Research On Flame And Smoke Detection Algorithm In Natural Scene

In recent years,fire and smoke warning has become an important part of modern security.The research and application in the field of prevention and control about fire disaster,has extremely urgent need and considerable significance for the development of modern society.And with the development of computer vision,deep learning technology has shown good performance in various fields and tasks.Meanwhile,flame and smoke detection technology based on deep learning,also gradually becomes hot issue that more and more researchers and scholars pay close attention to.Generally speaking,for flame and smoke detection models based on deep learning,the key to improve the model performance is how to extract more discriminant or informative target features according to the data characteristics of flame and smoke.In addition,it should not be ignored that datasets in the field of flame and smoke detection are very deficient and difficult to collect,which is a great challenge for training supervised deep models.The main research work of this thesis is as follows:(1)Construction of two datasets about flame and smoke detection.Due to the lack of public flame and smoke data and the limited scene,this thesis constructs two flame and smoke detection datasets in natural scene firstly.For indoor monitor perspectives of building,the Staircase dataset collects 18 680 samples,including 14 651 positive samples and 4 029 negative samples.For Internet social media images,the Multiburn dataset contains 14 108 images,including 11 608 positive samples and 2 500 negative samples.The constructed datasets provide a necessary research basis for flame and smoke detection.(2)The flame and smoke detection algorithm based on spatial feature transformation is proposed.Because of the visual difference,target scale and morphological diversity of flame and smoke targets,this thesis proposes a scale-adaptive object detection network model fused with spatial feature transformation,which verifies the effectiveness of the proposed model on two self-built datasets.And it analyzes the impact of different network structures on the performance of the flame and smoke detection model.Compared with the current classic object detection algorithms,the algorithm proposed in this thesis has stronger robustness.(3)The flame and smoke detection algorithm based on dynamic feature fusion is proposed.In order to fuse high-resolution feature and low-resolution feature in deep network,and enrich multi-scale feature information,this thesis proposes an object detection network model based on dynamic feature fusion combining with the feature fusion structure,which proves effectiveness of the proposed model on two self-built datasets.And it analyzes the experiment results of the proposed model.Meanwhile,compared with the current object detection algorithms,it proves the effectiveness of the algorithm proposed in this thesis.