| Research on definition of singular features and classification algorithms involves theknowledge about digital image processing, image pattern recognition and artificialintelligence. This paper combines the knowledge of the3domains, and probes into thepreprocessing of flame and smog image, the features selection, the search strategies andclassification algorithms of flame and smog regions.In the preprocessing part, flame and smog images in RGB, HSI, CMYK and YCbCrcolor models have been researched, and the13component images of them are processed byhistogram equalization because of the low contrast. Flame and smog regions are preliminarilysegmented by the dual-threshold segmentation method with the approriate thresholds and theright component images. The result images are better after further processed by Dilation,Opening, Closing and Hole Filling of image morphology.In the features selection part, covariance matrix descriptor are proposed for describingflame and smog regions after introducing color features, such as color moments, and motionfeatures. The vector form of covariance matrix descriptor is introduced to be applied toAdaBoost, support vector machine and sparse representation classification. In the searchstrategies part, exhaust algorithm, genetic algorithm, particle swarm optimization, grid tilingalgorithm and its improved form are researched into the search of flame and smog regions;their merits, demerits and applicable scenes are analyzed; and the parameters of geneticalgorithm and particle swarm optimization are probed to be adapted for the search of flameand smog.In the classification algorithms part, there are4classification algorithms, templateclassifier, AdaBoost, support vectors machine and sparse representation classification,proposed to classify flame regions against non-flame regions, and smog regions againstnon-smog regions. The parameters of AdaBoost and support vector machine are probed to bemore suitable for classification of flame and smog. Matching pursuit and orthogonal matchingpursuit are used to sovle sparse representation to classify flame against non-flame and smogagainst non-smog. In the experiments part, different features, different search strategies anddifferent classification algorithms are integrated, and the results of the experiments are shownand analyzed at the end of this part.At last, this paper integrates the image processing methods introduced before and theresults of the experiments, summarize that covariance matrix descriptor presents flame regionand smog region more accurately and sparse representation classification, solved by orthogonal matching pursuit, are more capable of classifying flame regions and smog regions.A detection software, which suits for complex environment and is with high robust, of flameand smog is developd based on therories of features extraction and classification algorithmsmentioned before. |
