Research On The Sparse Coding Based Video Flame Detection

This thesis studies the video automatic flame detection based on sparse coding technology. It is aimed at performing acquisition image process on a targeted scene or unit through a video capture device enabling contents and events analysis of video sequences, acquired by techniques such as image processing, pattern recognition, computer vision, and mathematical statistics. Fire detection, identification and judgment are implemented. Based on previous researches on fire detection algorithms, following works are conducted in this paper where the algorithm of sparse coding of flame is presented as the core research content.First of all. overseas and domestic research status is detailed in this thesis based on references of a large number of domestic and foreign literatures, related to flame detection. Commonly used methods for describing flame characteristics are concluded, including color feature, texture feature and shape feature, etc. Basic methodologies for fire detection based on fire movement characteristics are analyzed. Meanwhile, theoretical analysis on current popular algorithms of flame detection are performed, including flame detection algorithm, the SVM Haar-like the flame detection algorithm, primitive features detection algorithm, and summarizes these deficiencies. Disadvantages of these methods are also concluded.Secondly, in terms of existing disadvantages of traditional flame feature extracting, this thesis proposes a flame detection algorithm based on sparse coding. Starting with analyzing the principles of sparse coding, methods and procedures associated with application of sparse coding in flame are detailed here.Furthermore, this paper proposes a flame detection method based on Gabor sparse representation. The method builds a corresponding dictionary by using Gabor local characteristic. Disadvantages of the Fourier transform are not only improved, difficulties on texture expression and separation are also solved by utilizing their direction characteristics, scale invariance, translation invariance and rotation invariance properties, which heightening the accuracy of the flame image sparse representation.Finally, experimental results of various methods are presented.