Research On Straw Burning Smoke Detection Algorithm Based On Hard Sample Mining

Straw burning will pollute air,destroy soil structure,cause traffic accidents,and even fire.Smoke is one of the most significant characteristics in the early stage of straw burning.Through detecting smoke accurately in the field scene,finding the phenomenon of straw burning timely and warning can reduce the harm of straw burning.Therefore,it is of great practical significance to research a fast and effective smoke detection method.With the development of computer technology and artificial intelligence,smoke detection technology based on computer vision has been widely concerned.Different colors and changeable shape of smoke have brought huge challenge to smoke detection.At present,smoke detection technology mainly focuses on traditional image processing technology and deep learning.These existing methods are greatly affected by interference of external environment,leading to high false alarm rate.Therefore,it is of great research significance to propose an intelligent detection method with high detection accuracy and strong discrimination ability.In this paper,smoke detection algorithms based on hard sample mining are studied.The main work is as follows:(1)Establish smoke dataset.At present,there is a lack of publicly available standard data sets in the field of smoke research.At the same time,in existing smoke detection research,most of data sets used are taken from controllable experimental scenes,which are difficult to apply to complex and changeable scenes.Therefore,this paper collects a large number of smoke samples from cameras deployed on the signal towers,and standardizes this dataset with Pascal VOC2007 benchmark.The dataset contains a variety of complex environments,which can meet needs in real-life scenarios.(2)This paper proposes smoke detection algorithm based on hard negative sample mining.To tackle this problem of high false alarm rate in smoke detection,this algorithm integrates hard negative samples mining method into Faster R-CNN network,by using the rules for generating and selecting candidate regions of Faster R-CNN network to select harder negative samples.These samples information enhance the descriptive power of its characteristics,so as to reduce false alarm rate.(3)This paper proposes smoke detection algorithm based on dynamic hard sample mining.From another point of view to improve generalization ability and discrimination ability in network,by calculating samples loss function and dynamically adjusting the number of positive and negative samples to generate model with stronger discrimination ability.In the early stage of network training,more hard positive samples are selected to fit this network to improve generalization ability;in the late stage of network training,more hard negative samples are selected to improve discrimination ability to smoke similar targets.Compared with relevant algorithms,experimental results verify this proposed method effectiveness.