Research And Implementation Of Flame Detection Algorithm Based On Improved SSD

Fire poses an extremely serious hazard to human life and property safety,so the study of fire detection technology is of great importance.Flame as the main visual feature of fire,the real-time detection of flame target has become one of the most effective methods for disaster monitoring.At present,there are two main mainstream flame detection methods,sensor-based flame detection methods and flame detection methods based on traditional digital image processing technology,but they both have low detection accuracy,limited detection range,long delay and other problems.In this thesis,we propose a flame detection method based on Single Shot Multi Box Detector(SSD)to improve the accuracy of flame detection and identification,and to effectively reduce the time delay of flame detection and improve the response speed of the network,the main research contents are as follows.(1)The improvement of SSD algorithm in flame detection accuracy.To address the problem of poor detection of small flames by the SSD algorithm,we propose to introduce the Convolutional Block Attention Module(CBAM)into the SSD to obtain the improved model CBAM-SSD;to address the problem of low accuracy of overall flame detection by the SSD algorithm,we propose to introduce the Feature Pyramid Network(FPN)on the basis of the CBAM-SSD model to obtain the improved model FCB-SSD.Experimental results in the self-made flame dataset show that the two improved models are better for small flame detection,and the CBAM-SSD model is 2.02% and the FCBSSD model is 2.13% better than SSD in terms of average precision.(2)Improvement of SSD algorithm in flame detection speed.For the problem of slow detection speed of SSD algorithm and the inability to achieve real-time detection,a lightweight improvement of CBAM-SSD model is proposed.Mobilenetv2 is used to replace the backbone network and the inverted residual module is used to replace the common convolution module in the auxiliary feature extraction network to reduce the number of model parameters,computation and improve the model detection speed.At the same time,in order to reduce the impact of model lightweighting on detection accuracy,the Kmeans clustering algorithm is proposed to set the aspect ratio of the prior frame in the network,so as to obtain the improved model MCB-SSD.Experimental results on the self-made flame dataset show that the MCB-SSD model is 7.43% higher than the original Mobilenetv2-SSD model in terms of average precision,and 1.88% higher than the CBAM-SSD by 1.88%;in terms of detection speed the MCB-SSD model is three times higher than the CBAM-SSD model,reaching 13 frames per second.(3)Hardware construction of real-time flame detection system.The improved lightweight flame detection algorithm is mounted on the Jeston NX hardware development board,and the camera is used to capture the current scene in real time.At the same time,the Tkinter is used to design the interface of the flame detection system and build a human-computer interaction software system to detect and alarm the flame in the scene in real time.