding large deviation of the detection box. It can adapt to smoke target with fuzzy edge and improve the performance of detection. Finally, we build a smoke detection system based on the Internet of things. The experimental results show that our method achieves 87.26% mAP50 at 43 FPS on NVIDIA TITAN Xp. Compared with other mainstream methods, it has the better performance of quicker speed, higher accuracy and more accurate position of detection box.