| When there is a fire in forest,village and other outdoor environment,it is not easy to be detected as early as possible due to the lack of monitoring strength,so this paper will focus on the detection method of fire smoke in this environment.Through the study of outdoor fire smoke,it is found that it has persistence,spread,diversity,edge effect and drift phenomenon,the last four characteristics make the concentration,shape and color of fire smoke change irregularly and are not easy to be detected.In addition,cloud,as the most common distractor of fire smoke,often appears in the detection scene,this also poses a challenge for fire smoke detection.In view of the above difficulties,this paper proposes a Multi-level Saliency Target Detection Network(MSTDN)to detect outdoor fire smoke.The MSTDN network structure can synthetically identify the smoke area in fire scene by using the semantic information of different levels in feature mapping.Specifically,the network structure of MSTDN is firstly extracted the basic features,and then divided into two branchs to identify the fire smoke region,I.E.Low-level Network Branch structure and high-level network branch structure The semantic information in the low-level network branch contains three different primary feature maps,and the high-level network branch contains two more refined feature maps,these feature maps mainly contain color,texture and intensity information.Finally,the feature information of low-level network branches and high-level network branches is fused to judge the fire smoke region.In order to further improve the detection performance of MSTDN network for outdoor fire smoke,spatial attention mechanism and channel attention mechanism are introduced,and a MSTDN network structure based on attention mechanism is proposed.The experimental results show that the average crossover-to-parity ratio(Miou)of MSTDN can reach 87.71%,which is superior to other four network structures. |
PDF Full Text Request