Research On Fire Detection Algorithm Based On Infrared Video And Deep Learning

Fire is a kind of disaster that brings danger to people and property.It is important to work to detect and alarm the fire quickly.When the traditional fire detection device is used,it is often disturbed by the environment and can not be detected in time.At the same time,it will lead to missing and false alarm of the fire.To make the detection more reliable,a combination of multiple detection schemes can be used.The infrared spectral characteristics of fire are the key factor to distinguish it from the background,with the development of artificial intelligence,image fire detection using deep learning has also attracted the attention of researchers.In this paper,the fire is detected by combining infrared video and deep learning,and the algorithm is deployed offline.According to the spectral characteristics of fire,sunlight,and artificial light source,combined with the photosensitive response curve of the CMOS image sensor,a 940nm narrow-band infrared filter is arranged in front of the camera lens to obtain the detection video under infrared conditions.According to the characteristics of infrared imaging,edge detection is used to extract the target from the background,and the area,perimeter,width,height,and coordinate features are used to match the target between frames to generate the target candidate frame.The targets in the candidate box are input into the convolution neural network for target classification.To solve the problem of the unbalanced number of positive and negative samples and the difference in training difficulty,the focal loss is set as the loss function to balance the training data.Use data augmentation to solve the problems of fewer data and lack of generalization.The depthwise separable convolution module is introduced to reduce the computational complexity of convolution and the number of parameters used,which is convenient for the identification algorithm to be deployed on a small device.Use the attention mechanism to add channel attention and spatial attention to the network and strengthen the feature extraction ability of the target.After optimizing the network performance,pruning and quantization are used to further compress the volume of the model.Compared with the original network,the performance of the improved neural network is optimized,and the accuracy is increased to 96.1%.The results are compared with multiple models.Finally,the algorithm is deployed on a low-cost and small camera.The equipment equipped with an infrared filter can detect the fire at a certain distance in real-time and produce the prediction results.The time of detecting a fire is about 3 seconds,which achieves an ideal effect in accuracy and real-time.