Research On Smoke Detection Technology Based On Video Images

The video-based smoke detection is a new technology involving computer vision,digital image processing and pattern recognition.Compared with the traditional smoke detection method,the video-based smoke detection has wide detection rang,fast response and low cost.Based on the research of the existing video smoke detection technology,this paper deeply studies the video smoke detection algorithm based on image separation and convolutional neural network.In this paper,we first study the combination of moving target detection and color detection to extract suspected smoke blocks.In the aspect of moving object detection,the concept of original background is introduced and the background updating model is optimized.The image is divided into small non-overlapping 24x24 patches.The improved gain-based background subtraction is used to detect motion blocks and the area growth rate of moving area is calculated to confirm whether there is a moving target of suspected smoke.In color detection,the smoke color characteristics in the YUV color space is used to filter out non-smoky motion block.By combining motion detection and color detection,most non-smoke areas of the image are filtered out to obtain the suspected smoke area,laying a good foundation for further smoke identification.Based on the similarity of adjacent pixels in the smoke image,a pure smoke model based on local smoothness is constructed and the iterative formula for separating the smoke from the background is deduced.The original LBP operator is improved,and the non-redundant local binary pattern operator is proposed to extract the smoke texture feature.At last.SVM is used to train and test the smoke.The experimental results on a large number of synthetic smoke images show that the proposed method based on local smoothed smoke image segmentation algorithm has a very close relationship α with the true value α and the smoke estimation value S is highly correlated with the real value S.The detection results of several indoor,outdoor,thick smoke,light smoke,near view and distant view smoke videos show that using SVM classifier can obtain higher smoke detection accuracy.In this dissertation,a deep learning method based on convolutional neural network is studied,and a CNN model of video smoke detection convolution neural network based on LeNet-5 model is designed.Through the training and performance comparison of a large number of smoke images,the number of convolution kernels,the number of network layers,the number of hidden features,the method of pooling,the activation function and the number of neurons in all connected layers were determined.A large number of smoke images were trained and tested on the Caffe platform.The results show that the smoke detection using CNN method has higher detection accuracy than the SVM method.