Fire is one of the most common catastrophic events caused by natural or human factor,which poses a great threat to people’s production and living safety.Real-time smoke detection of surveillance video in various environments with high rate of accuracy is a significant means to reduce the risk of fire.In recent years,with the rapid development of computer vision and deep learning,the fire smoke detection algorithm based on video image content analysis has gradually replaced the traditional physical smoke sensor.At present,in the field of machine learning,vehicle identification,face recognition and other technologies like these have become more mature.But for the smoke detection of video images,a lot of image classification methods are not able to ensure the accuracy and the real-time process.Therefore,we propose following two smoke detection algorithm for video image:(1)Smoke detection algorithm of SVM based on feature fusion.After building the smoke image standard database,we extract the color feature and the texture feature of the positive samples.And the fusion feature is used as the input of training.Then we get the classifier with smoke detection function.And we propose that as the first step of smoke detection,we could extract the motion region of the video frames with mixed Gaussian model to improve the accuracy of this algorithm.(2)Smoke detection algorithm based on deep convolution neural network.After designing a deep CNN model,we show the basic network structure.We select the experiment data sets from the smoke image standard database we build and use the data sets to train the CNN model to make it obtain smoke detection function.The result on the test set shows a high accuracy of smoke detection.Then,we choose some existing classical deep CNN model for the contrast experiment.The results of the experiment shows that the proposed deep CNN model has high rate of accuracy and good real-time capability in such video image smoke detection problem. |