Research On Smoke Detection Algorithm Based On Semantic Segmentation Network

In the early stage of the fire,smoke often occurs first,followed by flame.Therefore,smoke detection can provide more timely warning clues than the flame,making it so significant to ensure the safety of people's lives and property and promote the development of the field of fire prevention.The traditional smoke detection methods realize alarm by monitoring the physical changes of smoke particles in the air,such as temperature,concentration and so on.However,such methods are limited in time and space.In terms of time,smoke concentration needs to reach a certain range before it can be perceived by sensors,leading to the missing of the best rescue time.From the space point of view,some open outdoor environment also brings difficulties to the installation and use of sensors.The development of video smoke detection technology solves many drawbacks of traditional methods,but the current smoke detection algorithms mostly rely on smoke feature selection in specific environments,with weak robustness and high error rate.In recent years,with the continuous progress of artificial intelligence industry,machine learning and deep learning algorithms have developed rapidly,convolutional neural network,as an important part of it,performs excellently in image classification,object detection,semantic segmentation and other large-scale video image processing.Through the training of massive data samples,the deep neural network can automatically learn the representation information reflecting the essence of the image,making it more suitable for smoke detection compared with the existing manual feature methods.Therefore,this paper applies the semantic segmentation network to the smoke detection task and builds a complete Deeplab V3 model to realize the semantic level of smoke classification,and can outline more detailed smoke contour information based on the location of the smoke area.Meanwhile,the algorithm is further optimized.First,a feature refinement module is added after the encoder structure to weaken the gridding effects caused by dilated convolutions.For the non-rigid objects such as smoke with variable scales and postures,the Atrous Spatial Pyramid Pooling module is combined with the deformable convolution to better adapt to the smoke deformation.And a channel attention decoder module is proposed to further restore the spatial details of smoke images.The final model has better detection accuracy and stronger generalization ability.According to the test results of smoke semantic segmentation data set,the average prediction time of our final model is 71.73 ms per image,the mean Pixel Accuracy is 97.78%,and the mean Intersection over Union is 91.21%,making it more suitable for smoke segmentation.Tests on public smoke videos show that this model can be applied to the complex environment better,and outperforms other video-based smoke detection methods on detection rate,with theoretical significance and application value.