Research And Development Of Smoke Video Detection System Based On Deep Learning

With the development of computer vision theory,smoke video detection technology has gradually replaced traditional smoke detection systems due to its advantages such as large coverage area,short response time,and low cost,and has attracted widespread attention in the industry However,due to the variety of smoke shapes,colors,and movements,the accuracy of smoke detection is not high,causing false alarms and false negatives,which affects the industrialization process of smoke detection systems.In addition,although the smoke detection algorithm based on deep learning improves the accuracy rate,it increases the network parameters,making it difficult to apply it on the terminal side in real time.Therefore,with the help of lightweight deep learning methods,a smoke video detection technology is studied in this thesis,in order to solve the above problems.The main tasks as follows?.Firstly,the basic content,theoretical research and system development of the smoke detection problem are summarized in this thesis.Then,the deep convolutional neural network and the deep object detection network are introduced in detail?.A smoke video detection algorithm based on improved MobileNetV2-SSD is researched Firstly,a new reconstructed pyramid structure is proposed to improve the detection accuracy of small smoke targets.Secondly,a candidate box parameter setting method based on smoke prior characteristics is put forward to quickly and accurately locate the smoke target.Then,a feature enhancement suppression mechanism based on the SE-Net module is introduced to effectively improve the ability of feature expression.Finally,through training and testing in the existing smoke data set,it was verified that the three improvements can improve the accuracy of smoke detection?.A smoke video detection algorithm based on improved 3D residual dense network is researched.Firstly,the location of suspected smoke area based on a priori scoring algorithm is proposed to achieve real-time positioning of smoke targets.Secondly,the lightweight 3D residual dense network is proposed to detect smoke more accurately.Finally,a dynamic detection strategy based on the time-varying features of smoke is proposed to achieve the best compromise between real-time and accuracy.Finally,through experimental comparison,the algorithm in this paper has improved in detection rate and accuracy?.A smoke video detection system based on deep neural network is developed and implemented.Firstly,the requirements and architecture of the system are described.Secondly,this thesis introduces the software environment and hardware resources of the system development,and clarifies the specific implementation process of the system.Finally,the actual operation effect of smoke video detection is showed in this thesis.