Research On Fire Detection Methods Based On Machine Learning

The fire causes serious casualties and property losses every year, so that fast and effective fire detection can rescue time and is of great significance to reduce fire losses. There are three main detection methods of fire detection technology, smoke detection, temperature detection and the visual detection technology. Visual detection technology is a new fire detection technology based on the development of computer digital image processing and pattern recognition, which can realize the non-contact detection, be of high intelligent and is not susceptible to highly toxic, explosive and the other space environment, so it has broad application prospects.Large space have open horizon and more possible interference.The traditional temperature or smoke detector is difficult to effectively cover.Image fire detection with artificial characteristics extracted is limited accuracy under the condition of the more interference. Aiming at this problem,this paper firstly puts forward traditional flame detection algorithm based on Dense-SIFT dictionary learning.Secondly,The deep learning method is innovative in applying to the flame recognition and analysis.At last, we compare the effect of deep learning with the traditional methods in flame recognition and analysis. The results show that the Dense-SIFT dictionary learning can significantly improve the recognition accuracy, timeliness,validity of judgment and stability,while,the deep learning algorithm has been able to achieve the effect of the traditional algorithm on the condition of having less hidden layers. The details are as follows:(1) The Dense-SIFT dictionary learning method to detect the fire is of three parts of moving object extraction, target recognition and analysis of the fire flame characteristics. First, Get the moving target based on the moving target detection algorithm of the improved background subtraction, using SILTP to transform the texture feature and combined with the spatial information of pixels. Secondly, the HSI color feature recognition could become flame aims for two times of the moving objects target. Finally, make the final decision for the gradient feature of the two-times goal and the scintillation characteristics by using the method of Dense-SIFT dictionary learning.(2) The deep learning method is innovative in applying to the flame recognition.Using the method of convolution neural network for flame recognition according to the characteristics of the building fire video monitoring system and video flame image. To do a comparison, respectively using Logic Regression and SVM classifier to classify the image based on the automatic feature extraction of CNN to realize the recognition of the flame.(3) We compared the traditional pattern recognition, machine learning and deep learning. The results showed that the deep learning algorithm had been able to achieve the effect of the traditional algorithm on the condition of having less hidden layers by contrast of Dense-SIFT dictionary learning and deep learning method in the effect of building fire flame detection of the video surveillance system, besides, if we increase the hidden layers it will improve the accuracy rate of recognition theoretically.