Research On Local Feature Modeling Methods For Video Smoke Detection

Computer vision based smoke detection methods have the ability of fast response to smoke,and have robustness against environmental disturbance.These advantages make video smoke detection an important sub-field of fire detection.Smoke varies in color,shape,flowing and flicker,etc.These factors make video based smoke detection be still challenging.Existing methods show high false alarm rates and missing alarm rates.Several reasons are responsible for this.First,many objects have similar color and texture with smoke.Second,smoke images are blurry and have low contrast.Besides,thin smoke will partly occlude background objects,thus the background objects become translucent.As a result,it is difficult to remove interference of background effectively.Local feature modeling is more suitable for video smoke detection.However,existing local feature modeling methods for smoke detection cannot achieve satisfying performance.In order to model smoke features,requiring higher levels information of local features.Many existing local features focused on the shallow features of the pixel level.Some shortcomings are concluded as follows.First,features from local regions are treated without discrimination and are combined using a single standard.The relationships of local features are lost.Second,each local region is considered independently in local feature extraction.So,the co-occurrence features of neighboring local regions are lost.Last,only single feature or partial features are modeled to avoid high dimensions.We focus on local feature modeling for video smoke detection.The coding model and modeling scheme are studied,and robust image smoke detection methods are proposed.We aim at improving the performance of smoke detection,reducing false alarm rates and error rates,and promoting the development of industrial application of video smoke detection.The detailed research of this paper is as follows:(1)A literature survey on video smoke setection.In order to comprehensively present the research results and the latest progress of video smoke detection,we summarize the general framework of video smoke detection on the basis of existing theories.Focusing on the main literatures published from 2014 to 2017,each processing step of the general framework is reviewed.The progresses and the problems are concluded,the key technologies of video smoke detection are condensed,and researches on local features are taken as the breakthrough point.(2)Re-coding modeling of local features for smoke detectionWe can model the features to capture more distinguishing information through studying the relationships between local features and exploring a re-coding modeling method with proper criteria.We propose sub oriented histograms of LBP for smoke detection and image classification.We first extract LBP codes from an image and compute the gradient of LBP codes,and then calculate sub oriented histograms to capture relations of LBP codes.Then,all the sub oriented histograms are concatenated together to form a robust feature vector,which is input into SVM for training and classifying.Experiments show that our approach not only has better performance than existing methods in smoke detection,but also has good performance in texture classification.(3)Co-occurrence local feature modeling for smoke detectionThe co-occurrence features of neighborhood regions capture higher levels of feature informations.We propose a novel feature extraction method based on similarity and dissimilarity matching measures of LBP.Given two bit-sequences of a pair of LBP codes,the similarity matching measure is defined as the ratio of the 1-1 bitwise matching number to the 0-0 bitwise matching number,and the dissimilarity matching measure is similarly defined as the ratio of the 1-0 number to the 0-1 number.We calculate the two measures between the LBP code of a center pixel and the LBP codes of its neighboring pixels in a local area,and compute the global mean of each measure.Then we compare local matching measures with its corresponding mean to propose Same Similarity Local Binary Patterns(SSLBP)and Different Similarity Local Binary Patterns(SDLBP).Finally,we concatenate the histograms of LBP,SSLBP and SDLBP to generate a final feature vector and adopt SVM for evaluation.Extensive experiments on smoke and texture datasets show that our method outperforms existing methods and has excellent generalization performance in smoke detection and texture classification.(4)Multiple local features modeling for smoke detectionCombining local features and their relationship with low dimension is the key to improving local feature combination modeling.We propose a novel feature extraction method by encoding high order directional derivatives at each pixel.We first quantize the directional derivatives into ternary values to generate Local Ternary Patterns(LTP).For the sake of simplification,each LTP code is usually decomposed into an upper LBP code and a lower LBP code,but this leads to loss of information.Hence,we use joint histograms to preserve the co-occurrence of upper and lower LBP codes for each order LTP.Then we concatenate all joint histograms from different orders to propose High-order Local Ternary Patterns(HLTP).We apply Locality Preserving Projection(LPP)to reduce the dimension of HLTP.To further improve performance,we present a noise resistant mechanism to remove noisy derivatives,and then propose HLTP based on Magnitudes of noise removed derivatives and values of Center pixels(HLTPMC).Finally,the Support Vector Machine(SVM)is used for training and classification.Experiments on large scale smoke data sets show that our method can achieve good performance.Experiments on a multi-class Brodatz texture data set also achieved good performance with low dimensional features.So our method has powerful discriminative capabilities and compact feature representation for multi-class image classification.The contributions of this paper are as follows:(1)We propose sub-oriented histograms of LBP codes according to two discrete orientations of LBP codes.The sub oriented histogram is used to capture the spatial distribution of LBP codes.Besides,Hamming distance of LBP codes with two coordinates systems is used for computation of orientations.Two coordinates systems increase the number of neighboring LBP codes for computation of orientations.(2)We propose similarity and dissimilarity matching measures to represent the relationship between a pair of LBP codes.Then we encode the two measures in a similar way of LBPs to propose Same Similarity Local Binary Patterns(SSLBP)and Different Similarity based Local Binary Patterns(SDLBP)respectively.We use different mapping modes and scalable neighborhoods to obtain rotation and scale invariances.We also analyze the relationship of SSLBP,SDLBP and LBP,and concatenate the histograms of three patterns to produce a robust feature vector for smoke and texture classification.SSLBP and SDLBP contain information on the spatial distribution of LBP codes.From another point of view,SMLBP and DMLBP can be regarded as second-order local binary patterns.(3)We present High order Local Ternary Patterns(HLTP)by quantizing high order directional derivatives into ternary values.We concatenate joint histograms of upper and lower LBP codes and use locality preserving projection to reduce the dimension of the concatenated histogram to propose efficient and compact features.In addition,we use a thresholding method to suppress noisy derivatives and then present high order local ternary patterns based on magnitudes of noise removed derivatives and values of center pixels(HLTPMC).