Research On Deep Neural Networks Based Video Smoke Detection

Video smoke detection technology is a kind of using computer vision technology to detect smoke in video technology. Because of its large coverage area, short response time, low cost advantages, and gradually replace the traditional sensors smoke detection system, and with the rapid development of computer vision technology, received extensive attention of the researchers. Some of the existing video smoke detection technology, mostly adopt the method of single, not form a complete set of video smoke detection system. The higher the rate of false positives is the main drawback of existing technology. On the existing framework of smoke detection technology, set up a deep neural network video smoke detection system.At present, deep neural network has become a research hotspot in the field of computer vision and machine learning.Convolution neural network as an important model of deep neural network, it performed very well in the aspect of image recognition.Deep convolutional neural network through learning from large data automatically to the image characteristics, compared with the traditional artificial feature extraction methods, it can better learning to depict image characterization of the nature of information, is advantageous to the classification.As a result of the smog area irregularity, is an effective detection method. Quick video preprocessing to significantly reduce the time complexity of the whole system, first of all to block video frame for motion detection, positioning the area of movement. Then, the motion area and do color analysis, filter out most of the smog area. Smoke texture analysis is the process of the whole system mainly, is also the process which the highest time complexity. Therefore, after the video preprocessing, analyze its texture feature, which can accurately detect smoke area, and to satisfy the requirement of system real time. Again in single frame based on texture analysis, static analysis of the dynamic information on the airspace at that time, and improve the detection accuracy, reduce the detection error rate.Using C++and Caffe realized based on the depth of the neural network based video smoke detection system, and can make use of GPU acceleration texture analysis part, the basic system to meet the real-time requirements, In some public video library on analysis of the algorithm performance, compared with the existing performance of the algorithm, the depth of the experimental data show that the neural network method to further improve on the accuracy, and error detection rate significantly reduced.