Fire Smoke Detection Research Based On Improved YOLOX

Smoke is used as early warning information by fire smoke detection algorithms for building fire alarm systems.Compared with sensor detection methods,smoke detection algorithms based on image processing technology have the advantages of low latency,wide coverage and low cost for large-scale applications.Fire smoke image recognition technology has been widely used in various fields,but there are still many problems in related research.Firstly,the scenes of smoke contained in the current public data set are very limited.The insufficient ability of smoke detection model to adapt to the real scene is caused by this,and there are problems of false alarms and missed alarms to be solved.Secondly,the detection accuracy has been improved or the problem of insufficient dataset has been alleviated by many deep learning improvement researches compared to the original method.However,no improved detection network has been developed specifically for the characteristics of smoke itself.In order to solve the above problems,an fire smoke detection method based on Improved YOLOX with the following main works is proposed in this thesis:(1)The detection model can be interfered with by the spatial characteristics of smoke,which include diffusion,translucency,and close combination with the background.ThetaEIOU bounding box regression loss function for smoke detection is designed to improve the prediction accuracy of smoke detection.Considering that the flame target in the image is also a crucial object that needs to be recognized by the early warning system.The Color-EIOU loss function is designed to fit the flame target.The BCE Loss function is replaced by the Focal Loss function to improve the performance of the smoke detection model caused by too many background samples in the training.(2)The characteristics of encoding and decoding module classes and location branches are analyzed in terms of network parameters,detection speed,and detection accuracy.The suitable decoupled head module is designed for the detection task.The feature fusion module of the YOLOX network is redesigned to enhance the detection of small targets.(3)Semi-supervised dataset cleaning is introduced to reduce the burden of manually labeling image target information and speed up the data set preparation.A fast and low-cost way is provided by it to expand the dataset for the maintenance and upgrading of the algorithm in real scenarios.(4)An expansion strategy is used to increase the type and number of small targets in the data set,and smoke and flame images in different scenes are taken to build a realistic and plentiful data set which can improve the generalization ability of the detection algorithm.A fire point estimation algorithm is proposed to assist the network in locating the fire point when the flame is obscured or no open flame is generated,which brings positive effects for early fire decision.(5)A comparison test of ablation experiments and other target detection algorithms are conducted to analyze the overall performance of the detection algorithm proposed in this work.The detection algorithm in this work is designed to be lightweight,to improve the inference speed of the algorithm on the edge computing platform,and to increase the practicality of the algorithm.Finally,the smoke detection algorithm proposed in this thesis is tested in a practical scenario.After the original target detection network is improved,a significant increase in the overall detection accuracy of the algorithm is observed in the experimental results.The generation of smoke can be identified in time and the current spatial position of smoke can be accurately located by the improved network.