Deep Learning Based Fire Detection Method

Fires are less harmful in the early stages,but if they are not contained in a timely and effective manner,they can easily cause harm to the safety of people and property and destabilize society.Therefore,it is of great importance for society to explore methods that enable timely and reliable fire detection.Fire detection tasks include both flame detection and smoke detection.This is because smoke detection becomes particularly important in the early stages of a fire,when a large amount of smoke is often generated and smoke is more easily observed.At the same time,it becomes critical to locate the source of the fire in order to be able to take remedial action in a timely manner.Therefore,flame detection is also an important part of the fire detection task.However,smoke detection is not sufficiently accurate due to the variable shape and scale of smoke and the tendency to form translucent smoke.In addition,the flame detection task has problems such as small target flames that are difficult to detect,unbalanced positive and negative flame samples,and easy bias in labeling information,which makes the flame detection process difficult.Therefore,specialized detection algorithms and technology research are needed for the characteristics of flame and smoke to improve the accuracy and sensitivity of fire detection,so as to better prevent and control the occurrence and hazards of fire.The focus of this thesis is to address the main difficulties in flame and smoke target detection by improving and optimizing the traditional target detection model with the aim of achieving timely and effective detection of fires.The main work of this thesis can be summarized as follows:(1)A smoke detection algorithm based on a multi-feature extraction method is proposed.A smoke detection model based on a multi-feature extraction method is proposed to overcome the problems of variable shape and scale of smoke and the difficulty of locating translucent smoke.The model uses re-parameterization convolution for smoke feature extraction,which can efficiently extract smoke features of different shapes.And an attention mechanism is introduced to enhance the spatial interaction of smoke texture features and improve the localization of translucent smoke.Finally,the problem of variable smoke scales is solved by obtaining rich multi-scale information through the convolution kernels of different scales in pyramidal convolution,which in turn improves the multi-scene smoke detection accuracy.(2)A flame detection algorithm for the flame sample problem is proposed.A flame target detection model for the flame sample problem is proposed in order to solve the problems of difficult detection of small target flames,positive and negative sample imbalance during the training process and differences in sample labeling.First,the downsampling module SPDConv is introduced in the backbone network and feature pyramid to avoid the loss of small target fine-grained information in the downsampling process.Second,the impact of positive and negative sample imbalance is reduced by assigning different weight ratios to the classification loss function.Finally,a generally distributed loss function is embedded into the border regression loss function to improve the robustness of the model to sample labeling inaccuracies.(3)A fire detection system is designed and developed by combining the algorithms proposed in this thesis with the actual requirements of the fire detection system.The system is developed with Py QT as the core technology to develop a fire detection interface that can achieve effective and fast response to fire events.